Tool wear prediction method based on FIR-Stacking

Abstract

The invention relates to a tool wear prediction method based on FIR-Stacking. The method comprises the following steps: collecting signal data in a tool machining process; de-noising the signal data by using the optimized band elimination filter; expanding the denoised signal data to obtain expanded signal data; respectively extracting statistics from the time domain, the frequency domain and the time-frequency domain of the extended signal data; taking the statistics as features, performing feature extraction on single-dimensional data in the expanded signal data, and splicing features of multi-dimensional data to obtain a feature data set; and inputting the feature data set into an integrated learning model based on Stacking to obtain a tool wear result. According to the method, the self-adaptive FIR filter is adopted to remove signal noise, and the Stacking integration method of multiple machine learning models is combined, so that the accuracy of tool wear prediction can be effectively improved. And through feature extraction and selection of the tool monitoring signal, the degradation feature of tool wear is better captured, so that the prediction result is more accurate and reliable.

Description

Technical Field

[0001] This patent application relates to a tool wear prediction method based on FIR-Stacking, which belongs to the field of industrial manufacturing. Background Art

[0002] Cutting tool wear has an important impact on production efficiency and product quality in the field of industrial manufacturing. However, in actual use, cutting tools often suffer from wear, breakage, and falling off, resulting in reduced cutting quality, extended processing time, reduced production efficiency, and increased costs. Therefore, accurate prediction of tool wear is of great significance for improving production efficiency, reducing costs, and extending the service life of cutting tools.

[0003] At present, the tool wear prediction method based on data drive has become a hot topic of research. This method collects and analyzes the cutting tool monitoring signal data, and builds a prediction model based on signal feature engineering to predict the tool wear amount. However, the existing methods have the following problems:

[0004] 1. Sensor signal noise: During the cutting process, signal data is often disturbed by the processing environment, resulting in a lot of noise in the signal. This noise will affect the accuracy and reliability of the prediction model.

[0005] 2. The prediction effect of a single traditional machine learning model is not ideal: Current methods usually use a single traditional machine learning model to predict tool wear, but these models often find it difficult to fully exploit the characteristics of the data, resulting in limited prediction effects. Summary of the invention

[0006] In order to solve the above-mentioned shortcomings and difficulties, the present invention provides a tool wear prediction method based on FIR-Stacking.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a tool wear prediction method based on FIR-Stacking, comprising the following steps:

[0008] 1) Collect signal data during tool processing;

[0009] 2) constructing a band-stop filter, optimizing the parameters of the band-stop filter through a Bayesian optimization algorithm, and using the optimized band-stop filter to denoise the signal data;

[0010] 3) Expanding the denoised signal data to obtain expanded signal data;

[0011] 4) extracting statistics from the time domain, frequency domain and time-frequency domain of the extended signal data respectively;

[0012] 5) Taking the statistic as a feature, extracting features from the single-dimensional data in the extended signal data, and concatenating the features of the multi-dimensional data to obtain a feature data set;

[0013] 6) Construct a Stacking-based ensemble learning model, input the feature data set into the Stacking-based ensemble learning model, and obtain the tool wear results.

[0014] The signal data includes multi-dimensional data: cutting force signal, vibration signal and acoustic emission signal.

[0015] The constructing of the band-stop filter specifically includes: constructing the band-stop filter using a FIR filter based on a Kaiser window function.

[0016] The optimization of the band-stop filter parameters by using the Bayesian optimization algorithm comprises the following steps:

[0017] The Bayesian optimization algorithm is used to take the order of the band-stop filter, the cutoff frequency, the bandwidth and the beta value of the Kaiser window function as parameters. By setting the parameter space of the above parameters and defining the maximization of the signal-to-noise ratio as the objective function, the parameter space is searched to obtain the optimal filter parameters.

[0018] The signal data is expanded to obtain the expanded signal data, specifically: the denoised signal data is divided into N signal data with a length of L.

[0019] The statistics include mean, skewness, kurtosis, standard deviation, energy, root mean square, peak-to-peak value, shape factor, pulse factor and margin factor;

[0020] Among them, mean, standard deviation and energy are basic statistics, which are used to describe the central trend, fluctuation degree and signal strength of signal data; skewness and kurtosis are used to describe the asymmetry and peak state of signal data; root mean square and peak-to-peak value are used to reflect the amplitude and variation range of signal data; waveform factor, pulse factor and margin factor are used to characterize the shape characteristics and variation law of signal data.

[0021] The construction of the stacking-based integrated learning model includes the following steps:

[0022] The signal data of the historical tool processing process is processed through steps 1) to 5) to obtain a feature data set; the Pearson correlation coefficient is used to select features for the training set in the feature data set, and the features with correlation coefficients greater than a threshold are selected as inputs of the ensemble learning model;

[0023] The screened training set is divided into three parts and input into LightGBM, SVR and MLP for training and prediction respectively to obtain the predicted value of tool wear. These predicted values ​​are added to the training set as new features and finally handed over to the Lasso model for regression prediction to obtain the final integrated learning model composed of LightGBM, SVR, MLP and Lasso models.

[0024] A tool wear prediction system based on FIR-Stacking, comprising:

[0025] A data acquisition module, used to collect signal data during tool processing;

[0026] The data denoising module is used to construct a band-stop filter and optimize the parameters of the band-stop filter through the Bayesian optimization algorithm, and use the optimized band-stop filter to denoise the signal data;

[0027] A data expansion module is used to expand the denoised signal data to obtain expanded signal data; extract statistics from the time domain, frequency domain and time-frequency domain of the expanded signal data respectively;

[0028] A feature extraction module is used to extract features from single-dimensional data in the extended signal data using statistics as features, and to concatenate features of multi-dimensional data to obtain a feature data set;

[0029] The tool wear prediction module is used to build an integrated learning model based on Stacking, input the feature data set into the integrated learning model based on Stacking, and obtain the tool wear results.

[0030] A tool wear prediction device based on FIR-Stacking comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement the tool wear prediction method based on FIR-Stacking when executing the computer program.

[0031] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the tool wear prediction method based on FIR-Stacking is implemented.

[0032] The present invention has the following beneficial effects and advantages:

[0033] 1. High-accuracy prediction: Adaptive FIR filters are used to remove signal noise, and the stacking integration method combined with multiple machine learning models can effectively improve the accuracy of tool wear prediction. By extracting and selecting the features of tool monitoring signals, the degradation characteristics of tool wear can be better captured, making the prediction results more accurate and reliable.

[0034] 2. Strong noise resistance: The adaptive FIR filter can automatically adjust the filter parameters according to the actual signal conditions, effectively reducing the impact of noise in the processing environment on the prediction results. By denoising the signal, the model's ability to interpret the signal is improved, and the stability and reliability of the prediction are improved.

[0035] 3. Comprehensive feature extraction: Using common statistics in the time domain, frequency domain and time-frequency domain as signal features can comprehensively and multi-dimensionally describe the characteristics of tool wear. By combining different features and screening effective features, it can better capture the changing laws of tool wear and improve prediction accuracy.

[0036] 4. Multi-model integration: The stacking integration method of multiple machine learning models can combine the advantages of each model to improve the robustness and generalization of predictions. By integrating the prediction results of the base model and the meta-model, the complementarity between models can be better utilized to improve the accuracy and reliability of the prediction results.

[0037] 5. Strong practicability: The method of the present invention is based on existing machine learning algorithms and signal processing technology, and has high practicability and operability. It provides a simple, efficient and reliable solution for tool wear monitoring and prediction in the industrial manufacturing field, which can help enterprises reduce unnecessary production downtime and maintenance costs, and improve production efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a module flow chart of the method of the present invention;

[0039] Figure 2 This is the structure diagram of the Stacking integration model. DETAILED DESCRIPTION

[0040] The method of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0041] A tool wear prediction method based on FIR-Stacking includes the following steps:

[0042] 1. Use adaptive FIR filter to remove signal noise;

[0043] 2. Calculate common statistics in the time domain, frequency domain, and time-frequency domain as signal features, and splice multi-source signal features;

[0044] 3. Use Pearson correlation coefficient to screen important features;

[0045] 4. Using LightGBM, SVR and MLP as base models and Lasso as meta-model, a Stacking integrated model is constructed for the final tool wear prediction.

[0046] like Figure 1 As shown, the method of the present invention comprises the following steps:

[0047] 1. In order to predict the amount of tool wear, a total of 7 channels of signals, including cutting force signals (3 channels, along the X, Y, and Z directions), vibration signals (3 channels, along the X, Y, and Z directions) and acoustic emission signals (1 channel) during tool processing, are collected, and the tool wear is measured offline after each round of tool processing is completed. The wear is used as a label for training the final model. The present invention predicts the amount of tool wear with 7-channel signal data. During tool processing, the feed and exit operations will cause fluctuations in signal data, which are regarded as outliers and need to be eliminated. Through observation and analysis, it is decided to eliminate outliers in the first 10% and last 10% stages of the signal data. The retained signal data has stability and continuity, and is more suitable for subsequent experimental research. In addition, considering that the signal lengths generated by different processing rounds are different, a signal with an intercept length of 100,000 is selected for subsequent experiments. These processing measures will ensure the accuracy and reliability of the experiment.

[0048] 2. For the signal data after outliers are removed, a FIR filter based on the Kaiser window function is used to design a band-stop filter to remove noise from the signal data. In order to remove the noise band as much as possible while retaining the original information, this study uses the Bayesian optimization algorithm to optimize the filter parameters. The Bayesian optimization algorithm uses the order, cutoff frequency, bandwidth and beta value of the Kaiser window function as parameters. By setting the parameter space of these parameters and defining the maximum signal-to-noise ratio as the objective function, the parameter space is searched to obtain a set of optimal filter parameters. The filter designed in this way can more effectively remove noise and improve signal quality.

[0049] 3. Use data enhancement technology to expand the data set to avoid problems such as model overfitting due to the limited number of samples. Specifically, the denoised signal data is divided into 10 parts, that is, the signal data with a length of 100,000 is divided into 10 signal data with a length of 10,000. In this way, the sample size of a single tool is expanded from 315 to 3150, thereby increasing the diversity and quantity of the data. This strategy further improves the accuracy and robustness of the model.

[0050] 4. Use Fourier transform to analyze the frequency domain information of the signal, and use short-time Fourier transform to obtain the time-frequency domain information of the signal. Ten common statistical quantities are extracted from the time domain, frequency domain and time-frequency domain of the signal, including mean, skewness, kurtosis, standard deviation, energy, root mean square, peak-to-peak value, waveform factor, pulse factor and margin factor.

[0051] These statistics reflect the characteristics of signal data in different domains and play an important role in the analysis and prediction of tool wear. Mean, standard deviation and energy are basic statistics used to describe the central trend, fluctuation degree and signal strength of the signal; skewness and kurtosis describe the asymmetry and peak state of the signal data; root mean square and peak-to-peak value reflect the amplitude and variation range of the signal; waveform factor, pulse factor and margin factor reveal the shape characteristics and variation law of the signal data.

[0052] By extracting features from single-dimensional signals, we obtained 30 features in the time domain, frequency domain, and time-frequency domain. Then, we spliced ​​the 7-dimensional signal features to achieve feature fusion, and obtained a total of 210 features.

[0053] 5. Randomly divide the data set obtained in the previous step into training set and test set, with the training set accounting for 80% and the test set accounting for 20%. Use the Pearson correlation coefficient to select features for the training set, and select features with a correlation coefficient greater than 0.2 as the input of the final model. Because the Pearson correlation coefficient is used to measure the degree of linear correlation between two variables, in this step, the Pearson correlation coefficient of each feature and the target variable (wear amount) is calculated, and the features with a correlation coefficient greater than 0.2 are considered to be important features and need to be retained. This treatment measure reduces the number of data features as much as possible to reduce the computational burden and improve model performance.

[0054] 6. Using the stacking-based ensemble learning model, the base model selects three strong regression and differentiated models in tool wear prediction: LightGBM, SVR and MLP, so that the ensemble model can learn the advantages of the above three models, and the meta-model selects a simple regression model: Lasso regression, so as to avoid model overfitting. Figure 2 As shown in the figure, the training set obtained in the previous step is divided into three parts, which are respectively trained and predicted by LightGBM, SVR and MLP to obtain the predicted values ​​of tool wear. These predicted values ​​are added to the training set as new features, and finally the Lasso model is used for regression prediction. During the model training process, a set of optimal model parameters is obtained by minimizing the prediction error of the model. The final model can be obtained by setting this set of model parameters. The final model is used to predict the test set to verify the accuracy of the model prediction.

[0055] The feature input obtained from step 1 to step 5 is used to predict the tool wear results through the final model. That is, the 7-channel signal data generated by each round of tool processing is characterized by steps 1 to 5, and the tool wear amount is obtained after inputting the model.

[0056] The tool wear prediction method of the present invention can not only improve the performance and efficiency of the model, but also adaptively remove the signal noise generated in different industrial environments. At the same time, since the model integrates three different models: LightGBM, SVR and MLP, the final model generalization ability will be better, thereby further improving the accuracy of tool wear prediction.

[0057] Through the comprehensive application of the above-mentioned processing technologies, the tool wear prediction method of the present invention can effectively remove the interference of signal noise generated in industrial environments on tool wear prediction, and the use of an integrated model improves the generalization of the model and improves the tool wear prediction effect, thereby having better performance and application prospects in industrial manufacturing tasks such as tool wear prediction and tool life prediction.

Claims

1. A tool wear prediction method based on FIR-Stacking, characterized in that: The following steps are involved: 1) Collect signal data during tool processing; 2) constructing a band-stop filter, optimizing the parameters of the band-stop filter through a Bayesian optimization algorithm, and using the optimized band-stop filter to denoise the signal data; 3) Expanding the denoised signal data to obtain expanded signal data; 4) extracting statistics from the time domain, frequency domain and time-frequency domain of the extended signal data respectively; 5) Taking the statistic as a feature, extracting features from the single-dimensional data in the extended signal data, and concatenating the features of the multi-dimensional data to obtain a feature data set; 6) Construct a Stacking-based ensemble learning model, input the feature data set into the Stacking-based ensemble learning model, and obtain the tool wear results.

2. The tool wear prediction method based on FIR-Stacking according to claim 1, characterized in that: The signal data includes multi-dimensional data: cutting force signal, vibration signal and acoustic emission signal.

3. The tool wear prediction method based on FIR-Stacking according to claim 1, characterized in that: The constructing of the band-stop filter specifically includes: constructing the band-stop filter using a FIR filter based on a Kaiser window function.

4. The tool wear prediction method based on FIR-Stacking according to claim 1, characterized in that: The optimization of the band-stop filter parameters by using the Bayesian optimization algorithm comprises the following steps: The Bayesian optimization algorithm is used to take the order of the band-stop filter, the cutoff frequency, the bandwidth and the beta value of the Kaiser window function as parameters. By setting the parameter space of the above parameters and defining the maximization of the signal-to-noise ratio as the objective function, the parameter space is searched to obtain the optimal filter parameters.

5. The tool wear prediction method based on FIR-Stacking according to claim 1, characterized in that: The expanding the denoised signal data to obtain the expanded signal data specifically includes: dividing the denoised signal data into N signal data of length L.

6. The tool wear prediction method based on FIR-Stacking according to claim 1, characterized in that: The statistics include mean, skewness, kurtosis, standard deviation, energy, root mean square, peak-to-peak value, shape factor, pulse factor and margin factor; Among them, mean, standard deviation and energy are basic statistics, which are used to describe the central trend, fluctuation degree and signal strength of signal data; skewness and kurtosis are used to describe the asymmetry and peak state of signal data; root mean square and peak-to-peak value are used to reflect the amplitude and variation range of signal data; waveform factor, pulse factor and margin factor are used to characterize the shape characteristics and variation law of signal data.

7. The tool wear prediction method based on FIR-Stacking according to claim 1, characterized in that: The construction of the stacking-based integrated learning model includes the following steps: The signal data of the historical tool processing process is processed through steps 1) to 5) to obtain a feature data set; the Pearson correlation coefficient is used to select features for the training set in the feature data set, and the features with correlation coefficients greater than a threshold are selected as inputs of the ensemble learning model; The screened training set is divided into three parts and input into LightGBM, SVR and MLP for training and prediction respectively to obtain the predicted value of tool wear. These predicted values ​​are added to the training set as new features and finally handed over to the Lasso model for regression prediction to obtain the final integrated learning model composed of LightGBM, SVR, MLP and Lasso models.

8. A tool wear prediction system based on FIR-Stacking, characterized in that: include: A data acquisition module, used to collect signal data during tool processing; The data denoising module is used to construct a band-stop filter and optimize the parameters of the band-stop filter through the Bayesian optimization algorithm, and use the optimized band-stop filter to denoise the signal data; A data expansion module is used to expand the denoised signal data to obtain expanded signal data; extract statistics from the time domain, frequency domain and time-frequency domain of the expanded signal data respectively; A feature extraction module is used to extract features from single-dimensional data in the extended signal data using statistics as features, and to concatenate features of multi-dimensional data to obtain a feature data set; The tool wear prediction module is used to build an integrated learning model based on Stacking, input the feature data set into the integrated learning model based on Stacking, and obtain the tool wear results.

9. A tool wear prediction device based on FIR-Stacking, characterized in that: It comprises a memory and a processor; the memory is used to store a computer program; the processor is used to implement a tool wear prediction method based on FIR-Stacking as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by the processor, the tool wear prediction method based on FIR-Stacking as described in any one of claims 1 to 7 is implemented.

Images