CN108764597A - A kind of product quality control method based on integrated study - Google Patents

Abstract

The invention discloses a product quality control method based on integrated learning, which comprises the following steps for predicting key quality indicators (yield rate) of products under different progress in the production process: (1) data analysis based on injection molding process data; (2) Feature engineering analysis and construction; (3) model design based on ensemble learning; (4) data imbalance processing; (5) multi-model fusion processing scheme. Recommending optimal preset values for process parameters in the production process includes the following steps: (6) recommending overall process adjustable parameters; (7) recommending process adjustable parameters for specific process non-adjustable parameters. The present invention is suitable for dealing with the characteristics of data imbalance in industrial data, breaks through the single parameter analysis method of traditional product quality control, and discovers the abnormalities in the production process by digging out the inherent characteristic relationship between parameters by using the feature engineering construction of machine learning , Improve product quality control.

Description

A Product Quality Control Method Based on Ensemble Learning

technical field

The invention relates to the technical field of data mining, in particular to a product quality control method based on integrated learning.

Background technique

Machine learning is currently an important research field in the application of artificial intelligence, and its development is very active, while ensemble learning is a popular research direction in machine learning. "Made in China 2025" proposes to use in-depth integration of informatization and industrialization to lead and drive the development of the entire manufacturing industry and transform the manufacturing industry to Industry 4.0. However, due to the fact that the networking and intelligence of injection molding machinery has just started, the industry's information service level is low, and industry resources lack unified planning, resulting in high overall labor costs, low information level, and product added value in plastic-related industries. Low-level problems have seriously restricted the overall development of Made in China 2025.

Cloud computing and big data are the key technologies to realize Industry 4.0. The continuous update and improvement of the big data platform has driven the continuous progress of machine learning and data mining. On the basis of massive data in the injection molding industry, using big data technology and machine learning methods to solve practical industrial problems is of great significance for optimizing production and increasing production capacity. Industrial data is characterized by data imbalance, which has a great impact on the processing of machine learning algorithms. However, the popular ensemble learning algorithm in machine learning algorithms can solve the problems caused by data imbalance to a certain extent by using its own algorithm characteristics. The impact is well applied to industrial big data.

The so-called quality control refers to the operating techniques and activities adopted to meet the quality requirements. That is to say, quality control is to eliminate factors that cause unqualified or unsatisfactory effects in all stages of the quality loop by monitoring the quality formation process. Traditional product quality control analysis is more about analyzing the impact of parameters on quality indicators one by one. However, such analysis is difficult to find the internal relationship between parameters, and it is not universal.

To sum up, the problem of quality control of injection molding products is transformed into a typical machine learning problem. Based on the method of integrated learning, the data is mined, and the inherent characteristics of the data are extracted to find abnormalities in the production process and improve product quality control. For optimization production, increasing production capacity is extremely important.

Contents of the invention

The object of the present invention is to provide a product quality control method based on integrated learning in order to solve the above-mentioned defects in the prior art.

The purpose of the present invention can be achieved by taking the following technical solutions:

A kind of product quality control method based on integrated learning, described product quality control method comprises the following steps:

S1. Based on data analysis of injection molding process data, analyze mixed variables, feature discrimination, and data distribution according to injection molding process parameters;

S2. Feature engineering analysis and construction, the process is as follows:

S21. Specify the feature usage plan, that is, predict the key quality indicators of the product under different generation progress;

S22. Feature cleaning, removing some abnormal samples;

S23. Feature processing, including categorical variable processing, numerical variable processing, and time-series state monitoring index data processing. Among them, categorical variable processing is to encode categorical variables before inputting them into the model; numerical variable processing is to only contain A limited number of numerical variables are coded as categorical variables, but the original values are retained. For other numerical variables, the original values are retained. For missing values, the median value is used to fill in the processing; the data processing of time series status monitoring indicators is to process the data of time series indicators through Extract statistical values of each parameter in time stages, including mean, median, mode, maximum and minimum values, and variance;

S24. Feature selection, extracting features from the time series state index data, performing embedded feature selection methods, selecting tree model XGBoost and random forest model design methods, obtaining feature importance by using tree model XGBoost, and sorting the features, Eliminate features with low importance and reduce the feature dimension;

S3. Model design based on ensemble learning, using the arithmetic mean of the RMSE value of the evaluation index through the predicted value and the actual value as the evaluation standard. During the model training process, regarding the classification model, K cross-validation is used as the evaluation method, and AUC is selected as the performance Measurement method; for the regression model, choose K cross-validation as the evaluation method, and choose RMSE as the performance measurement method;

S4. Data imbalance processing, specifically:

S41. Data and algorithm level:

S411. By performing combined sampling on the unbalanced time series model, sampling a large sample set, combining with a small sample set to form a new sample, performing model training on the new sample set, and finally carrying out Bagging;

S412. Select the XGBoost algorithm and the DART algorithm;

S413. Perform model training on the sample set by adopting a cost-sensitive learning method, and adopt different penalty coefficients for different types of data in the XGBoost algorithm;

S414, adopting the tree model Dart that introduces deep learning, and introducing the Dropout method of deep learning for processing to prevent model overfitting;

S42. At the model fusion level, the classification model and the regression model are fused: for the key quality index prediction, the key quality index of each batch of products is predicted through the regression model. Category, use the classification model for model prediction, and finally use the method shared by classification and regression methods for data processing;

S5. Multi-model fusion processing, specifically:

S51, regression model fusion adopts the method of weighted average;

S52. The classification model fusion adopts two binary classification models. After the model training is completed, the test set is predicted to obtain the probability that the key_index of each sample is lower than 0.92 or higher than 0.98, and the predicted value of the sample with high confidence is limited to 0.92 or 0.98.

Further, the described product quality control method also includes the following steps:

R2. Define the feature usage plan, that is, recommend the optimal preset value of the process parameters in the production process to obtain better key quality indicators, as follows:

R21. The overall process adjustable parameter recommendation, specifically:

Dig out the best parameter combination that maximizes the yield rate, group the parameter combinations, get all the parameter combinations that appear in the training data, and calculate the mean, median, maximum, minimum, and value of each combination's yield rate Combine the number of occurrences, get the statistical table, sort according to the average yield rate from large to small, accumulate each adjustable parameter in the Top20/30/40 parameter combination, and find out the majority value of each adjustable parameter as a recommendation;

R22. For the non-adjustable parameters of a specific process, recommend the adjustable parameters of the process, specifically:

Firstly, from the training data, product batches whose yield rate is greater than a certain threshold (threshold can be adjusted according to demand) are screened out, and the parameters of these product batches are used as candidate values; then, for new product batches, the process non-adjustable parameter table is used as Features, find the most similar or Top k most similar samples from the candidate samples, and take its adjustable parameters as recommendations.

Further, in step R21 , in the overall process adjustable parameter recommendation, for the double-type adjustable parameters, the median or mean value is taken as the recommendation.

Further, in step R22, for a specific process non-adjustable parameter, the problem is transformed into a similarity measurement problem in recommending process adjustable parameters, and different distance measurement methods are used according to the type of value to obtain the process non-adjustable parameter. The importance of the numerical parameters in the parameters is assigned weights, that is, the weighted Euclidean distance is completed.

Further, the injection molding process parameters include injection pressure, injection time, injection temperature, holding pressure and time, back pressure, and rotation speed.

Further, in the step S22, for the characteristics of unbalanced data, a combination of up and down sampling is adopted at the data level; at the algorithm level, the Boosting ensemble learning algorithm is selected as the basic algorithm model, and AUC is used as the evaluation index of the classification result.

Further, in the step S23, the categorical variable processing and the numerical variable processing adopt one-hotencode to code the variables.

Further, the regression model selects three models of XGBoost, DART, and RandomForest, and selects a common parameter for model tuning: min_child_weight. For regression problems, this parameter corresponds to the minimum number of samples on each leaf node;

The classification model selects two binary classification models, one of which predicts whether the key_index of the sample is lower than 0.92, and the other predicts whether the key_index of the sample is higher than 0.98. Both models use "binary:logistic" as the objective function and AUC as an evaluation indicator.

Further, the above step S411 is realized by using the Random Forest model.

Compared with the prior art, the present invention has the following advantages and effects:

(1) The present invention creatively proposes a product quality control method based on ensemble learning, which breaks the traditional product quality control analysis method of parameter-by-parameter analysis, and uses machine learning data mining methods to analyze data and construct feature engineering, Extract the hidden features, and analyze the internal relationship between these parameters and the impact on quality indicators through model training.

(2) Aiming at the unbalanced problem of industrial data, the present invention adopts a solution strategy of multi-faceted fusion of data level and algorithm level, and the combination of up and down sampling in the data aspect can adjust the problem of uneven distribution of data sets from the data level; while the algorithm level adopts multiple An integrated learning algorithm, a cost-sensitive learning method, an AUC model evaluation index, and a multi-model fusion method can be modified to adapt the algorithm to the imbalance of the data, so that the data can be predicted and analyzed more accurately.

(3) The present invention uses a variety of efficient integrated learning methods to complete model training. On the one hand, it can be used to adapt to the above-mentioned data imbalance, and can also prevent data overfitting, obtain higher prediction accuracy, and have efficient operations Efficiency can reduce resource utilization and improve performance; RandomForest and DART reduce data overfitting through different aspects, and distributed XGBoost can have better accuracy and higher computing performance.

(4) The present invention adopts the parameter recommendation mode combining the whole and the part. In terms of parameter recommendation, the present invention not only recommends the adjustable parameters according to the result of the yield rate, but also recommends the influence of the overall non-adjustable parameters, like this The influence of all parameters can be better and more comprehensively considered, and a parameter combination with better effect can be obtained; at the same time, in the parameter recommendation, a combination of multiple methods is used to obtain the optimal parameter combination through problem conversion.

Description of drawings

Fig. 1 is the work flowchart of the product quality control method based on integrated learning of the embodiment of the present invention;

Fig. 2 is a data distribution scatter analysis diagram of the present invention;

Fig. 3 is a feature importance ranking diagram of the present invention;

Fig. 4 is a diagram of the model fusion scheme of the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments It is a part of embodiments of the present invention, but not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

Example

This embodiment takes the product quality control of the injection molding industry as the research object, based on 1) predicting the key quality indicators (yield rate) of products under different progress in the production process; 2) recommending the optimal preset value of the process parameters in the production process, In order to obtain better key quality indicators, two methods are used to conduct applied research on the quality control of injection molding products. Use big data technology and machine learning methods to discover abnormalities in the production process and improve product quality control.

As shown in Figure 1, the product quality control method based on integrated learning in this implementation includes two aspects: 1) predicting the key quality indicators (yield rate) of products under different progress in the production process; 2) analyzing the process parameters in the production process Recommend optimal preset values to obtain better key quality indicators.

Among them, based on problem 1, the following steps are included:

S1. Data analysis based on injection molding process data, as follows:

The production process of the manufacturing industry generally includes several links such as material selection, processing, and product quality inspection, and the injection molding industry is no exception. The injection molding process refers to the process of making semi-finished parts of a certain shape from molten raw materials through operations such as pressurization, injection, cooling, and detachment. Injection molding process parameters mainly include injection pressure, injection time, injection temperature, holding pressure and time, back pressure, speed, etc. The production process data of manufacturing products are divided into process adjustable parameters and process non-adjustable parameters. , Timing status monitoring indicators. By observing and analyzing the process parameter data, it is found that the data has the following characteristics: mixed variables (numerical, categorical, etc.), insufficient feature discrimination, and unbalanced data distribution.

S2. Feature engineering analysis and construction, the specific steps are as follows:

S21. Specify the feature usage plan, that is, predict the key quality index (yield rate) of the product under different generation progress.

S22. Feature cleaning, eliminating some abnormal samples. For the characteristics of unbalanced data, a combination of up and down sampling is adopted at the data level; at the algorithm level, the method of cost-sensitive learning is selected, and the Boosting ensemble learning algorithm is selected as the basic algorithm model, AUC As the evaluation index of classification results. Algorithm level and algorithm multi-faceted solution strategies are adopted.

S23. Feature processing, mainly including categorical variable processing, numerical variable processing, and time series state monitoring index data processing. The categorical variable processing mainly includes categorical variables in both adjustable and non-adjustable variable values, and encodes them before inputting into the model, using the commonly used one-hot encode. In the processing of numerical variables, there are only a limited number of reference values, and they are also treated as category types for one-hot encoding, but the original values are retained. For other numerical variables, keep the original value, and for the missing value, fill it with the median value. Time series state monitoring index data processing: In the time series index data, many index data, for this part of the index related to time series, the statistical value of each parameter is extracted by time stages, including mean, median, mode, maximum/minimum value ,variance. Specifically, according to add_time, the processing process of each product_no is evenly divided into 10 stages, and various statistics of each parameter in each time period are counted. In addition, the parameter statistics of the whole processing process and the processing progress of 50% are also extracted. Although these statistical features are simple, they are discriminative to a certain extent. For example, if the temperature variance is too large within a certain period of time, it may have a relatively large impact on the quality of the product. Through this staged approach, the characteristics of time series data are reflected. And construct the approximate total processing time and parameter difference features.

S24. Feature selection: Several hundred-dimensional features are extracted from the time series state index data. Many features are redundant and easily cause overfitting. Here, an embedded feature selection method is used because the tree is selected in the model design part. Model XGBoost and random forest, by using the tree model to get the feature importance, and sort the features, remove the low importance features, and finally retain only about 100-dimensional features

Step S3, model design based on integrated learning, specifically:

S31. Model evaluation plan

Based on the problem of predicting the key quality index value (yield rate) of different progress in the production process, the final evaluation index uses the arithmetic mean of the RMSE value of the predicted value and the actual value as the evaluation standard. In the process of model training, for the classification model, K cross-validation is used as the evaluation method, and AUC is selected as the performance measurement method; for the regression model, K cross-validation is also selected as the evaluation method, and RMSE is used as the performance measurement method.

S32. Regression model

Three models of XGBoost, DART, and RandomForest were selected for the regression model. Regarding model tuning, since the model uses an integrated learning algorithm, the three tree models have a common and important parameter: min_child_weight. For regression problems, this parameter corresponds to the minimum number of samples on each leaf node. , the smaller this parameter is, the easier it is to overfit. The min_child_weight set in this scheme is 5, which is a big improvement in RMSE compared to the default value of 1.

S33. Classification model

The regression model selects the XGBoost model, and the predicted value of the regression model is concentrated between [0.92,0.98]. Therefore, two binary classification models are established, one of which predicts whether the key_index of the sample is lower than 0.92, and the other predicts whether the key_index of the sample is Is it higher than 0.98. The features used by these two classification models are the same as those used by the regression model. The model also uses XGBoost, but the model uses "binary:logistic" as the objective function and AUC as the evaluation indicator.

Step S4, data imbalance processing, specifically:

S41. Data and algorithm level:

S411, by performing combined sampling on the unbalanced time series model, sampling a large number of sample sets (such as sampling 20%, the specific ratio needs to be cross-validated), and combining with a small sample set to form a new sample (such as a 5-person sample set ), model training is carried out for the new sample set, and finally Bagging is carried out. Here, the Random Forest model is mainly used.

S412. Select a Boosting-related algorithm. Here, the XGBoost algorithm and the DART algorithm are mainly selected. For the classification model, AUC is selected as the evaluation index of the classification results.

S413. Perform model training on the sample set by using a cost-sensitive learning method, and adopt different penalty coefficients for different types of data in the XGBoost algorithm.

S414. Adopting the tree model Dart introduced with deep learning. After the data is under-sampled and processed, if the amount of data is not large enough, it is easy to overfit. The dropout method of deep learning can be used to prevent the model from overfitting.

S42. Model fusion level:

S421. Because the above algorithm selection uses an ensemble learning algorithm, there are two methods for model fusion: one is the fusion method of linearly weighting different ensemble learning algorithms; Model fusion is performed by linear model cross fusion.

S422. The method of integrating classification model and regression model: For the prediction of key quality indicators, the key quality indicators of each batch of products are predicted through the regression model. Due to the characteristics of data imbalance, the unprocessed few-sample data can be regarded as a small category , use the classification model for model prediction, and finally use the shared method of classification and regression methods for data processing.

Step S5, multi-model fusion processing scheme, specifically:

S51. Regression model fusion:

The three models XGBoost, Dart, and Random Forest are all tree-based models. XGBoost and DART are boosting algorithms that focus on reducing model bias. Random forest is a bagging algorithm that focuses on reducing the variance of the model. Combining these several models can further improve the performance of the model. The commonly used fusion method with better effect is stacking or blending, and multi-level stacking/blending learning is easy to overfit, so this solution only uses a simple weighted average method: 0.4* XGB+0.4*DART+0.3*RF, the weight is adjusted according to the offline performance of the model.

S52. Classification model fusion:

Two binary classification models, after the model training is completed, predict the test set, get the probability that the key_index of each sample is lower than 0.92 or higher than 0.98, and limit the predicted value of those samples with high confidence (high probability) to 0.92 Or 0.98. Using this post-processing scheme has greatly improved the final evaluation results.

For problem two, the product quality method based on integrated learning disclosed by the present invention also includes the following steps:

R2. Define the feature usage plan, that is, recommend the optimal preset value of the process parameters in the production process to obtain better key quality indicators, as follows

R21. Overall process adjustable parameter recommendation;

R22. For the non-adjustable parameters of a specific process, recommend the adjustable parameters of the process.

Step R21, overall process adjustable parameter recommendation, specifically:

Dig out the best parameter combination that maximizes the yield rate, and group the parameter combinations to get all the parameter combinations that appear in the training data, and calculate the mean, median, maximum, and minimum value of the yield rate of each combination, and The number of occurrences of each combination. Get the statistical table, there are only 91 parameter combinations in total, sorted according to the average yield rate from large to small, accumulate each adjustable parameter in the Top20/30/40 parameter combination, and find out the value of each adjustable parameter The mode value is used as a recommendation. For double-type adjustable parameters, the median or mean value can also be used as a recommendation. In addition, some fixed parameter combinations are also considered when recommending.

Step R22, recommending process adjustable parameters for specific process non-adjustable parameters, specifically:

The solution is to first screen out product batches with a yield rate greater than a certain threshold (threshold can be adjusted according to demand) from the training data, and use the parameters of these product batches as candidate values. For a new product batch (ie, a test sample), the process non-adjustable parameter table is used as a feature, and the most similar (or Top k most similar) samples are found from the candidate samples, and its adjustable parameters are used as recommendations. The problem is transformed into a similarity measure (distance measure) problem. According to the type of value, different distance measurement methods are used, such as MinkovDM, to obtain the importance of the numerical parameters in the process non-adjustable parameters, and to assign weights, that is, to complete the weighted Euclidean distance.

The above-mentioned embodiment is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited by the above-mentioned embodiment, and any other changes, modifications, substitutions, combinations, Simplifications should be equivalent replacement methods, and all are included in the protection scope of the present invention.

Claims (9)

1. a kind of product quality control method based on integrated study, which is characterized in that the product quality control method packet Include the following steps：

S1, the data analysis based on Shooting Technique data, according to molding proces s parameters, analysis mixed type variable, feature decision, Data distribution；

S2, Feature Engineering analysis and structure, process are as follows：

S21, clear feature operational version, that is, predict product Key Quality Indicator under different manufacturing schedules；

S22, feature cleaning, reject the abnormal sample in part；

S23, characteristic processing, including class variable processing, the processing of numeric type variable, time sequence status monitor control index data processing, In, class variable processing is that coded treatment is carried out before input model to classification type variable；The processing of numeric type variable is to value It contains only limited several numeric type variable and carries out coded treatment as classification type variable, but retain raw value, for other Numerical variable keeps initial value, for missing values, is handled with intermediate value filling；Time sequence status monitor control index data processing is referred to sequential Mark data pass through the statistical value for dividing time phase to extract parameters, including mean value, intermediate value, mode, maximum and small value, variance；

S24, feature selecting extract feature from time sequence status achievement data, carry out Embedded feature selection approach, selection tree The design methods of model XGBoost and random forest obtain feature importance by using tree-model XGBoost, and right Feature is ranked up, and is rejected the low feature of importance, is reduced intrinsic dimensionality；

Evaluation metrics are passed through the arithmetic equal value of the RMSE value of predicted value and actual value by S3, the modelling based on integrated study As evaluation criteria, during model training, about disaggregated model, appraisal procedure is used as by K cross validations, selects AUC As performance metric method；About regression model, selects K cross validations as appraisal procedure, select RMSE as performance metric Method；

S4, data nonbalance processing, specially：

S41, data and algorithm level：

S411, by doing com bined- sampling to unbalanced time series models, more sample sets is sampled, with few sample Collection is combined into new sample, carries out model training for new sample set, finally carries out Bagging；

S412, selection XGBoost algorithms and DART algorithms；

S413, model training is carried out to sample set by using cost sensitive learning method, in XGBoost algorithms, to difference The data of classification take different penalty coefficients；

S414, using introduce deep learning tree-model Dart, and the Dropout methods for introducing deep learning carry out processing prevent Model over-fitting；

S42, Model Fusion level, disaggregated model and regression model blend：Key Quality Indicator is predicted, by returning mould Type is predicted per a batch of product Key Quality Indicator, due to data nonbalance feature, is worked as to untreated few sample data Small classification is done, model prediction is carried out using disaggregated model, it is final to be carried out at data in such a way that classification and homing method share Reason；

S5, multi-model fusion treatment, specially：

S51, regression model fusion use average weighted method；

S52, disaggregated model, which merge, uses two two disaggregated models, after the completion of model training, predicts test set, obtains every A sample key_index is less than 0.92 or the predicted value of the high sample of confidence level is limited to 0.92 by the probability higher than 0.98 Or 0.98.

2. a kind of product quality control method based on integrated study according to claim 1, which is characterized in that described Product quality control method further includes the following steps：

R2, clear feature operational version carry out optimal preset value recommendation to technological parameter in production process, preferable to obtain Key Quality Indicator, it is specific as follows：

R21, whole technique adjustable parameters are recommended, specially：

It excavates so that yields maximum optimal parameter combination, is grouped parameter combination, obtains owning in training data The parameter combination of appearance, and calculate the mean value of yields of each combination, intermediate value, maximum value, minimum value and each be combined into Existing number, obtains statistical form, sorts from big to small according to yields mean value, to each of Top20/30/40 parameter combinations Adjustable parameter adds up, and finds out the mode value of each adjustable parameter as recommendation；

R22, parameter is cannot be adjusted for specific technique, technique adjustable parameter recommended, specially：

The product batches that yields is more than certain threshold value are filtered out in training data first, are made with the parameter of these product batches For candidate value；Then new product batches are found out therewith using the non-adjustable parameter list of technique as feature in candidate samples Most like or k most like samples of Top, take its adjustable parameter as recommendation.

3. a kind of product quality control method based on integrated study according to claim 2, which is characterized in that described Step R21, during whole technique adjustable parameters are recommended, for the adjustable parameter of double types, intermediate value or mean value conduct are taken Recommend.

4. a kind of product quality control method based on integrated study according to claim 2, which is characterized in that described Step R22, parameter is cannot be adjusted for specific technique, problem is converted to a phase in recommending technique adjustable parameter The problem of being measured like property uses different distance measure according to the type of numerical value, obtains the numerical value in the non-adjustable parameter of technique The importance of shape parameter carries out weight assignment, that is, completes weighted euclidean distance.

5. a kind of product quality control method based on integrated study according to claim 1, which is characterized in that described Molding proces s parameters include injection pressure, injection time, injection temperature, dwell pressure and time, back pressure, rotating speed.

6. a kind of product quality control method based on integrated study according to claim 1, which is characterized in that described In step S22, for unbalanced data feature, data plane, the mode for taking down-sampling to be combined；Algorithm level is chosen Boosting Ensemble Learning Algorithms are as basic algorithm model, judging quotas of the AUC as classification results.

7. a kind of product quality control method based on integrated study according to claim 1, which is characterized in that described Class variable processing and the processing of numeric type variable carry out coded treatment using one-hot encode to variable in step S23.

8. a kind of product quality control method based on integrated study according to claim 1, which is characterized in that described Regression model selects tri- kinds of models of XGBoost, DART, RandomForest, and common parameters are selected about the participation in the election of model tune：min_ Child_weight, for regression problem, it is number of samples minimum on each leafy node that the parameter is corresponding；

The disaggregated model selects whether two two disaggregated models, the key_index of one of forecast sample are less than 0.92, Whether the key_index of another forecast sample is higher than 0.98, and two models are with " binary:Logistic " is used as target letter Number, using AUC as evaluation index.

9. a kind of product quality control method based on integrated study according to claim 1, which is characterized in that described Random Forest model realizations are utilized in step S411.