TWI709054B - Building device and building method of prediction model and monitoring system for product quality - Google Patents

Abstract

A building devie, a building method of prediction model, and a product quality monitoring system are provided. The buidingling parses a quality detection data set generated by detecting a plurality of products and a manufacturing data set related to production of the product. The building device includes a generating module of strong classifier. The generating module of strong classifier includes plural generators and a pre-seletion module. The generators generate respectively a plurality of candidate strong classifier groups according to different classifier strategies, the manufacturing data set and the quality detection data set. The pre-selection module determins whether the candidate strong classifier groups satisfiy with a pre-selection condition according to the quality detection data set.

Description

Predictive model establishment device, establishment method and product quality monitoring system

The present invention relates to a predictive model establishment device, establishment method, and product quality monitoring system, and in particular to a predictive model establishment device, establishment method, and product quality monitoring that can predict product quality without actual quality inspection system.

Manufacturing is an indispensable part of an industrialized society. Although different types of manufacturing produce different products, their essence is to produce products after processing production materials. Of course, the manufacturer hopes that the quality of the products produced are qualified. However, there are many variables in the production process that make the quality of the product unstable. As long as the quality of any product may have defects, the manufacturer must consider the following issues. For example, how to determine whether a product is defective, whether it is necessary to conduct quality inspections on all products, and where the problem occurs in the production process, which leads to product quality defects. In order to ensure that the quality of products meets regulations, quality testing is quite necessary for the manufacturing industry.

Please refer to Figure 1, which is a schematic diagram of conventional technology testing products through quality testing devices. After the production material 11 is processed by the production equipment 13, a product 19 is produced. After the product 19 undergoes quality inspection by the quality inspection device 17, quality inspection data 15 are generated. The quality inspection data 15 shows that the quality of the product 19 is qualified or defective.

However, the method of using the quality inspection device 17 to detect the quality of the product not only takes a considerable amount of time and money, but also cannot help the manufacturer to identify an abnormality in the production process that causes the quality of the product 19 to be defective. In other words, the production process of a product involves a lot of links, and which is the actual cause of the defects in the product 19 is not an easy problem for manufacturers to judge.

The invention relates to a predictive model establishment device, establishment method and product quality monitoring system. The predictive model establishing device, the establishing method and the product quality monitoring system of the present invention establish the predictive model so that the manufacturer of the product can only rely on the quality prediction results and defect source tracking information produced by the predictive model. The use of predictive models allows manufacturers to quickly grasp the quality of products, which not only saves a lot of time and cost for quality inspection, but also provides relevant analysis information for abnormal production processes.

According to the first aspect of the present invention, a device for establishing a prediction model is provided. The predictive model establishment device performs analysis based on the quality inspection data set generated by inspecting multiple products and the processing feature data set related to the production of these products. The establishment device includes: a strong classifier generation module. The strong classifier generation module includes: The first generator, the second generator, the third generator, and the primary selection module. The first generator generates a first candidate strong classifier group including K first candidate strong classifiers according to the first classifier strategy, the processing feature data set and the quality inspection data set, where K is a positive integer. The second generator generates a second candidate strong classifier group including K second candidate strong classifiers according to the second classifier strategy, the processing feature data set, and the quality inspection data set. The third generator generates a third candidate strong classifier group including K third candidate strong classifiers according to the third classifier strategy, processing feature data set, and quality inspection data set. The primary selection module is electrically connected to the first generator, the second generator and the third generator. The preliminary selection module determines whether the first candidate strong classifier group, the second candidate strong classifier group, and the third candidate strong classifier group meet the preliminary selection conditions according to the quality detection data set.

According to the second aspect of the present invention, a method for establishing a prediction model is proposed. The method of establishing the predictive model is analyzed based on the quality inspection data set generated by the inspection of multiple products and the processing feature data set related to the production of these products. The establishment method includes the following steps. First, according to the first classifier strategy, the processing feature data set, and the quality detection data set, a first candidate strong classifier group including K first candidate strong classifiers is generated, where K is a positive integer. Secondly, according to the second classifier strategy, the processing feature data set and the quality detection data set, a second candidate strong classifier group including K second candidate strong classifiers is generated. Then, according to the third classifier strategy, the processing feature data set, and the quality detection data set, a third candidate strong classifier group including K third candidate strong classifiers is generated. In addition, according to the quality The detection data set determines whether the first candidate strong classifier group, the second candidate strong classifier group, and the third candidate strong classifier group satisfy the primary selection condition.

According to the third aspect of the present invention, a product quality monitoring system is provided. The product quality monitoring system includes: quality inspection device, data pre-processing device and model building device. The quality inspection device inspects multiple products and generates a quality inspection data set. The data pre-processing device receives multiple production parameters related to the production of these products, and generates processing feature data sets based on them. The model building device includes: a strong classifier generation module. The strong classifier generation module includes: a first generator, a second generator, a third generator, and a primary selection module. The first generator generates a first candidate strong classifier group including K first candidate strong classifiers according to the first classifier strategy, the processing feature data set, and the quality detection data set, where K is a positive integer. The second generator generates a second candidate strong classifier group including K second candidate strong classifiers according to the second classifier strategy, the processing feature data set, and the quality inspection data set. The third generator generates a third candidate strong classifier group including K third candidate strong classifiers according to the third classifier strategy, processing feature data set, and quality inspection data set. A primary selection module is electrically connected to the first generator, the second generator and the third generator. The preliminary selection module determines whether the first candidate strong classifier group, the second candidate strong classifier group, and the third candidate strong classifier group meet the preliminary selection conditions according to the quality detection data set.

In order to have a better understanding of the above and other aspects of the present invention, the following specific examples are given in conjunction with the accompanying drawings to describe in detail as follows:

11, 31, 21: production materials

13, 23: production equipment

19, 29, P1, P100, bMP, uMP, eMP: products

17, 248: Quality inspection device

15: Quality inspection data

331: Blank material heating furnace

31a, 31b: semi-finished products

333: Warm forging press machine

335: static cooling zone

231, 232, 241: sensors

20: Factory

24: Product quality monitoring system

243: Data pre-processing device

245: Model Building Device

247: Model Use Device

249: Model Evaluation Device

PP: production parameters

2430: receiving module

2431: preprocessing module

2435: Key data selection module

2433: database

2437: Feature Conversion Module

2439: Production condition selection module

corePP1, corePP2, corePP3, corePP4, corePP5, corePP6, corePP7, corePP8, corePP9, corePP10: key production parameters

corePF1, corePF2, corePF3: key production factors

PMF1, PMF2, PMF3, PMF4, PMF5, P1MF1, P1MF2, P1MF3, P1MF4, P1MF5, P100MF1, P100MF2, P100MF3, P100MF4, P100MF5: Processing features

P1MFG, P100MFG: Processing feature data

PMFGall: Processing feature data set

bM: model building mode

reSClf1, reSClf2: check strong classifier

uM: model usage mode

eM: model evaluation mode

C1: Node path

1, 2-1, 2-2, 3-1, 3-2, 3-3, 3-4: Node

411a, 411b, 411c, 431a, 431b, 431c: processing characteristics of product bMP

413a: N1 classification conditions

413b: N2 classification conditions

413c: N3 classification conditions

415a, 415b, 415c, 433a, 433b, 433c: weak classifier

435a, 435b: Adjust the weight

bMP_origPP, uMP_origPP, eMP_origPP: original production parameters

245a: strong classifier generation module

2451a, 2451b, 2451c, 2451d, 2451e: strong classifier generator

2453: Primary Selection Module

2453a: Accuracy calculation module

2453b: Strategy Selection Module

2455: Check Module

2455a: Verification sequence calculation module

2455c: correlation calculation module

2455e: strong classifier selection module

245b: strong classifier parsing module

2457: Rule Comparison Module

2457a: Support calculation module

2457c: Coverage calculation module

2457b: Weight calculation module

2458: Node Analysis Module

2458a: Rule Item Set Conversion Module

2458b: Rule reading module

S41, S43, S45, S43a, S43c, S43e, S43g, S43h, S45, S47, S501, S503, S506, S505, S507, S509, S511, S513, S5031, S5033, S5035, S5037, S31, S33, S35, S37, S331, S333, S335, S337, S801, S802, S803, S804, S805, S807, S811, S813, S815, S817: steps

preTstDAT_1, preTstDAT_10: preliminary test data

canSClfGp_A: Candidate strong classifier group

canSClf_A1, canSClf_A10: Candidate strong classifier

pdtA1, pdtA10: individual accuracy rate

reTrnDAT: Check training data

reTstDAT: Check test data

reTst_QC: Check the quality inspection data of the verification product

QpdtA, QpdtB, QpdtD: check the quality prediction result of the verification product

preSClf_A, preSClf_B, preSClf_D: primary selection of strong classifiers

seqA, seqB, seqD: verification sequence

2471: product feature receiving module

2473: Classification rule receiving module

2475: Comparison module of product features and classification rules

2477: Similarity calculation module

2479: Quality Prediction Module

2470: Defect Source Tracking Module

Figure 1 is a schematic diagram of the quality inspection of products through quality inspection devices using conventional technology.

Figure 2 is a schematic diagram of an embodiment of the product quality monitoring system conceived in this case in the production process of bicycle shoulder blades.

Figure 3 is a block diagram of the data pre-processing device.

Figure 4 is a schematic diagram of the processing features of the predictive model input generated by the preprocessing module.

Figure 5 is a schematic diagram illustrating the processing feature data set PMFGall corresponding to the product.

Figure 6A is a schematic diagram of the product quality monitoring system in the model building mode bM, and the data pre-processing device provides the product bMP processing feature data set bMPMFGall as a model building purpose.

Figure 6B is a schematic diagram of the uMP processing feature data set provided by the data pre-processing device when the product quality monitoring system is in the model use mode uM.

Figure 6C is a schematic diagram showing that the product quality monitoring system is in the model evaluation mode eM, and the data preprocessing device provides the processing feature data set of the product eMP as a model evaluation purpose.

Figure 7 is a schematic diagram of the prediction model using a binary tree structure to classify processing features and quality inspection data.

Figure 8A is a schematic diagram of using random forest as a classifier strategy.

Figure 8B is a schematic diagram of using adaptive enhancement as a classifier strategy.

Figure 9 is a schematic diagram of using processing features to build a predictive model when the product quality monitoring system is in model building mode (bM).

Figure 10 is a block diagram of the model building device.

Figure 11 is a flowchart of the strong classifier generator generating the candidate strong classifier group canSClfGp.

Figure 12 is a schematic diagram illustrating how the primary selection module determines how to select the classifier strategy as the primary selection strategy by taking the classification result generated by the candidate strong classifier group canSClfGp_A as an example.

Figure 13 is a flow chart for the primary selection module to determine whether the classifier strategy can be used as the primary selection strategy based on the quality prediction results generated by the candidate strong classifier group canSClfGp.

Figure 14 is a schematic diagram of the strong classifier generator establishing the preliminary strong classifier preSClf, the verification sequence calculation module and the preliminary strong classifier preSClf generate the verification sequence, and the correlation calculation module calculates the correlation coefficient of the verification sequence .

Figure 15 is a flowchart of the strong classifier generation module.

Figure 16 is a schematic diagram of predicting the quality of the product uMP based on the prediction model and product processing characteristics when the product quality monitoring system is in the model use mode (uM).

Figure 17 is a block diagram of the model using device.

Figure 18 is a schematic diagram of checking whether the prediction model needs to be updated when the product quality monitoring system is in the model evaluation mode (eM).

Figure 19 shows the flow chart of the product quality monitoring system in model evaluation mode (eM).

As mentioned earlier, the product manufacturer must be able to grasp the quality of the product, but the cost of using quality inspection devices for quality inspection is too high, and it cannot effectively analyze the source of defects in the production process. For this reason, this case proposes a product quality monitoring system with production equipment. In order to explain the use of this kind of product quality monitoring system, this article will use the production process of bicycle shoulder blades as an example to illustrate how to apply the product quality monitoring system of this case.

Please refer to Figure 2, which is a schematic diagram of an embodiment of the product quality monitoring system conceived in this case in the production process of the bicycle shoulder blade. In Figure 2, it is assumed that the product quality monitoring system 24 and the production equipment 23 for producing bicycle shoulder blades are both installed in the factory 20. In actual application, the product quality monitoring system 24 can also be set outside the factory building 20, and then connected to the sensors 231, 232 and the production equipment 23 via the network or the like.

In short, the production process of bicycle shoulder blades is roughly divided into three steps. First, the billet heating furnace 331 heats the production material (aluminum block) 31 to produce a semi-finished product (aluminum block) 31a. Next, the warm forging press table 333 pressurizes and heats the semi-finished product (aluminum block) 31a to be shaped to produce a shoulder-shaped semi-finished product (aluminum block) 31b. Next, after placing the semi-finished product (aluminum block) 31b in the standing cooling zone 335 for a period of time, the production of the product 29 is completed. The quality defects of the bicycle shoulder blades may refer to the phenomenon that the size of the product 35 is deformed, bruised, or cracked.

When the production equipment 23 processes and manufactures the production material 21 to transform the production material 31 into a product 29, the production equipment 23 itself may have some machine setting parameters that need to be set. Alternatively, multiple sensors can be installed inside or around the production equipment 23 232. These sensors 232 may be used to sense the machine status of the production equipment 23 (for example, the temperature and pressure of the machine), or to sense the temperature and other characteristics of the workpiece (the production material 31 or the semi-finished products 31a, 31b). In addition, sensors 231 such as thermometers or hygrometers may also be installed in the factory building 20. For ease of description, the sensing parameters or setting parameters from different sources are collectively referred to as production parameters PP.

In Figure 2, it is assumed that the billet heating furnace 331 heats the production material 31 in four stages, including: heating at 470°C for the first stage; heating at 480°C for the second stage; heating at 490°C for the third stage ; And, the fourth stage heating at 500 ℃. After the semi-finished product 31a is taken out from the billet heating furnace 331, an infrared sensor can be used to sense the feeding temperature of the semi-finished product 31a (for example, between 440°C and 460°C). Furthermore, in the warm forging press table 333, a temperature sensor and a pressure sensor can be provided. Generally, the temperature of the upper die of the warm forging press station 333 is between 120°C and 150°C; the temperature of the lower die is between 150°C and 190°C; and the maximum forging pressure is 6 tons. Accordingly, the production parameter PP can be matched with the sensing result of the sensor 232 related to the production equipment 23.

The product quality monitoring system 24 includes a data preprocessing device 243, a model building device 245, a model using device 247, a model evaluation device 249, and a quality inspection device 248. The data pre-processing device 243 is electrically connected to the sensors 231, 232, the production equipment 23, the model building device 245 and the model using device 247; the model using device 247 is electrically connected to the model building device 245 and the model evaluating device 249; and the quality The detection device 248 is electrically connected to the model establishment device 245 and the model evaluation device 249.

The core of the product quality monitoring system 24 is a predictive model established for the purpose of product quality and analysis required by the manufacturer. With different purposes such as the establishment, use, and maintenance of the predictive model, the product quality monitoring system 24 may alternately be in the model establishment mode bM, the model usage mode uM, or the model evaluation mode eM.

For the convenience of explanation, this article puts the product produced by the production equipment 23 in the model establishment mode bM when the product quality monitoring system 24 is in the model establishment mode bM; when the product quality monitoring system 24 is in the model use mode uM, the products produced by the production equipment 23 Expressed in uMP; and, when the product quality monitoring system 24 is in the model evaluation mode eM, the products produced by the production equipment 23 are expressed in eMP.

In actual application, in each mode of the product quality monitoring system 24, the number of products bMP, uMP, and eMP produced by the production equipment 23 does not need to be limited, but can be determined by the manufacturer based on its production process or product characteristics. select. Or, you can select products bMP, uMP, eMP with sampling method. Regarding the actual application, the selection and quantity of the products bMP, uMP, eMP, and when the product quality monitoring system 24 switches the operation mode, etc., can be adjusted according to the needs of the manufacturer, and will not be detailed here.

Since the product quality monitoring system 24 must use the data pre-processing device 243 in each mode. Here, the third, fourth, and fifth figures are used to illustrate how the data pre-processing device 243 processes the production parameter PP. Next, we will use Figures 6A, 6B, and 6C to explain how to build or use a predictive model in the model building mode (bM), model use mode (uM), or model evaluation mode (eM); in the seventh, 8A, and 8B Figures illustrate the basic structure of the prediction model; Figures 9-15 illustrate the model building mode (bM); Figures 16 and 17 illustrate the model usage mode (uM); Figures 18 and 19 illustrate the model evaluation mode (eM).

Please refer to Figure 3, which is a block diagram of the data pre-processing device. The data preprocessing device 243 includes a receiving module 2430, a preprocessing module 2431, a key data selection module 2435, a database 2433, a feature conversion module 2437, and a production condition selection module 2439. Among them, the pre-processing module 2431 is electrically connected to the receiving module 2430 and the database 2433; the key data selection module 2435 is electrically connected to the database 2433 and the feature conversion module 2437; and the production condition selection module 2439 is electrically connected to the feature conversion Module 2437. In addition, the receiving module 2430 can receive production parameters PP from sensors 231 and 232 through electrical connection or signal connection; and, production condition selection module 2439 can process bMP, uMP, eMP products through electrical connection or signal connection, etc. The feature data set is sent to the model using device 247.

First, based on the consideration of reducing the amount of data, the production parameters PP sensed by the sensors 231 and 232 need to be sampled first. For example, only one production parameter per hour is taken as the sampling production parameter smpPP. In addition, the number of sensors 231 and 232 used in the production process may be quite large. Therefore, manufacturers can selectively provide selection strategies based on their mastery and understanding of the production line. The key data selection module 2435 selects which production parameters PP sensed by the sensors 231 and 232 are relatively important according to the selection strategy, and defines it as the key production parameter corePP. Alternatively, the key data selection module 2435 can be omitted, and all the sampling production parameters smpPP are directly regarded as the key production parameters corePP. In Figure 4, assume that the key data selection module 2435 selects a total of 10 key production parameters corePP1 to corePP10.

Because some of the key production parameters corePP1~corePP10 may be related to each other. For example, it may be the temperature sensed at different locations inside the same machine. Therefore, the manufacturer can provide a feature conversion formula based on the understanding and mastery of the production process, and integrate these mutually related key production parameters corePP into a key production factor PF through methods such as property induction. For example, assume that the feature conversion module 2437 integrates the key production parameters corePP3, corePP4, and corePP5 into the key production factor corePF1; integrates the key production parameters corePP6 and corePP7 into the key production factor corePF2; and integrates the key production parameters corePP9 and corePP10 as the key Production factor corePF3. The step of generating the key production factor corePF by the feature conversion module 2437 can also be omitted in practical applications.

Thereafter, the production condition selection module 2439 can select the key production parameter corePP and/or the key production factor corePF according to different uses of the prediction model, and select those that are more suitable for the use of the prediction model as processing features. Figure 4 assumes that the production condition selection module 2439 selects key production parameters corePP1, corePP5, corePP8, and key production factors corePF1 and corePF3 as processing features PMF1~PMF5.

It should be noted that as the purpose of establishing the prediction model is different (for example, to analyze the yield rate, or to analyze the root cause of the defect), the production condition selection module 2439 may have different considerations for selecting processing features. Therefore, although the key production parameters corePP2~corePP4, corePP6, corePP7, corePP9, corePP10, and the key production factor corePF2 are not selected as processing features in Figure 4, they may still be used when building predictive models for other purposes. As illustrated in Figure 4, the conversion of the production parameter PP by the data pre-processing device 243 only takes a single product as an example. In actual application, the conversion of these parameters needs to be applied to other products produced on the production line. For ease of description, Figure 5 will illustrate the correspondence between products and their corresponding processing features.

Please refer to Figure 5, which is a schematic diagram illustrating the processing feature data set corresponding to the product. In Figure 5, it is assumed that the product number is represented by 1~100, and each product P1~P100 corresponds to the five processing features PMF1~PMF5. For example, product P1 corresponds to processing features P1MF1 to P1MF5; product P100 corresponds to processing features P100MF1 to P100MF5.

In order to distinguish the processing features corresponding to different products P1~P100, the processing features corresponding to different products P1~P100 are defined here. For example, the processing features P1MF1~P1MF5 corresponding to product P1 are jointly defined as the processing features corresponding to product P1 Information P1MFG. For another example, the processing features P100MF1 to P100MF5 corresponding to the product P100 are collectively defined as the processing feature data P100MFG corresponding to the product P100. In addition, the collection formed by the processing features of multiple products P1~P100 can be defined as the processing feature data set PMFGall.

In Figures 4 and 5, the product quantity, product number, number of parameters, etc., are used as examples only. Figures 6A, 6B, and 6C illustrate how the data preprocessing device 243 generates processing feature data sets corresponding to the products bMP, uMP, and eMP in the model building mode bM, the model use mode uM, and the model evaluation mode eM, and then It is provided to the model building device 245, the model using device 247, and the model evaluating device 249 for use.

Please refer to Figure 6A, which is a schematic diagram of the product quality monitoring system in the model building mode bM, and the data pre-processing device provides the product bMP processing feature data set as a model building purpose. In the model establishment mode bM, the model establishment device 245 establishes a prediction model based on the processing feature data set of the product bMP and the quality inspection data set of the product bMP. The content of the prediction model drawn here (check the strong classifiers reSClf1, reSClf2, rule item sets and weights, etc.) will be explained later.

Please refer to Figure 6B, which is a schematic diagram of the uMP processing feature data set provided by the data preprocessing device when the product quality monitoring system is in the model use mode uM. In the model usage mode uM, the model usage device 247 uses the prediction model established based on the product bMP, takes the processing feature data set of the product uMP as input, classifies the product uMP, and uses the classification result as the quality prediction result of the product uMP . In addition, the model using device 247 can also generate relevant defect source analysis information for the product uMP classified as the defect according to the quality prediction result of the product uMP.

Please refer to Figure 6C. When the product quality monitoring system is in the model evaluation mode eM, the data pre-processing device provides the product eMP processing feature data set as the model evaluation Schematic diagram of estimated use. In the model evaluation mode eM, the model using device 247 uses the prediction model established based on the product bMP to classify the processing feature data set of the product eMP, and uses the classification result as the quality prediction result of the product eMP. On the other hand, the quality inspection device 248 will perform quality inspection on the product eMP and generate a quality inspection data set of the product eMP. Thereafter, the model evaluation device 249 compares the quality prediction result of the product eMP with the quality inspection data set, and then determines whether the previously established prediction model can be used.

In Figures 6A, 6B, and 6C, the prediction model established based on the product bMP includes multiple strong classifiers reSClf1 and reSClf2, and a list of the relationship between the rule item set and the rule weight. Among them, it is assumed that the re-selected strong classifier reSClf1 contains T1 weak classifiers; and it is assumed that the re-selected strong classifier reSClf2 contains T2 weak classifiers. In practical applications, the number of strong multiple classifiers included in the prediction model does not need to be limited. In addition, based on simplified considerations, it can be assumed that T1=T2.

Next, the basic structure of the prediction model of this case is explained. In short, the embodiment of the present invention can build multiple strong classifiers reSClf1 and reSClf2 that are generated after aggregating multiple weak classifiers using a binary tree architecture through machine learning. How to generate the reSClf1 and reSClf2 multiple-selection strong classifiers in the prediction model will be explained in Figures 11-15. Here, we will first explain the composition of the weak classifier.

Please refer to Figure 7, which is a schematic diagram of the prediction model using a binary tree structure to classify the processing characteristics and quality inspection data of the product. In the default situation, it is assumed that the strong classifier is composed of 100 weak classifiers (T=100), and the depth of each weak classifier is three layers (D=3).

According to the concept of the present invention, regardless of the operating mode of the product quality monitoring system, the weak classifiers in the reSClf check strong classifier all use a binary tree structure to classify the processing feature data sets of the products bMP, uMP, and eMP . Model building Under bM, the input of the strong classifier reSClf is the processing feature data set of the product bMP; under the model usage mode uM, the input of the strong classifier reSClf is the processing feature data set of the product uMP; and, in the model evaluation mode Under eM, the input of the strong classifier reSClf is selected as the processing feature data set of the product eMP. Among them, each node of the weak classifier is equivalent to clfCOND (classification condition of prediction model) for classifying processing features.

In the model building mode bM, the nodes of the weak classifier are generated according to the selected classifier strategy. The classification conditions represented by the nodes of the weak classifier will be repeatedly modified in the model building mode bM until the reSClf is generated. On the other hand, once the strong classifier reSClf is generated, the nodes of the weak classifiers contained in it will be used as the classification of products uMP and eMP under the model use mode uM and the model evaluation mode eM. Therefore, the weak classifier settings (number T, depth D, classification conditions represented by nodes, etc.) in the check strong classifier reSClf selected when the model was built in the model bM will not be used in the model. Modified under eM model evaluation mode.

Taking Figure 7 as an example, node 1 is the first-level node; nodes 2-1 and 2-2 are the second-level nodes; nodes 3-1, 3-2, 3-3, and 3-4 are the third-level nodes (Terminal node). In addition, the circles extending below nodes 3-1, 3-2, 3-3, and 3-4 are used to indicate the number of products that conform to the path of these nodes; the bottom box in Figure 7 indicates that the Among the number of products of these node paths, the number of products confirmed to be qualified (indicated by the white mesh bottom) and defective (indicated by the dotted mesh bottom) after quality inspection.

For ease of explanation, the following uses an example to illustrate the structure of the weak classifier and how to use the processing feature data set of the product bMP to judge. It is assumed here that under the model building mode bM, there are a total of 80 product bMPs that need to be classified. In addition, the classification conditions corresponding to each node assumed here are summarized in Table 1.

It is also assumed here that the processing characteristics of the product bMP include: the feeding temperature of the semi-finished product 31a, the heating temperature of the first to fourth stages of the blank heating furnace 331, the maximum pressure of the warm forging press 333, and the warm forging press The temperature of the upper mold and the lower mold of the table 333. In the weak classifier, the product bMP is classified according to the processing characteristics of these products and the classification conditions listed in Table 1. To simplify the description, here only the node path C1 is taken as an example to illustrate how to use the weak classifier to classify the product bMP.

First, read whether the feed temperature of the 80 bMP products is greater than 450°C (node 1). Among them, suppose that there are 35 bMPs of products with a feed temperature greater than 450°C, and 45 bMPs of products with a feed temperature less than or equal to 450°C. Next, among the 35 bMP products with a feed temperature greater than 450°C, the first stage heating temperature of those products bMP is higher than 470°C. It is assumed here that among 35 bMP products with a feed temperature greater than 450°C, there are 25 bMP products whose first stage heating temperature is higher than 470°C, and the remaining 10 bMP first stage heating temperatures are less than or equal to 470°C. After that, at node 3-1, it is judged whether the corresponding lower mold temperature is for 25 products whose feed temperature is greater than 450°C and the first stage heating temperature is greater than 470°C Above 150°C. It is assumed here that among the 25 product bMPs with the feed temperature greater than 450°C and the first stage heating temperature higher than 470°C, a total of 23 products have bMP lower mold temperatures higher than 150°C, and the remaining two products bMP lower mold temperatures Less than or equal to 150°C. Furthermore, among the 23 bMP products with the feed temperature greater than 450°C, the first stage heating temperature greater than 470°C, and the lower mold temperature greater than 150°C, a total of 15 products were judged as qualified by the quality inspection device 248, and 8 One is judged as a defect by the quality inspection device 248.

Incidentally, it is assumed that there are 10 bMP products that meet node 1 but do not meet node 2-1 (that is, the input temperature is higher than 450°C and the first stage heating temperature is less than or equal to 470°C). If the quality inspection results of the 10 bMP products are all qualified, there is no need to judge by other classification conditions at node 3-2.

Next, take Figures 8A and 8B as an example to illustrate how to generate the weak classifier shown in Figure 7. In Figures 7, 8A, and 8B, the vertical grid bottom represents the first-tier nodes; the horizontal grid bottom represents the second-tier nodes; and the top-right-bottom-left grid bottom represents the third-tier nodes. In addition, when the node is a terminal node, it is marked with a thicker frame.

For ease of description, this article assumes that the model building device 245 provides five classifier strategies (candidate strategies) as a strong classifier. In actual application, the number and types of classifier strategies are not limited to the examples in this article. Figures 8A and 8B show two examples of classifier strategies.

Please refer to Figure 8A, which is a schematic diagram of using random forest as a classifier strategy. This diagram only takes weak classifiers 415a, 415b, and 415c as an example. When the random forest strategy is adopted, the weak classifiers 415a, 415b, and 415c respectively select a part of the product bMP from the processing characteristics of the product bMP, and use it with those selected The processing characteristics of the product bMP taken. In addition, the weak classifiers 415a, 415b, and 415c are independently selected and generated from the same N production conditions.

For example, suppose there are a total of M product bMPs, and the random forest strategy provides a total of N classification conditions. Then, the input of the weak classifier 415a is to randomly select the processing feature 411a of M1 product bMP from the M product bMP, and its node system is N1 classification conditions 413a randomly selected from N classification conditions. Then, the weak classifier 415a classifies the input processing features 411a of the M1 products bMP according to the selected N1 classification conditions 413a, as shown in FIG. 7. Furthermore, the input of the weak classifier 415b is to randomly select the processing features 411b of M2 product bMPs from the M product bMP, and the node system is N2 classification conditions 413b randomly selected from the N classification conditions. Then, the weak classifier 415b classifies the input processing features 411b of the M2 products bMP according to the selected N2 classification conditions 413b as shown in FIG. 7. Similarly, the input of the weak classifier 415c is to randomly select the processing features 411c of M3 product bMPs from the M product bMPs, and the node system is N3 classification conditions 413c randomly selected from N classification conditions. Then, the weak classifier 415c classifies the input processing features 411c of the M3 product bMP according to the selected N3 classification conditions 413c as shown in FIG. 7.

Please refer to Figure 8B, which is a schematic diagram of using Adaptive Boosting as a classifier strategy. This diagram only shows three weak classifiers 433a, 433b, and 433c. When the adaptive enhancement strategy is adopted, the inputs of the weak classifiers 433a, 433b, and 433c are not completely the same.

It is also assumed that there are M product bMPs, and the adaptive enhancement strategy provides a total of N classification conditions. Then, although the weak classifier 433a is derived from the M products bMP, they are The machine selects the used M1 product bMP processing features 431a as input, but the weak classifier 433b will generate the product bMP adjustment weight 435a according to the classification result of the weak classifier 433a, and the adjustment weight 435a and the product bMP processing The feature 431a, which jointly produces the processing feature 431b of the product bMP after the first weight adjustment. The processed feature 431b of the product bMP after the first weight adjustment is regarded as the input of the weak classifier 433b. The reason for pre-weighting the input of the weak classifier 433b here is to prevent the weak classifier 433b from classifying the samples incorrectly by the weak classifier 433a and again classifying the product bMP incorrectly. The classification error mentioned here refers to the classification results generated after the processing characteristics of the product bMP are classified, and the quality inspection results generated after the actual product bMP inspection is compared, and the two are confirmed to be inconsistent .

In the same way, the weak classifier 433c adjusts the adjustment weight 435b of the product bMP again according to the classification result of the weak classifier 433b, and, based on the adjustment weight 435b, and the processing feature 431b of the product bMP adjusted for the first time, the second Processing feature 431c of the product bMP with secondary weight adjustment. The processed feature 431c of the product bMP after the second weight adjustment will be used as the input of the weak classifier 433c. Similarly, the reason for re-weighting the input of the weak classifier 433c is to prevent the weak classifier 433c from classifying the sample incorrectly by the weak classifier 433b to classify the product bMP again. That is, the weak classifier generated according to the adaptive enhancement strategy has the ability to correct the misjudgment result of the previous weak classifier, so it can achieve the effect of iterative correction.

According to the description of Figures 8A and 8B, it can be known that when different classifier strategies are used, different classification results can be generated for the same processing features and classification conditions. According to the concept of the present invention, the classification condition represented by the node of the weak classifier is based on the model The type establishment mode bM decides. In the model use mode uM and the model evaluation mode eM, the classification conditions represented by the nodes of the weak classifier will be directly used to classify the processing features of the products uMP and eMP. After that, we will further compare the weak classifiers established using these classifier strategies to classify the processing features, which one is more in line with the actual situation.

Next, this article will use Figures 9 to 15 to illustrate the model building mode bM of the product quality monitoring system 24; use Figures 16 and 17 to illustrate the model usage mode uM of the product quality monitoring system 24; and, use Figures 18 and 19 The model evaluation mode eM of the product quality monitoring system 24 is explained.

Please refer to Figure 9, which is a schematic diagram of the process of establishing a predictive model based on the original production parameter bMP_origPP of the product bMP when the product quality monitoring system is in the model establishment mode (bM). Please also refer to Figure 6A. In the model building mode bM, the production equipment 23 processes the production material 21 to produce a product (bMP) 29a. While producing the product bMP, the sensor 241 (which may be the sensors 231 and 232 in FIG. 2) generates the original parameter bMP_origPP to the data preprocessing device 243. Then, the data pre-processing device 243 transmits the converted product bMP processing feature data set to the model building device 245. On the other hand, the quality inspection device 248 performs quality inspection on the product bMP. The quality inspection data generated after the quality inspection device 248 performs quality inspection on the product bMP will be further sent to the model establishment device 245 and provided to the model establishment device 245 as a reference. Therefore, the model building device 245 receives the processing feature data set of the product bMP from the data preprocessing device 243, and the quality inspection data set of the product bMP from the quality inspection device 248.

Please refer to Figure 10, which is a block diagram of the model building device. The model building device 245 includes a strong classifier generation module 245a and a strong classifier analysis module 245b. Here only Briefly introduce the internal structure of these modules and their connection relationship. The operation details of these modules will be explained later.

The strong classifier generation module 245a includes: strong classifier generators 2451a to 2451e, a primary selection module 2453, and a multiple selection module 2455. Among them, the primary selection module 2453 includes: an accuracy rate calculation module 2453a, and a strategy selection module 2453b. Among them, the accuracy rate calculation module 2453a is electrically connected to the strong classifier generators 2451a to 2451e, the strategy selection module 2453b, and the check module 2455. The check module 2455 further includes: a verification sequence calculation module 2455a, a correlation calculation module 2455c, and a strong classifier selection module 2455e. The correlation calculation module 2455c is electrically connected to the verification sequence calculation module 2455a and the selection module 2455e, and the check module 2455 is electrically connected to the strong classifier analysis module 245b. The check module 2455 is electrically connected to the strong classifier generators 2451a~2451e and the strong classifier parsing module 245b.

The strong classifier analysis module 245b includes a rule comparison module 2457 and a node analysis module 2458. Among them, the rule comparison module 2457 includes: a support calculation module 2457a, a coverage calculation module 2457c, and a weight calculation module 2457b. The support calculation module 2457a and the coverage calculation module 2457c are both electrically connected to the node analysis module 2458 and the weight calculation module 2457b at the same time.

For ease of description, the process of establishing a prediction model by the model establishing device 245 is divided into three stages (primary selection stage STG1, multiple selection stage STG2, and analysis stage STG3). Among them, the strong classifier generation module 245a is related to the primary selection stage STG1 and the multiple selection stage STG2; the strong classifier analysis module 245b is related to the analysis stage STG3.

In the primary selection stage STG1, strong classifier generators 2451a~2451e respectively establish candidate strong classifier groups according to classifier strategies A~E canSClfGp_A~canSClfGp_E, and the primary selection module 2453 selects the primary selection strategy according to the candidate strong classifier group canSClfGp_A~canSClfGp_E. In the multiple selection stage STG2, the strong classifier generators 2451a~2451e establish the primary strong classifiers preSClf1~preSClf3 according to the primary selection strategy, and the check module 2455 selects the multiple strong classifiers from the primary strong classifiers preSClf1~preSClf3器reSClf1, reSClf2. In the parsing stage STG3, the strong classifier parsing module 245b reads the production conditions contained in the check strong classifiers reSClf1 and reSClf2, and the production conditions represented by the nodes of each weak classifier, and further compares the check strong classifiers reSClf1 After the rule item set of the weak classifier included in reSClf2, the rule weight corresponding to the rule item set is calculated.

Next, the operation of STG1 in the preliminary selection stage will be described. In the embodiment of the present invention, it is assumed that five different classifier strategies A to E are provided. First, use different strong classifier generators 2451a~2451e to generate five candidate strong classifier groups canSClfGp_A~canSClfGp_E each containing K strong classifiers. According to the concept of the present invention, the candidate strong classifier groups canSClfGp_A to canSClfGp_E generated by the strong classifier generators 2451a to 2451e each include K candidate strong classifiers (as shown in Table 2).

For ease of illustration, suppose there are 100 product bMPs in total and divide them into K parts (here, suppose K=10). Based on this assumption, each part contains 10 product bMPs. Among them, each product bMP has its corresponding processing characteristics and quality inspection data. In addition, the combination of processing features corresponding to all product bMP is defined as the processing feature data set bMPMFGall of product bMP. Then, the processing feature data set bMPMFGall is divided into K equal parts. And, these equal divisions are defined as K processing feature sub-data sets DATdiv(1)~DATdiv(K) corresponding to K candidate strong classifiers respectively. Then, according to the different value of k, select (K-1) among the self-processing feature subset data sets DATdiv(1)~DATdiv(K) as the primary training data preTrnDAT_k; and, the self-processing feature subset data set DATdiv (1) Choose one of ~DATdiv(K) as the preliminary test data preTstDAT_k.

Table 3 takes K (assuming K=10) candidate strong classifiers canSClf_A1~canSClf_A10 of the candidate strong classifier group canSClfGp_A as an example to illustrate the initial training corresponding to the kth strong candidate classifier (k=1~K) Data preTrnDAT_k and preliminary test data preTstDAT_k. Based on the assumption that there are 100 products bMP and K=10, it can be known that the processing feature sub-data contained in each preliminary training data preTrnDAT_k The material set corresponds to the processing features corresponding to 90 product bMPs; and the processing feature sub-data set contained in each preliminary test data preTstDAT_k contains the processing features corresponding to 10 product bMPs.

Then, the strong classifier generators 2451b~2451e will use different classifier strategies B~E and the preliminary training data preTrnDAT_B~preTrnDAT_E to generate candidate strong classifier groups canSClfGp_B~canSClfGp_E in a relationship similar to Table 4. According to the foregoing description, it can be known that after the initial training data preTrnDAT_1 corresponding to k=1 and the classifier strategies A~E are respectively matched, candidate strong classifiers canSClf_A1, canSClf_B1, canSClf_C1, canSClf_D1, and canSClf_E1 will be generated respectively. In the same way, the remaining preliminary training data preTrnDAT_2~preTrnDAT_10 are matched with the classifier strategies A~E respectively, and they also correspond to their corresponding candidate strong classifiers.

Please refer to Figure 11, which is a flowchart of the strong classifier generator generating the candidate strong classifier group. Each strong classifier generator 2451a~2451e will execute this process respectively. First, initialize the strong classifier counter (k=1) (step S41). Secondly, the k-th candidate strong classifier in the candidate strong classifier group is formed (step S43). Next, it is determined whether all (K) candidate strong classifiers in the candidate strong classifier group have been formed (step S45). If all K candidate strong classifiers have been formed, the process ends. If the judgment result of step S45 is negative, the strong classifier counter is accumulated (k++) (step S47), and step S43 is executed again.

Step S43 further includes the following steps: First, initialize the weak classifier counter (t=1) (step S43a). Secondly, using the preliminary training data preTrnDAT and the classifier strategy corresponding to the strong classifier generator to form the t-th weak classifier among the k-th candidate strong classifiers of the classifier strategy (step S43c). After that, it is judged whether all the weak classifiers in the k-th candidate strong classifier have been formed (assuming that each strong classifier includes T weak classifiers) (step S43g).

If the judgment result of step S43g is negative, after the weak classifier counter (t) is accumulated (step S43e), step S43c is executed again. If the judgment result of step S43g is yes Therefore, the T weak classifiers included in the k-th candidate strong classifier in the candidate strong classifier group can be collectively aggregated into the k-th candidate strong classifier in the candidate strong classifier group (step After S43h), step S43 is ended.

Please refer to Fig. 12, which takes the classification result generated by the candidate strong classifier group canSClfGp_A as an example to illustrate how the primary selection module determines how to select the classifier strategy as the primary selection strategy. As mentioned above, the candidate strong classifier group canSClfGp_A includes candidate strong classifiers canSClf_A1~canSClf_AK. For each candidate strong classifier canSClf_A1~canSClf_AK in the candidate strong classifier group canSClfGp_A, test with the preliminary test data preTstDAT_1~preTstDAT_K respectively, and obtain the individual accuracy pdtA1~pdtAK corresponding to the candidate strong classifier canSClf_A1~canSClf_AK .

Taking the candidate strong classifier canSClf_A1 as an example, the calculation method of its corresponding individual accuracy pdtA1 is as follows. Continuing the previous assumptions and the examples in Table 3, suppose there are a total of 100 products bMP1~bMP100 used as a predictive model. Then, the processing features of the products bMP1~bMP90 are used as the preliminary training data preTrnDAT_1 used in the training candidate strong classifier canSClf_A1; and the processing features of the products bMP91~bMP100 are used as the individual accuracy pdtA1 of the test candidate strong classifier canSClf_A1 The preliminary test data preTstDAT_1.

After inputting the processed features of the products bMP91~bMP100 into the candidate strong classifier canSClf_A1, multiple classification results can be generated according to the node path of the candidate strong classifier canSClf_A1. These classification results are equivalent to the quality prediction results of the products bMP91~bMP100. After that, the classification results of the products bMP91~bMP100 are compared with the quality inspection results of the products bMP91~bMP100. After the comparison, it can be known that the strong candidate classifier Among the classification results generated by canSClf_A1 corresponding to the products bMP91~bMP100, the number of correctly predicted product bMP quality (for example: 5). Then, divide the number of bMP products that can predict the quality correctly by the number of products bMP91~bMP100 (10) to obtain the individual accuracy corresponding to the candidate strong classifier canSClf_A1 (for example: 5/10*100% =50%). Similarly, the individual accuracy pdtA2~pdtAK corresponding to the candidate strong classifier canSClf_A2~AK can be calculated in a similar way.

The result of summing up the individual accuracy pdtA1~pdtAK corresponding to the candidate strong classifiers canSClf_A1~canSClf_AK is divided by the number (K) of the candidate strong classifiers canSClf in the candidate strong classifier group canSClfGp_A, you can get The group average accuracy rate pdtGA corresponding to the candidate strong classifier group canSClfGp_A. Similarly, for the candidate strong classifier group canSClfGp_B~canSClfGp_E, the group average accuracy rate pdtGB~pdtGE can also be calculated in a similar way. After that, the average accuracy of each group, pdtGA~pdtGE, will be compared with the average accuracy threshold.

Please refer to Figure 13, which is a flow chart for the primary selection module to determine whether the classifier strategy can be used as the primary selection strategy based on the quality prediction results generated by the candidate strong classifier group. The strong classifier counter (k=1) is initialized (step S501). Calculate the individual accuracy pdtk corresponding to the k-th candidate strong classifier (S503).

Next, it is determined whether all strong classifiers in the candidate strong classifier group canSClfGp have been generated (strong classifier counter k==K?) (step S505). If not, after accumulating the classifier counter k (step S506), step S503 is executed again.

If the judgment result of step S505 is affirmative, the strategy selection module 2453b sums and averages the individual accuracy rates pdt1~pdtK of the K strong classifiers in the candidate strong classifier group canSClfGp, and then obtains the same as the candidate strong classifier group Group canSClfGp correspondence The group average accuracy rate pdtG (step S507). In this article, based on the comparison between the group average accuracy pdtG and the average accuracy threshold, it is judged whether the candidate strong classifier group canSClfGp meets the primary selection conditions.

That is, the primary selection condition is equivalent to judging whether the group average accuracy pdtG of the candidate strong classifier group canSClfGp is higher than the average accuracy threshold (step S509). If the group average accuracy pdtG of the candidate strong classifier group canSClfGp is indeed higher than or equal to the average accuracy threshold, the strategy selection module 2453b confirms that the classifier strategy can be included in the primary selection strategy (step S511). At this time, it is confirmed that the candidate strong classifier group canSClfGp satisfies the primary selection conditions.

Conversely, if the group average accuracy pdtG of the candidate strong classifier group canSClfGp is lower than the average accuracy threshold, the strategy selection module 2453b confirms that the classifier strategy should not be included in the primary selection strategy (step S513). At this time, it is confirmed that the candidate strong classifier group canSClfGp does not meet the primary selection conditions.

Please refer to Table 4, which is a continuation of the comparison of the group average accuracy pdtG and the average accuracy threshold corresponding to the candidate strong classifier group in the foregoing example. In Table 4, assume that the average accuracy threshold is 80%, and Y represents the classifier strategy selected as the primary strategy; and N represents the classifier strategy that is not selected as the primary strategy.

The group average accuracy pdtGA of the candidate strong classifier group canSClfGp_A is 80%, which is equal to the average accuracy threshold. Therefore, the classifier strategy A can be regarded as Primary selection strategy. The group average accuracy pdtGB of the candidate strong classifier group canSClfGp_B is 90%, which is higher than the average accuracy threshold. Therefore, the classifier strategy B can be regarded as the primary selection strategy. The group average accuracy pdtGC of the candidate strong classifier group canSClfGp_C is 70%, which is lower than the average accuracy threshold. Therefore, the classifier strategy C is not selected as the primary strategy. The group average accuracy pdtGD of the candidate strong classifier group canSClfGp_D is 80%, which is equal to the average accuracy threshold. Therefore, the classifier strategy D can be regarded as a primary selection strategy. The group average accuracy pdtGE of the candidate strong classifier group canSClfGp_E is 60%, which is lower than the average accuracy threshold. Therefore, the classifier strategy E should not be regarded as a primary selection strategy.

In summary, classifier strategies A, B, and D will be regarded as primary selection strategies, while classifier strategies C and E will not be regarded as primary selection strategies. According to the concept of this case, those who are selected as the primary selection strategy will once again use the corresponding strong classifier generator to generate the primary strong classifier. For example, in this example, the strong classifier generators 2451a, 2451b, and 2451d are used to generate the corresponding primary strong classifiers preSClf_A, preSClf_B, and preSClf_D for the classifier strategies A, B, and D that are regarded as primary selection strategies.

According to the concept of the present invention, in the primary selection stage STG1 and the multiple selection stage STG2, the same product bMP processing feature data set bMPMFGall is used. However, in the primary selection stage STG1 and the multiple selection stage STG2, the training data and test data defined according to the processing feature data set bMPMFGall are not the same.

In the preliminary selection stage STG1, the strong classifier generators 2451a~2451e use the preliminary training data preTrnDAT and the preliminary test data preTstDAT to generate the candidate strong classifier group canSClfGp. In the STG1 segment of the preliminary selection stage, it is used to generate candidates The preliminary training data preTrnDAT and the preliminary test data preTstDAT of the selected strong classifier canSClf will change with the strong classifier counter (k) of the candidate strong classifier canSClf in the candidate strong classifier group canSClfGp.

On the other hand, in the check stage STG2. The strong classifier generators 2451a~2451e use the randomly selected multiple training data reTrnDAT as input according to the primary selection strategy to generate the primary strong classifier preSClf. After that, the check module 2455 uses the check test data reTstDAT and the check verification product quality test data reTst_QC for testing. Therefore, in the multiple selection stage STG2, the multiple training data reTrnDAT used to generate the primary strong classifier preSClf and the multiple testing data reTstDAT used to evaluate the primary strong classifier preSClf remain unchanged.

Please refer to Figure 14. After the strong classifier generator establishes the preliminary strong classifier preSClf, the verification sequence calculation module works with the preliminary strong classifier preSClf to generate the verification sequence, and the correlation calculation module calculates the correlation coefficient of the verification sequence The schematic diagram. The strong classifier generator 2451a generates the primary strong classifier preSClf_A according to the classifier strategy A and the re-selected training data reTrnDAT. The strong classifier generator 2451b generates the primary strong classifier preSClf_B based on the classifier strategy B and the re-selected training data reTrnDAT. The strong classifier generator 2451d generates the primary strong classifier preSClf_D according to the classifier strategy D and the re-selection training data reTrnDAT.

In the reselection stage STG2, check the quality test data of the verification product by reselecting the test data reTstDAT and the multiple corresponding to the reTstDAT (for example: selecting 10 of the 100 products bMP) reTst_QC as input. The primary strong classifiers preSClf_A, preSClf_B, and preSClf_D respectively predict the quality of the multiple-selected verification product based on the multiple-selected test data reTstDAT, and then generate and The quality prediction result QpdtA of multiple check verification products corresponding to the preliminary strong classifier preSClf_A, the quality prediction result QpdtB of multiple check verification products corresponding to the preliminary strong classifier preSClf_B, and the quality prediction result QpdtB corresponding to the preliminary strong classifier preSClf_D Multiple checks verify the product quality prediction result QpdtD. Then, the verification sequence calculation module 2455a compares the quality prediction results QpdtA, QpdtB, and QpdtD of the check-checked products with the quality inspection data reTst_QC of the check-checked products, and then generates multiple verification results corresponding to the check-checked products. . And, after listing these verification results one by one, the verification sequences seqA, seqB, and seqD shown in Table 5 are formed.

In Table 5, "1" and "0" respectively represent the verification sequence calculation module 2455a confirms the quality inspection data reTst_QC of the double-selected and verified products and the quality prediction results of the double-selected and verified products QpdtA, generated by the primary strong classifier preSClf. The verification result of QpdtB and QpdtD compliance. "1" means that the verification result is in compliance, and "0" means that the verification result is not in compliance. Further, the verification results are aggregated into verification sequence seq corresponding to each preliminary strong classifier preSClf. In Table 5, the verification sequence seqA corresponding to the primary strong classifier preSClf_A is {1,0,1,1,1,0,1,1,1,1}; the verification corresponding to the primary strong classifier preSClf_B The sequence seqB is {1,0,1,1,1,1,1,1,1,1}; the verification sequence seqD corresponding to the primary strong classifier preSClf_D is {1,1,0,1,1,1 ,0,1,1,1}.

When the verification result is in line ("1"), it represents the quality prediction result of the re-selected verification product predicted by the preSClf strong classifier based on the re-selected test data reTstDAT Results (for example, QpdtA, QpdtB, QpdtD) are consistent with the quality inspection data reTst_QC of the check-checked product. That is, the preliminary strong classifiers preSClf_A, preSClf_B, and preSClf_D predict that the check-validated product is qualified, and the quality inspection data of the check-checked product reTst_QC also shows that the check-checked product is qualified; or, the preliminary strong classifier preSClf_A, preSClf_B and preSClf_D predict that the check-checked product is defective, and the quality inspection data reTst_QC of the check-checked product also shows that the check-checked product is defective.

When the verification result is non-conformance ("0"), it represents the quality prediction result of the re-selected verification product (for example, QpdtA, QpdtB, QpdtD) predicted by the preSClf strong classifier preSClf according to the re-selected test data reTstDAT, and the re-selected Verify that the product quality inspection data reTst_QC is inconsistent. That is, the preliminary strong classifiers preSClf_A, preSClf_B, and preSClf_D predict that the check-checked product is defective, but the quality inspection data reTst_QC of the check-checked product shows that the check-checked product is defective; or, after the preliminary strong classifier preSClf_A, preSClf_B and preSClf_D predict that the check-checked product is defective, but the quality inspection data reTst_QC of the check-checked product shows that the check-checked product is of good quality.

After the verification sequence calculation module 2455a generates verification sequences seqA, seqB, and seqD corresponding to the primary strong classifiers preSClf_A, preSClf_B, and preSClf_D, respectively, the correlation calculation module 2455c calculates the verification sequences seqA, seqB, and seqA by the correlation calculation formula. Verification sequence correlation coefficients between seqDs (verification sequence correlation coefficient AB/BA, verification sequence correlation coefficient AD/DA, and verification sequence correlation coefficient BD/DB). Table 6 is a list of the correlation coefficient of the verification sequence and the verification sequence.

Please refer to Figure 14 and Table 6 at the same time. The verification sequence correlation coefficient A-B/B-A represents the correlation between the verification sequences seqA and seqB. The verification sequence correlation coefficient A-D/D-A represents the correlation between the verification sequences seqA and seqD. The verification sequence correlation coefficient B-D/D-B represents the correlation between the verification sequences seqB and seqD. According to the verification sequences seqA, seqB, and seqD shown in Table 5, the correlation coefficient AB/BA of the verification sequence can be calculated with the correlation calculation formula to be 0.666667; the correlation coefficient AD/DA of the verification sequence is 2.5; and the correlation coefficient BD of the verification sequence /DB is 0.16667.

The strong classifier selection module 2455e will further compare the correlation coefficients of the verification sequence AB/BA, AD/DA, BD/DB, and confirm that according to the primary selection strategy preStg1 (classifier strategy A), the primary selection strategy preStg3 (classifier strategy D ) The generated primary strong classifier preSClf has the lowest correlation with each other (the value of the correlation coefficient AD/DA of the verification sequence differs the most from 1). Therefore, the strong classifier selection module 2455e will select candidate strong classifiers canSClf_A and canSClf_D from the preliminary strong classifiers preSClf_A, preSClf_B, and preSClf_D as the reSClf1 and reSClf2 candidates. Next, how to The primary selection strategy preSelStg is selected according to multiple classifier strategies, and the process of generating the primary strong classifier preSClf is summarized in Figure 15.

Please refer to Figure 15, which is a flowchart of the strong classifier generation module. The generation of the primary strong classifier mainly includes four steps. First, the strong classifier generators 2451a to 2451e generate multiple candidate strong classifier groups (for example, candidate strong classifier groups canSClfGpA to canSClfGpE) according to the classifier strategy (step S31). Secondly, the primary selection module 2453 determines the primary selection strategy according to the candidate strong classifier group (step S33). Then, the strong classifier generators 2451a~2451e establish the preliminary strong classifiers (for example, the preliminary strong classification preSClf_A, preSClf_B, preSClf_D) according to the preliminary selection strategy (step S35); and, the check module 2455 self-selects strong Among the classifiers, two are selected as the reSClf1 and reSClf2.

Step S33 further includes the following steps. First, the accuracy calculation module 2453a calculates the group average accuracy pdtG for each candidate strong classifier group canSClfGp, and determines whether the classifier strategy should be selected as the primary selection strategy according to the group average accuracy pdtG (step S331 ). Next, it is determined whether the number of primary selection strategies is sufficient (for example, whether it is greater than or equal to 3) (step S333).

If the judgment result of step S333 is affirmative, step S35 is executed. On the contrary, the strategy selection module 2453b needs to control the corresponding strong classifier generators 2451a~2451e to modify the model structure parameters for the classifier strategies that have not been selected as the primary selection strategy, and the strong classifier generators 2451a~ 2451e generates the candidate strong classifier group canSClfGp again (step S335). The model structure parameters here are, for example, the number of weak classifiers included in the strong classifier (T), the depth of nodes included in the weak classifier (D), and so on. Step S335 The details of is and step S331, the difference is that the model structure parameters of the strong classifier generators 2451a~2451e used to generate the strong classifiers are changed.

Then, when the model structure parameters are modified and the candidate strong classifier group canSClfGp is regenerated, the accuracy calculation module 2453a recalculates the group average accuracy pdtG for the regenerated candidate strong classifier group canSClfGp, and then judges that it is not. Whether the selected classifier strategy can be selected as the primary selection strategy (step S337).

After the strong classifier selection module 2455e selects the check strong classifier reSClf, the node analysis module 2458 will further read the multiple weak classifiers contained in each check strong classifier reSClf. The root node The node path to the end node. If the previous example is continued, the rule reading module 2458b is equivalent to reading the T (for example, T=100) weak classifiers in the candidate strong classifiers canSClf_A and canSClf_D from the root node to multiple terminal nodes. Multiple node paths.

The node path from the root node to each terminal node of the weak classifier is composed of multiple nodes. As described in Figure 7, each node represents a classification condition. Therefore, the node path from the root node to each terminal node of the weak classifier must satisfy the classification condition represented by each node on the node path. For ease of description, the node path from the root node of the weak classifier to each terminal node is defined as a classification rule.

As shown in FIG. 10, the strong classifier analysis module 245b includes a node analysis module 2458 and a rule comparison module 2457. Among them, the node analysis module 2458 further includes a rule reading module 2458b and a rule item set conversion module 2458a; the rule comparison module 2457 further includes a support calculation module 2457a, a coverage calculation module 2457c, and a weight calculation module 2457b.

The rule reading module 2458b reads the classification rules represented by the respective weak classifiers of the check strong classifiers reSClf1 and reSClf2, and each node path from the root node to the terminal node. Since the nodes of each weak classifier are classified conditions, and each node path includes an unequal number of nodes, the classification conditions represented by multiple nodes on the node path can be combined to form a set of classification rules. In addition, the rule item set conversion module 2458a converts each production condition in the classification rule in a frequent item set (A-priori) manner. Therefore, the relationship shown in Table 7 is obtained.

Since the candidate strong classifiers canSClf_A and canSClf_D contain a large number of weak classifiers (for example, 100), and the node path of each weak classifier can be parsed, multiple rule item sets can be obtained, here Not to be detailed. In Table 7, several examples of the rule item sets included in the weak classifiers 1~3 and the production conditions included in the rule item sets are listed.

Please refer to Table 8, which is an example of the support calculation module 2457a calculating the support based on the rule item set.

It is assumed here that there are five rule item sets. The classification range of the item set corresponding to the rule item set 1-1 is that the feed temperature is less than the critical temperature pdtTmpTh_a. The item set classification range corresponding to the rule item set 2-1 is that the feed temperature is between the critical temperatures pdtTmpTh_a and pdtTmpTh_b, and the maximum pressure is less than the critical pressure Pth_a. The item set classification range corresponding to the rule item set 2-2 is that the feed temperature is less than the critical temperature pdtTmpTh_a, and the upper mold temperature is less than the critical temperature mlduTmpTh_a. The item set classification range corresponding to the rule item set 3-1 is that the feed temperature is less than the critical temperature pdtTmpTh_a, and the lower mold temperature is less than the critical temperature mldlTmpTh_a. The classification range of the item set corresponding to the rule item set 3-2 is that the maximum pressure is between the critical pressures Pth_a and Pth_b.

The support in Table 8 is equivalent to the proportion of the item set classification range corresponding to the rule item set in the item set classification range corresponding to all the rule item sets. For example, in Table 8, the classification range of the item set corresponding to the rule item set 1-1 (ie, the feed temperature is less than the critical temperature pdtTmpTh_a) is also repeated in the rule item set 2-2 and the rule item set 3-1 . Therefore, the item set classification range corresponding to the rule item set 1-1 (that is, the feed temperature is less than the critical temperature pdtTmpTh_a) appears three times in the total number of all rule item sets (5). Therefore, the support of the item set classification range corresponding to the rule item set 1-1 (that is, the feed temperature is less than the critical temperature pdtTmpTh_a) is 3/5.

On the other hand, the item set classification ranges corresponding to the rule item sets 2-1, 2-2, 3-1, 3-2 do not overlap with the item set classification ranges of other rule item sets. That is to say, the item set classification range corresponding to the rule item set 2-1 (ie, the feed temperature-(pdtTmpTh_a, b), the maximum pressure-(Pth_a)) only appears in the rule item set 2-1, not seen in In the item set classification range of the other four rule item sets; the item set classification range corresponding to the rule item set 2-2 (ie, feed temperature-(pdtTmpTh_a), upper mold temperature-(mlduTmpTh_a)) only appears in the rule item Set 2-2, not found in the item set classification range of the other four rule item sets; the item set classification range corresponding to the rule item set 3-1 (ie, feed temperature-(pdtTmpTh_a), and lower mold temperature-( mldlTmpTh_a)), the maximum pressure-(Pth_a)) only appears in the rule item set 3-1, not in the item set classification scope of the other four rule item sets; and, the item set corresponding to the rule item set 3-2 The classification range (ie, the maximum pressure-(Pth_a, Pth_b)) only appears in the rule item set 3-2, and is not found in the item set classification range of the other four rule item sets. Therefore, the support of rule item sets 2-1, 2-2, 3-1, and 3-2 are all 1/5.

Please refer to Table 9, which is an example of the coverage calculation module 2457c calculating the coverage based on the rule item set. The coverage rate means that among the bMP products judged to be defective, the ratio of those that meet the classification range of the item set corresponding to the rule item set to the total number of defective products bMP.

The coverage calculation module 2457c receives the rule-item set from the rule-item set conversion module 2458a, and receives the quality inspection data about the product bMP from the quality inspection module (for example, how many product bMPs are judged to be defective, and how many are judged The processing characteristics corresponding to the defective product bMP, etc.). The coverage calculation module 2457c compares the quality inspection data of the product bMP with the classification range of the item set corresponding to the rule item set, and determines whether the product bMP whose quality inspection data is judged to be defective and whether its processing characteristics conform to each rule item set ( For example, the classification range of the rule item set 1-1, 2-1, 2-2, 3-1, 3-2); and, in the product bMP whose quality inspection data is judged to be defective, the rule item set The number of itemset classification ranges.

According to the different rule item sets 1-1, 2-1, 2-2, 3-1, 3-2, the coverage calculation module 2457c will use the quality inspection data to determine the products that meet the rule set in bMP. Divide the number of bMPs by the total number of bMP products that are judged to be defective by the quality inspection data, and get the corresponding rule item sets 1-1, 2-1, 2-2, 3-1, 3-2 Coverage. As shown in Table 9, the coverage rate corresponding to rule item set 1-1 is 0.1; the coverage rate corresponding to rule item set 2-1 is 0.32; the coverage rate corresponding to rule item set 2-2 is 0.98; The coverage rate corresponding to item set 3-1 is 0.36; and the coverage rate corresponding to rule item set 3-2 is 0.08.

The weight calculation module 2457b, the self-support calculation module 2457a receives the support corresponding to the rule set 1-1, 2-1, 2-2, 3-1, 3-2, and the self-coverage calculation module 2457c After the coverage rates corresponding to the rule item sets 1-1, 2-1, 2-2, 3-1, and 3-2, the coverage ratios corresponding to the rule item sets 1-1, 2-1, 2-2, and 3- 1. Rule weight corresponding to 3-2.

Please refer to Table 10, which is an example of calculating the rule weight corresponding to the rule item set by the weight calculation module 2457b according to the support and coverage ratio and the difference of the rule item set.

The weight calculation module 2457b directly multiplies the support (0.6) corresponding to the rule item set 1-1 with the coverage rate (0.1) to obtain the original rule weight (0.6*0.1=0.06); it will be compared with the rule item set 2. After the support (0.2) corresponding to -1 is directly multiplied by the coverage (0.32), the original rule weight (0.2*0.32=0.064) is obtained; the support (0.2) corresponding to the rule set 2-2 is compared with the coverage After the rate (0.98) is directly multiplied, the original rule weight (0.2*0.98=0.196) is obtained; it will be The support (0.2) corresponding to the rule set 3-1 is directly multiplied by the coverage (0.36) to obtain the original rule weight (0.2*0.36=0.072); and the support corresponding to the rule set 3-2 After the degree (0.2) and the coverage rate (0.08) are directly multiplied, the original rule weight (0.2*0.08=0.016) is obtained. After that, the weight calculation module 2457b sums the original rule weights corresponding to the rule item sets 1-1, 2-1, 2-2, 3-1, 3-2 to obtain the sum of the original rule weights; and, The weight of each original rule is divided by the sum of the weight of the original rule to obtain the weight proportion of the original rule weight corresponding to the rule item set 1-1, 2-1, 2-2, 3-1, 3-2.

In Table 10, the range of support is between 0 and 1, and the range of coverage is between 0 and 1. Accordingly, the weight of the original rule based on the product of the two ranges from 0 to 1. In order to avoid the situation where the value of one of the support or coverage is too large, but the value of the other is just 0, resulting in the calculation result of the original rule weight being 0, the support shift amount ( For example: 0.5), and the amount of coverage shift (for example: 0.5). That is, the support degree after the translation is obtained by adding the support degree and the translation amount of the support degree, and the coverage ratio after the translation is obtained by adding the coverage ratio and the translation amount of the coverage ratio. Then, multiply the support degree after translation and the coverage ratio after translation to obtain the weight of the rule after translation. Thereafter, similar to the description in Table 10, the sum of the rule weights after translation and the proportion of the rule weights after translation corresponding to the rule item set can be calculated.

After translation processing, the range of support after translation is between 0.5 and 1.5, and the range of coverage after translation is also between 0.5 and 1.5. According to this, the range of the rule weight after translation represented by the product of the two multiplied together will be within the range of 0.5*0.5=0.25 and 1.5*1.5=2.25, so the situation where the original rule weight is 0 can be avoided.

According to the concept of the present invention, the higher the weight ratio of the original rule weights corresponding to the rule item sets 1-1, 2-1, 2-2, 3-1, 3-2, the higher the weight ratio of the original rule items set 1-1, 2-1, 2-2, 3-1, 3-2 , 2-1, 2-2, 3-1, 3-2 corresponding to the item set classification range and the higher the correlation with the defect. For example, in Table 11, the weight ratio (0.48) of the original rule weight corresponding to the rule item set 2-2 is the highest. Therefore, according to Table 8, if the processing characteristics of a certain product uMP meet the requirements of the feed temperature≦430°C and the upper mold temperature≦155°C, the chance of the product uMP being defective is also higher.

The above explains how to build the prediction model. Then continue to explain how the established prediction model can be used to predict whether the product uMP needs quality inspection; and if the product uMP is confirmed to require quality inspection, the prediction model can also be used to analyze the production equipment 23 to provide the source of defects The analysis function assists the manufacturer to determine which processing link is more likely to cause defects in the quality of this product uMP.

Please refer to Figure 16, which is a schematic diagram of predicting product quality based on the prediction model and product processing characteristics when the product quality monitoring system is in model use mode (uM). Please refer to Figure 6B and Figure 16.

The production material 21 is processed by the production equipment 23 into a product uMP. In the production process of manufacturing the product uMP, the sensor 241 provides the sensed original production parameter uMP_origPP to the data preprocessing device 243. The processing feature data set of the product uMP is converted by the data preprocessing device 243. In the model use mode uM, the model using device 247 will use the prediction model established according to the product bMP to analyze and compare the processing feature data set of the product uMP. After that, the model using device 247 will generate the quality prediction result of the product uMP and the defect source analysis information of the product uMP.

When the quality prediction result of the product uMP shows that the product uMP does not need to be sampled for quality testing, the product uMP can be shipped directly. When the quality prediction result of the product uMP shows that the product uMP needs to be sampled for quality inspection, the quality inspection device 248 performs quality inspection on the product uMP. In addition, the defect source analysis information can be used as a reference for the user to inspect the production equipment 23.

Please refer to Figure 17, which is a block diagram of the device used in the model. The model using device 247 includes a product feature receiving module 2471, a classification rule receiving module 2473, a product feature and classification rule comparison module 2475, a similarity calculation module 2477, a quality prediction module 2479, and a defect source tracking module 2470. The product feature and classification rule comparison module 2475 is electrically connected to the similarity calculation module 2477, the product feature receiving module 2471, and the classification rule receiving module 2473. The quality prediction module 2479 is electrically connected to the similarity calculation module 2477 and the defect source tracking module 2470.

The product feature receiving module 2471 is electrically connected to the data preprocessing device 243, and receives the processing feature data set of the product uMP from the data preprocessing device 243. On the other hand, the classification rule receiving module 2473 is electrically connected to the model establishing device 245, and receives from the model establishing device 245 the item set classification range corresponding to the rule item set and the weight ratio of the original rule weight corresponding to the rule item set . The processing characteristics of the product uMP received by the product feature receiving module 2471 are shown in Table 11.

In actual application, there are quite a lot of products uMP that the model using device 247 needs to analyze. For the sake of example, only products uMP1~uMP6 are taken as examples, and it is assumed that the processing characteristics include the feed temperature, the maximum pressure, and the upper mold temperature. And lower mold temperature.

After the product feature receiving module 2471 receives the processing features of each product uMP1~uMP6 listed in Table 11 (for example, the feed temperature, the maximum pressure, the upper mold temperature and the lower mold temperature), it first sets the frequent items The processing characteristics of the products uMP1~uMP6 are converted in a way. Then, compare the classification ranges of the rule item sets 1-1, 2-1, 2-2, 3-1, 3-2 listed in Table 7 with the processing characteristics of the products uMP1~uMP6. That is, confirm whether the input temperature, maximum pressure, upper mold temperature and lower mold temperature of the products uMP1~uMP6 comply with the rule set 1-1, 2-1, 2-2, 3-1, 3-2 item set classification range.

Table 12 is a confirmation table for confirming whether the product complies with the processing rules of rule item sets 1-1, 2-1, 2-2, 3-1, 3-2. Among them, Y represents the processing feature of the product uMP conforms to the item set classification range of the rule item set. The product feature and classification rule comparison module 2475 is used to perform the comparison shown in Table 12.

In Table 12, the processing feature of product uMp1 meets the item set classification range of rule item set 2-1 and 3-1; the processing feature of product uMp2 meets the item set classification of rule item set 3-2 Scope; the processing feature of product uMp3 meets the item set classification scope of rule item set 1-1, 2-1, 2-2; the processing feature of product uMp4 meets the item set classification scope of rule item set 3-1, 3-2; The processing feature of product uMp5 meets the item set classification range of rule item sets 1-1 and 3-2; and the processing feature of product uMp6 meets the item set classification range of rule item sets 1-1 and 2-1.

The product feature and classification rule comparison module 2475 generates the comparison result shown in Table 12, and then transmits the comparison result to the similarity calculation module 2477. Next, the similarity calculation module 2477 receives from the weight calculation module 2457b, as shown in the rightmost column of Table 10, the weight ratio of the original rule weight corresponding to the rule item set, and the comparison module from the product features and classification rules After 2475 receives the comparison result, it will be used to calculate the relationship between the product uMP and the weight proportion of the original rule weight corresponding to the rule item set.

Please refer to Table 13. The similarity calculation module calculates the list of defect similarity based on the product uMP and the weight proportion of the original rule weight corresponding to the rule item set. Table 13 is generated according to Table 10 and Table 12. The method is to fill in the weight percentage of the original rule weight corresponding to the rule item set calculated in Table 10 for the column listed as Y in Table 12 . Then, according to the difference of each product uMP1~uMP6, after calculating the similarity of the corresponding flaws, Table 13 can be obtained. The defect similarity can be regarded as the amount of the processing feature data set corresponding to the product uMP in the process of the production equipment 23 producing the product uMP, and the amount of the item set classification range corresponding to the rule item set.

When the defect similarity is higher, it means that among the processing features generated in the production process of the product uMp, the more processing features that are easier to cause the product uMp to produce defects, the higher the chance that the product uMp may be a defect can be predicted. Conversely, when the defect similarity is higher and lower, it means that in the process of producing product uMp, there are not too many processing features that are likely to cause defects in the product uMp. Therefore, the better chance that the product uMp may be defective is predicted. low.

As shown in Table 12, the processing characteristics of product uMP1 conform to the item set classification range of rule item sets 2-1 and 3-1. Therefore, in Table 13, the defect similarity corresponding to the product uMP1 is based on the weight ratio (0.16) of the original rule weight corresponding to the rule item set 2-1 and the rule item set 3- 1 corresponds to the sum of the weight proportions (0.18) of the original rule weights ((0.16+0.18)*100%=34%).

As shown in Table 12, the processing characteristics of product uMP2 conform to the item set classification range of rule item set 3-2. Therefore, in Table 13, the defect similarity corresponding to product uMP2 is derived from the weight ratio of the original rule weight corresponding to the rule item set 2-1 in Table 11 (0.04*100%=4%) .

As shown in Table 12, the processing characteristics of product uMP3 conform to the item set classification range of rule item sets 1-1, 2-1, and 2-2. Therefore, in Table 13, the defects corresponding to the product uMP3 The similarity is in Table 11, the weight ratio of the original rule weight corresponding to the rule item set 1-1 (0.15), the weight ratio of the original rule weight corresponding to the rule item set 2-1 (0.16), and The weight ratio (0.48) of the original rule weight corresponding to the rule item set 2-2 is added up to obtain ((0.15+0.16+0.48)*100%=79%).

As shown in Table 12, the processing characteristics of product uMP4 conform to the item set classification range of rule item sets 3-1 and 3-2. Therefore, in Table 13, the defect similarity corresponding to the product uMP4 is based on the weight ratio (0.18) of the original rule weight corresponding to the rule item set 3-1, and the original rule corresponding to the rule item set 3-2 The weight of the weight ratio (0.04) is the sum of ((0.18+0.04)*100%=22%).

As shown in Table 12, the processing characteristics of the product uMP5 meet the classification scope of the rule item sets 1-1 and 3-2. Therefore, in Table 13, the defect similarity corresponding to product uMP5 is based on the weight ratio (0.15) of the original rule weight corresponding to the rule item set 1-1, and the original rule corresponding to the rule item set 3-2 The weight ratio (0.04) is the sum of the weight ratio ((0.15+0.04)*100%=19%).

As shown in Table 12, the processing characteristics of the product uMP6 conform to the item set classification range of rule item sets 1-1 and 2-1. Therefore, in Table 13, the defect similarity corresponding to the product uMP6 is based on the weight ratio (0.15) of the original rule weight corresponding to the rule item set 1-1, and the original rule corresponding to the rule item set 2-1 The weight of the weight ratio (0.16) is the sum of ((0.15+0.16)*100%=31%).

After the similarity calculation module 2477 respectively calculates the similarity of the defects corresponding to the products uMP1~uMP6, the quality prediction module 2479 is used to resemble the defects The degree is compared with the threshold of defect similarity, and based on the comparison result, it is judged whether the product needs quality inspection.

Please refer to Table 14, which is a list for the quality prediction module 2479 to compare the defect similarity with the defect similarity threshold shown in Table 13 to determine whether the product quality inspection is required. The quality prediction module 2479 compares the defect similarity of the products uMP1 to uMP6 with the defect similarity threshold (for example, 30%). In practical applications, the value of the defect similarity threshold varies depending on the product type, yield, and other considerations.

As shown in Table 14, the flaw similarity of product uMP1 is 34%, the flaw similarity of product uMP3 is 79%, and the flaw similarity of product uMP6 is 31%, which is higher than the flaw similarity threshold (30%). Therefore, the quality prediction module 2479 determines that the products uMP1, uMP3, and uMP6 should undergo quality inspection. On the other hand, the defect similarity of product uMP2 is 4%, the defect similarity of product uMP4 is 22%, and the defect similarity of product uMP5 is 19%, which are all below the defect similarity threshold (30%). Therefore, the quality prediction module 2479 determines that the products uMP2, uMP4, and uMP5 do not require quality inspection.

In addition to generating quality prediction results based on whether the product itself needs to be quality tested, the model using device 247 can also use the prediction model to provide a defect source tracking function. According to the embodiment of the present invention, the manufacturer can provide a checklist for defect source tracking Module 2470. The root cause comparison table represents the relationship between the rule item sets 1-1, 2-1, 2-2, 3-1, 3-2 and the relevant production equipment that affects the production rules. In actual application, the root cause comparison table can be provided by the manufacturer based on experience or statistical values. Table 15 is an example of a root cause comparison table.

In Table 15, the root cause of the rule item set 1-1 is only related to the billet heating furnace 331. Therefore, when the quality of a product uMP is predicted by the prediction model to be defective because the processing characteristics meet the feature range of the rule item set 1-1, it means that the risk of the product uMP caused by the blank heating furnace 331 is 100%.

In Table 15, the root of the rule set 2-1 is due to the pressure of the billet heating furnace 331 and the warm forging press 333. Among them, the chance caused by the billet heating furnace 331 is 70%, and the chance caused by the pressure of the warm forging press 333 is 30%. Therefore, when the quality of a product uMP is predicted by the predictive model to be defective due to its processing characteristics conforming to the feature range of the rule item set 2-1, it means that the risk of defects in the uMP of this product caused by the billet heating furnace 331 is 70%. The risk of defects in the uMP of this product caused by the pressure of the warm forging press 333 is 30%.

In Table 15, the root of the rule set 2-2 is heated by the billet heating furnace 331 and the warm forging press 333. Among them, the chance caused by the billet heating furnace 331 is 50%, and the chance caused by the heating of the warm forging press table 333 is 50%. Therefore, when a product's uMP processing feature meets the feature range of the rule item set 2-2, the prediction model predicts its quality as a defect At the time, it means that the uMP defect risk of this product caused by the blank heating furnace 331 is 50%, and the uMP defect risk of this product caused by the warm forging press 333 heating is 50%.

In Table 15, the root cause of the rule set 3-1 is only related to the heating of the warm forging stamping machine 333. Therefore, when the quality of a product uMP is predicted to be a defect due to its processing characteristics conforming to the feature range of the rule item set 3-1, it means that the risk of defects in the uMP of this product caused by the heating of the warm forging stamping machine 333 is 100%.

In Table 15, the root cause of the rule set 3-2 is only related to the pressure of the warm forging press 333. Therefore, when the quality of a product uMP is predicted to be a defect due to its processing features meeting the feature range of the rule item set 3-2, it means that the risk of defects for this product uMP caused by the pressure of the warm forging stamping machine 333 is 100%.

As in Table 14, the product uMP3 has the highest defect similarity (79%). Therefore, taking the product uMP3 as an example, we will explain how the defect source tracking module 2470 is matched with the root cause comparison table shown in Table 15 to further analyze the product uMP3 production process, which is more likely to cause the product uMP3 as the defect source .

As shown in Table 12, product uMP3 is related to rule item sets 1-1, 2-1, 2-2. Therefore, for ease of explanation, the rule weights shown in Table 11 and the weight proportions of the original rule weights corresponding to the rule item sets are quoted here, and the comparison relationship between the rule item sets and root causes shown in Table 15 is summarized in Table 16. .

According to Table 16, the total weight proportion of the original rule weight corresponding to the rule item set can be calculated as 0.15+0.16+0.48=0.79. In addition, it can be seen from Table 15 that the root causes that cause the product uMP3 to be predicted to be defective include the blank heating furnace 331, the pressure of the warm forging press 333, and the heating of the warm forging press 333. Therefore, the defect source tracking module 2470 will determine the rate at which the product uMP3 may be predicted to be defective due to the pressure of the blank heating furnace 331, the warm forging press table 333 and the heating of the warm forging press table 333.

Please refer to Table 17, which is a list of the root causes and their ratios that may cause the uMP3 product to undergo quality testing. This table is aimed at the root causes that can make the product uMP3 defective, and calculates the chances of the product uMP3 being defective.

According to Table 17, the product uMP3 may be due to the failure of the billet heating furnace 331, resulting in the probability that the product uMP3 is predicted (classified) as a defect based on the rule item set 1-1 is 0.15*1; The chance that the product uMP3 is predicted (classified) as a defect is 0.16*0.7; and, the chance that the product uMP3 is predicted (classified) as a defect based on the rule item set 2-2 is 0.48*0.5. The result of adding up these three is divided by the original After the sum of the weights of the rule weights (0.79), the probability that the product uMP3 is classified as a defect due to the failure of the billet heating furnace 331 is 0.63.

According to Table 17, the product uMP3 may be classified as a defect based on the rule item set 2-1 due to abnormal pressure of the warm forging press 333, and the chance of classifying the product uMP3 as a defect is 0.16*0.3. The chance of classifying product uMP3 as a defect (0.16*0.3) is divided by the sum of the weight proportions of the original rule weight corresponding to the rule item set (0.79), and the product uMP3 can be obtained by warm forging stamping machine 333 The probability of abnormal pressure being classified as a defect is 0.63.

According to Table 17, the product uMP3 may be abnormally heated by the warm forging stamping machine 333, resulting in the classification of the product uMP3 based on the rule set 2-2, the chance of classifying the product uMP3 as defective is 0.48*0.5. In other words, the chance of classifying the product uMP3 as a defect (0.48*0.5) is divided by the sum of the weight proportions of the original rule weight corresponding to the rule item set (0.79), and then the product uMP3 can be obtained due to warm forging. The probability that the press table 333 is abnormally heated and classified as a defect is 0.31.

For the sake of comparison, the calculation results in Table 17 are summarized in Table 18. According to the concept of the present invention, the defect source tracking module 2470 can present the result to the user for reference in the form of a pie chart or the like. Therefore, the manufacturer can grasp which link of the production equipment 23 should be overhauled without going through a complicated analysis.

After the prediction model is established, it can further evaluate its prediction effect, and then confirm whether its prediction result still meets the characteristics of the production process. Evaluate the predictive model The exact timing of whether it is applicable or not can be adjusted according to the manufacturer's needs, product characteristics and other considerations, and does not need to be limited. For example, every interval is a fixed period, or after a certain number of products are produced.

Please refer to Figure 18, which is a schematic diagram of checking whether the prediction model needs to be updated when the product quality monitoring system is in the model evaluation mode (eM). Please refer to Figure 6C and Figure 18. While the production equipment 23 produces the product eMP, the sensor 241 also correspondingly generates the original production parameter eMP_origPP corresponding to the product eMP. After the data preprocessing device 243 receives the original production parameter eMP_origPP, it converts it into the product processing feature of the product eMP. The model using device 247 generates a quality prediction result corresponding to the product eMP according to the product processing characteristics and the prediction model. On the other hand, the quality inspection device 248 generates a quality inspection data set after performing quality inspection on the product eMP.

The model evaluation device 249 receives the product eMP quality prediction result from the model using device 247 and the product eMP quality inspection data set from the quality detection device 248, and compares the two to generate a model evaluation result. After that, it is determined according to the model evaluation result whether it is necessary to notify the model building device 245 to rebuild the prediction model, or to notify the model using device 247 to continue using the prediction model.

Please refer to Figure 19, which is the flow chart of the product quality monitoring system in model evaluation mode (eM). First, the update counter of the prediction model is initialized (step S801). Then, after the sensor detects the production process of the product eMP, it generates the original production parameter eMP_origPP (step S802); the data pre-processing device 243 performs data pre-processing on the original production parameter eMP_origPP to generate the processing feature data set of the product eMP (Step S803); and after the model using device 247 uses the processing feature data set of the product eMP as the input of the prediction model, the prediction model is used to generate the quality prediction result corresponding to the product eMP (step S804). On the other hand, the quality inspection device 248 detects the product eMP and generates quality inspection data corresponding to it (step S805).

After the quality prediction result and the quality detection data set are generated separately, the model evaluation device 249 compares whether the quality prediction result of the product eMP and the quality detection data are different (step S807). If the comparison result in step S807 shows that the quality prediction result still matches the quality detection result, the model evaluation device 249 informs the model using device 247b that the prediction model can still be used (step S817).

If the comparison result of step S807 is considered to be divergent, it is first judged whether the prediction model update counter has reached the preset update threshold (for example, twice) (step S811). If the judgment result of step S811 is negative, the model evaluation device 249 notifies the model creation device 245 that the prediction model needs to be rebuilt, and accumulates the prediction model update counter (step S813). After the model establishment device 245 re-establishes the prediction model, the execution from step S804 is repeated. Conversely, if the judgment result of step S811 is affirmative, after reselecting the product eMP for evaluating the prediction model (step S815), the flow of FIG. 19 is re-executed.

In conclusion, the product quality monitoring system 24 of this case first obtains the processing characteristics of the product bMP production process and the product quality inspection data set under the model building mode (bM). Establish a predictive model by analyzing the correlation between the quality of the product bMP and its processing characteristics. Secondly, under the model use mode (uM), the processing characteristics that accompany the production of the product uMP are obtained, and when the prediction model predicts that a part of the product uMP may have defects, further quality inspections are carried out for the product uMP with a higher risk of defects. And provide defect source analysis information so that the manufacturer can maintain or overhaul the production equipment 23. In addition, the product quality monitoring system 24 also provides For model evaluation mode (eM) to maintain the prediction quality of the prediction model.

Please also note that although the above description takes the production of bicycle parts as an example, the product quality monitoring system 24 of this case can be applied to various types of manufacturing factories. For the product quality monitoring system 24, although the production parameters that may be obtained by the production equipment 23 of different types of production plants are different, the steps of generating processing features performed by the data preprocessing device 243 need to be modified according to the characteristics of the production plant and the product category . However, after being converted into processing features, the operation modes of the subsequent model building device 245, model using device 247, and model evaluation device 249 are still similar. Therefore, the product quality monitoring system 24 of the present invention can be applied to various manufacturing industries.

In summary, the model establishment mode (bM), model usage mode (uM), and model evaluation mode (eM) provided by the product quality monitoring system 24 of the present invention can enable the prediction model to maintain its accuracy in predicting product quality. Since the predictive model can be used to predict the quality of the product steadily, manufacturers will be able to significantly save the cost and time required for quality inspection of product quality.

Those of ordinary knowledge in the field can understand that in the above description, the various logic blocks, modules, circuits and method steps as examples can be implemented by electronic hardware, computer software, or a combination of the two. In addition, the connection between these implementations, regardless of whether the above description adopts terms such as signal connection, connection, coupling, electrical connection or other types of alternatives, the purpose is only to illustrate the implementation of logic blocks, modules, In the circuit and method steps, different means, such as wired electronic signals, wireless electromagnetic signals, and optical signals, can be used to exchange signals in direct and indirect ways to achieve the purpose of exchange and transmission of signals, data, and control information. . Therefore, the language used in the manual does not There will be no restrictions on the realization of the connection relationship in this case, and it will not deviate from the scope of this case due to the difference in connection methods.

In summary, although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Those who have ordinary knowledge in the technical field to which the present invention belongs can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention shall be subject to those defined by the attached patent application scope.

20: Factory

231, 232: Sensor

23: Production equipment

24: Product quality monitoring system

29: Products

245: Model Building Device

243: Data pre-processing device

249: Model Evaluation Device

247: Model Use Device

31: Production materials

248: Quality Inspection Device

31a, 31b: semi-finished products

331: Blank material heating furnace

335: static cooling zone

333: Warm forging press machine

Claims (14)

A device for establishing a predictive model is analyzed based on a quality inspection data set generated by detecting a plurality of products and a processing feature data set related to the production of these products, wherein the establishing device includes: a strong classification The generator generation module includes: a first generator, which generates K first candidate strong candidates according to a first classifier strategy using a binary tree structure, the processing feature data set, and the quality inspection data set One of the classifiers is a first candidate strong classifier group, wherein each of the K first candidate strong classifiers includes T1 first weak classifiers, and K is a positive integer; a second generator, which is based on A second classifier strategy of the binary tree structure, the processing feature data set, and the quality inspection data set generate a second candidate strong classifier group including K second candidate strong classifiers, wherein each of the K Each of the second candidate strong classifiers includes T2 second weak classifiers; a third generator, which is based on a third classifier strategy using the binary tree structure, the processing feature data set, and the quality inspection data Set to generate a third candidate strong classifier group including one of K third candidate strong classifiers, wherein each of the K third candidate strong classifiers includes T3 third weak classifiers; and, a selected model Group, electrically connected to the first generator, the second generator and the third generator, The primary selection module determines whether the first strong classifier candidate group, the second strong classifier candidate group, and the third strong classifier candidate group satisfy a primary selection condition based on the quality detection data set. The establishment device described in claim 1, wherein, when the first candidate strong classifier group satisfies the primary selection condition, the first generator is based on the first classifier strategy and a multiple selection Training data to generate a first primary strong classifier; when the second candidate strong classifier group satisfies the primary selection condition, the second generator is based on the second classifier strategy and the multiple selection training data Generate a second primary strong classifier; and when the third candidate strong classifier group satisfies the primary selection condition, the third generator is generated according to the third classifier strategy and the multiple selection training data A third primary strong classifier. For example, in the establishment device described in item 2 of the scope of patent application, the multiple selection training data is randomly selected from the processing feature data set, and the processing feature data set includes the multiple selection training data and a multiple selection test data. As described in item 3 of the scope of patent application, the strong classifier generating module further comprises: a multiple selection module electrically connected to the first generator, the second generator and the third generator , Which is based on the multiple selection test data and the quality inspection data set, selecting two of the first primary strong classifier, the second primary strong classifier, and the third primary strong classifier As a first multiple selection strong classifier and a second multiple selection strong classifier. Such as the establishment device described in item 4 of the scope of patent application, wherein the multi-select module includes: A verification sequence calculation module, electrically connected to the first generator, the second generator, and the third generator, generates a first verification sequence corresponding to the first strong classifier, and the second The strong classifier corresponds to a second verification sequence, and the third strong classifier corresponds to a third verification sequence. For example, the establishment device described in item 5 of the scope of patent application, wherein the multi-selection module further includes: a correlation calculation module electrically connected to the verification sequence calculation module, which calculates the first verification sequence and the first verification sequence A first verification sequence correlation coefficient between two verification sequences, calculation of a second verification sequence correlation coefficient between the first verification sequence and the third verification sequence, and calculation of the second verification sequence and the third verification sequence A third verification sequence correlation coefficient between, and a strong classifier selection module, electrically connected to the correlation calculation module, which is based on the correlation coefficient of the first verification sequence, the correlation coefficient of the second verification sequence and the The comparison of the correlation coefficients of the third verification sequence determines the first multiple-selection strong classifier and the second multiple-selection strong classifier. The establishment device described in item 4 of the scope of patent application, wherein the establishment device further comprises: a strong classifier analysis module electrically connected to the multi-selection module, which reads the first multi-selection strong classifier The paths of the plurality of first weak classifiers included, and after reading the paths of the plurality of second weak classifiers included in the second multiple-selected strong classifier, a plurality of classification rules are obtained, and the classification Multiple rule weights corresponding to the rule. For the establishment device described in item 1 of the scope of patent application, wherein the quality inspection data set includes K pieces of quality inspection data, and the processing feature data set includes K pieces of processing feature data, wherein the K first candidates are strong The k-th first candidate strong classifier in the classifier is based on (K-1) quality inspection data in the K quality inspection data, and (K-1) processing features in the K processing feature data. Data, and the first classifier strategy; a k second candidate strong classifier in the K second candidate strong classifiers is based on the (K-1) quality test data, the (K- 1) Pieces of processing feature data and the second classifier strategy are generated; and a k-th third candidate strong classifier among the K third candidate strong classifiers is detected according to the (K-1) quality Data, the (K-1) processing feature data, and the third classifier strategy. For example, the establishment device described in item 1 of the scope of patent application, wherein, when the primary selection module determines that the first candidate strong classifier group does not satisfy the primary selection condition, the first generator is modified and the first After at least one of the plurality of first model structural parameters related to a classifier strategy, the K first candidate strong classifiers are updated; when the primary selection module determines that the second candidate strong classifier group does not satisfy the In the initial selection condition, the second generator updates the K second candidate strong classifiers after modifying at least one of the plurality of second model structure parameters related to the second classifier strategy; When the selection module determines that the third candidate strong classifier group does not satisfy the primary selection condition, the third generator is modifying at least one of the plurality of third model structure parameters related to the third classifier strategy After that, the K third candidate strong classifiers are updated. For the establishment device described in item 1 of the scope of patent application, each of the T1 first weak classifiers has a first depth, each of the T2 second weak classifiers has a second depth, and each of the T3 Each of the third weak classifiers has a third depth, wherein the first model structure parameters include T1 and the first depth, the second model structure parameters include T2 and the second depth, and the first model structure parameters include T2 and the second depth. The three model structure parameters include T3 and the third depth. The establishment device described in item 10 of the scope of patent application, wherein T1 is equal to T2. The establishment device described in claim 10, wherein the first depth is equal to the second depth. A method for establishing a predictive model, which is based on a quality inspection data set generated by testing a plurality of products and a processing feature data set related to the production of these products. The establishment method includes the following steps: According to a first classifier strategy using a binary tree structure, the processing feature data set, and the quality detection data set, a first candidate strong classifier group including K first candidate strong classifiers is generated, wherein each The K first candidate strong classifiers all include T1 first weak classifiers, and K is a positive integer; according to a second classifier strategy using the binary tree structure, the processing feature data set and the quality detection data Set to generate a second candidate strong classifier group including one of K second candidate strong classifiers, wherein each of the K second candidate strong classifiers includes T2 second weak classifiers; According to a third classifier strategy using the binary tree structure, the processing feature data set, and the quality detection data set, a third candidate strong classifier group including K third candidate strong classifiers is generated, wherein each The K third candidate strong classifiers all include T3 third weak classifiers; and, according to the quality detection data set, determine the first strong candidate classifier group, the second strong candidate candidate group, and the first candidate strong classifier group. Whether the three candidate strong classifier groups meet a primary selection condition. A product quality monitoring system includes: a quality inspection device, which detects a plurality of products and generates a quality inspection data set, and a data preprocessing device, which receives a plurality of production parameters related to the production of these products, and data To generate a processing feature data set; and, a model building device, including: a strong classifier generation module, including: a first generator, which is based on a first classifier strategy using a binary tree structure, The processing feature data set and the quality detection data set generate a first candidate strong classifier group including one of K first candidate strong classifiers, wherein each of the K first candidate strong classifiers includes T1 first Weak classifier, and K is a positive integer; a second generator, which is based on a second classifier strategy using the binary tree structure, the processing feature data set, and the quality detection data set to generate K-th A second candidate strong classifier group, one of the two strong candidate classifiers, wherein each of the K second strong candidate classifiers includes T2 second weak classifiers; A third generator, which generates a third candidate including K third candidate strong classifiers based on a third classifier strategy using the binary tree structure, the processing feature data set, and the quality detection data set A strong classifier group, wherein each of the K third candidate strong classifiers includes T3 third weak classifiers; and, a preliminary selection module electrically connected to the first generator, the second generator, and The third generator, wherein the preliminary selection module determines the first candidate strong classifier group, the second candidate strong classifier group, and the third candidate strong classifier group according to the quality detection data set Whether to meet a primary selection condition.

Images