CN111475921A - A Tool Remaining Life Prediction Method Based on Edge Computing and LSTM Network - Google Patents

Abstract

The invention relates to a tool residual life prediction method based on edge computing and LSTM network, and belongs to the field of tool life prediction of numerical control machine tools. The method includes: first, data cleaning and feature extraction are performed directly at the edge to reduce transmission time, save transmission costs, and improve the real-time performance of life prediction; then, after further feature extraction and selection in the cloud, establish a three-layer LSTM recurrent neural network The network model is used to predict the real-time remaining life of the tool. The invention improves the real-time performance and accuracy of tool life prediction by using edge computing and LSTM methods.

Description

A Tool Remaining Life Prediction Method Based on Edge Computing and LSTM Network

technical field

The invention belongs to the field of tool life prediction of numerically controlled machine tools, and relates to a method for predicting the remaining life of a tool based on edge computing and LSTM network.

Background technique

As an important tool in the industrial manufacturing process, the tool life and wear state affect the production quality of the workpiece, the production efficiency and the health of the lathe. If the remaining life of the tool can be accurately predicted, it will effectively reduce the cost of industrial manufacturing.

Experience-based, physics- and data-driven models have been used for tool remaining life prediction over the years, respectively. Among them, research based on data-driven methods has become more and more mature with the development of big data and artificial intelligence. Zhuang et al. used the nonlinear mapping characteristics of neural network to establish a tool failure prediction model for shield machine based on ACO-BP algorithm. Sun et al. proposed a deep transfer learning network based on sparse autoencoders for tool life prediction. Zhao et al. proposed a gated regression unit network based on local features to predict the tool wear state in machine health. Drouille et al. used neural network (NN) technology to predict the remaining useful life (RUL) of the tool based on the power value of the machine tool spindle.

In cloud services, users can obtain high-quality services at lower costs, and can flexibly control computing resources. However, cloud computing has limitations and bottlenecks arise when large amounts of data are transferred to a single cloud center and processed there. If all data is processed in the cloud, it will cause data delay and affect real-time performance. For the real-time monitoring of the working state of the tool, such a problem is a fatal flaw.

Based on the defects of the prior art, the present invention proposes a method for predicting the remaining life of a tool by combining edge data in an industrial physical network and an LSTM network, so as to solve the above-mentioned technical defects.

SUMMARY OF THE INVENTION

In view of this, the purpose of the present invention is to provide a general prediction model suitable for key equipment milling tools, by analyzing the current tool state monitoring technology, selecting the tool indirect measurement index with ideal effect, using data cleaning, feature extraction and The LSTM cyclic neural network method establishes a tool remaining life prediction model, and provides a method for predicting the remaining tool life of CNC machine tools.

To achieve the above object, the present invention provides the following technical solutions:

A tool residual life prediction method based on edge computing and LSTM network, which specifically includes the following steps:

S1: Collect PLC controller signals and external sensor signals to obtain sensing data;

S2: Data cleaning and feature extraction are performed at the edge of the data;

S3: transfer the feature data obtained in step S2 to the cloud, and perform feature extraction and feature screening again;

S4: Normalize the data obtained in step S3, construct the training data of the LSTM neural network, establish the LSTM cyclic neural network, and perform real-time prediction of the remaining life of the tool.

Further, in step S1, the sensor data are mainly current signals and vibration signals in three directions, that is, the x-axis, the y-axis and the z-axis; the PLC controller signal is mainly the load of tool machining.

Further, in step S2, the data cleaning is to first fill in the missing values with the mean value, then use the boxplot to remove the outliers, and finally perform moving average filtering on the data; the extracted time domain features mainly include mean value, root mean square value, variance , slope, kurtosis, crest factor and margin factor, etc.

Further, in the step S3, feature extraction and feature screening are performed again, which specifically includes the following steps:

S31: Perform clustering per second, extract statistical features such as mean, maximum, minimum, median, three-quarter digit, quarter digit, and variance at the time, and perform more detailed data on the data description of;

S32: Send the extracted feature data into the LightGBM model to obtain features whose feature importance is greater than zero; first, use Bayesian optimization to adjust the parameters of the lightGBM model to obtain the optimal parameters; then use the five-fold cross-validation method to train The model obtains the importance of each feature; finally, feature screening is performed to obtain the features required by the final model.

Further, in the step S4, establishing an LSTM cyclic neural network specifically includes the following steps:

S41: Construct the input of the LSTM recurrent neural network, and use the feature information of the past 20 seconds to predict the remaining life of the next second;

S42: Build LSTM network: build a three-layer LSTM network structure, and use the grid search method to search for the optimal hyperparameters. The dimensions of the two hidden layers are set to 256 and 256, respectively, and the activation function selects the PRelu function; in addition, in RMSProp , Adam optimizer is selected from Adam and Adadelta optimizer; In addition, in order to solve the problem of easy overfitting of neural network, Dropout layer and Batch normalization layer are added, and the parameter of Dropout layer is set to 0.5; In training, the mean absolute value error loss function is selected as The training error is calculated, and the method of early stopping is used to obtain the optimal model. The number of early stopping steps is set to 20 steps, that is, when the loss does not decrease in 20 trainings, the training will be stopped.

The beneficial effects of the present invention are: the present invention performs data cleaning and feature extraction at the edge end, reduces the transmission time, saves the transmission cost, and improves the real-time performance of life prediction. The invention also establishes an LSTM cyclic neural network model in the cloud, which can better predict and monitor the remaining life of the tool in real time.

Other advantages, objects, and features of the present invention will be set forth in the description that follows, and will be apparent to those skilled in the art based on a study of the following, to the extent that is taught in the practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the following description.

Description of drawings

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be preferably described in detail below with reference to the accompanying drawings, wherein:

1 is a schematic diagram of a model framework for predicting the remaining service life of a CNC machine tool tool according to the present invention;

FIG. 2 is a data cleaning diagram of a model edge end according to an embodiment of the present invention;

FIG. 3 is a comparison diagram of the prediction curve of the remaining life value of the tool according to the embodiment of the present invention.

Detailed ways

The embodiments of the present invention are described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the drawings provided in the following embodiments are only used to illustrate the basic idea of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

Please refer to Fig. 1 to Fig. 3. Fig. 1 is a preferred method of the present invention for predicting the remaining service life of a CNC machine tool tool. As shown in Fig. 1, first, the controller (PLC) signal and the external sensor (Sensor) are collected according to the CPS framework. Signal, collect working condition information and sensor data during processing, sensor data is mainly current signal, sensor data is mainly current signal and vibration signal in three directions, namely x-axis, y-axis and z-axis. Then, the data cleaning and time domain feature extraction are performed directly at the edge of the data. The data cleaning first uses the mean to fill in the missing values, then uses the boxplot to remove the outliers, and then performs moving average filtering on the data, and the extracted time domain Features mainly include mean, root mean square, variance, slope, kurtosis, crest factor and margin factor. The current signal is compared before and after cleaning, as shown in Figure 2. The obtained feature data is transmitted to the cloud, and the cloud builds the LSTM model. The steps are as follows:

1) Cluster by second, extract statistical features such as mean, maximum, minimum, median, three-quarter digit, quarter digit, and variance in the time, and update the data. detailed description;

2) The extracted feature data is sent to the LightGBM model to obtain features whose feature importance is greater than zero. First, the lightGBM model is adjusted by Bayesian optimization to obtain the optimal parameters, and then the five-fold cross-validation method is used for training. The model obtains the importance of each feature, and finally performs feature screening to obtain the features required by the final model;

3) Construct the input of the LSTM recurrent neural network, and use the feature information of the past 20s to predict the remaining life of the next 1 second.

4) A three-layer LSTM network structure is constructed, and the grid search method is used to search for the optimal hyperparameters. A three-layer LSTM network structure is constructed, and the grid search method is used to search for the optimal hyperparameters. The dimensions of the two hidden layers are set to 256 and 256 respectively, and the PRelu function is selected as the activation function. Also, Adam optimizer is chosen among RMSProp, Adam and Adadelta optimizers. In addition, in order to solve the problem of easy overfitting of the neural network, the Dropout layer and the Batch normalization layer are added. The parameter of the dropout layer is set to 0.5. In the training, the average absolute value error loss function is selected as the training error, and the method of early stopping is used to obtain the optimal model. The number of early stopping steps is set to 20 steps, that is, when the loss does not decrease in 20 trainings, the training is stopped. .

In order to verify the feasibility and accuracy of the method, the following test experiments are carried out, and this model is compared with some commonly used machine learning models. The data source is the actual CNC machining process, a brand new tool starts the normal machining program, and stops the data collection when the tool life ends. In terms of data sampling frequency, the sampling frequency of PLC signal is 33Hz, and the sampling frequency of vibration sensor is 25600Hz. Three sets of data are selected for prediction, and the prediction result curve is shown in Figure 3. The model of the present invention is compared with the prediction using the lightgbm model and the ordinary neural network model. The calculation model calculates the mean absolute value error (mean_absolute_error) between the proportion of remaining tool life and the actual proportion of remaining life of the tool. The results are shown in Table 1.

Table 1 Error result comparison table

Model the first the second the third average value lightgbm 0.033 0.092 0.045 0.057 NN 0.016 0.106 0.076 0.066 LSTM 0.025 0.066 0.032 0.041

Combining Figure 3 and Table 1, it can be seen that the remaining life value prediction effect of the three knives at multiple times can be maintained at a good level by LSTM, which is better than the lightgbm model and the ordinary neural network.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be Modifications or equivalent replacements, without departing from the spirit and scope of the technical solution, should all be included in the scope of the claims of the present invention.

Claims (5)

1. a tool residual life prediction method based on edge computing and LSTM network, is characterized in that, this method specifically comprises the following steps: S1: Collect PLC controller signals and external sensor signals to obtain sensing data; S2: Data cleaning and feature extraction are performed at the edge of the data; S3: transfer the feature data obtained in step S2 to the cloud, and perform feature extraction and feature screening again; S4: Normalize the data obtained in step S3, construct the training data of the LSTM neural network, establish the LSTM cyclic neural network, and perform real-time prediction of the remaining life of the tool. 2. a kind of tool remaining life prediction method based on edge computing and LSTM network according to claim 1, is characterized in that, in step S1, described sensor data is current signal and three directions namely x-axis, y-axis and The vibration signal of the z-axis; the PLC controller signal is the load of the tool processing. 3. a kind of tool remaining life prediction method based on edge computing and LSTM network according to claim 1, is characterized in that, in step S2, described data cleaning is to first use mean to fill missing values, then use boxplot to remove. Outliers, and finally perform moving average filtering on the data; the extracted time domain features include mean, root mean square, variance, slope, kurtosis, crest factor and margin factor. 4. a kind of tool residual life prediction method based on edge computing and LSTM network according to claim 1, is characterized in that, in described step S3, carry out feature extraction and feature screening again, specifically comprises the following steps: S31: perform clustering per second, and extract the mean, maximum, minimum, median, three-quarter digit, quarter-digit, and variance in the time; S32: Send the extracted feature data into the LightGBM model to obtain features whose feature importance is greater than zero; first, use Bayesian optimization to adjust the parameters of the lightGBM model to obtain the optimal parameters; then use the five-fold cross-validation method to train The model obtains the importance of each feature; finally, feature screening is performed to obtain the features required by the final model. 5. a kind of tool remaining life prediction method based on edge computing and LSTM network according to claim 1, is characterized in that, in described step S4, establish LSTM cyclic neural network, specifically comprises the following steps: S41: Construct the input of the LSTM recurrent neural network, and use the feature information of the past 20 seconds to predict the remaining life of the next second; S42: Build LSTM network: Build a three-layer LSTM network structure, and use the grid search method to search for optimal hyperparameters. The dimensions of the two hidden layers are set to 256 and 256, respectively, and the activation function selects the PRelu function; select Adam optimizer , adding the Dropout layer and the Batch normalization layer, the parameter of the Dropout layer is set to 0.5; in the training, the average absolute value error loss function is selected as the training error, and the method of early stopping is used to obtain the optimal model, and the number of early stopping steps is set as Step 20 means that the training is stopped when the error of loss no longer decreases in 20 training sessions.

Images