Multi-feature fusion numerical control machine tool cutter wear prediction method

Abstract

The invention discloses a multi-feature fusion numerical control machine tool cutter wear prediction method. The method comprises the following steps: firstly, preprocessing original data by adopting a third quartile and a Hampel filtering method; secondly, carrying out time domain, frequency domain, time-frequency domain and other analysis on the preprocessed data, and extracting a plurality of transform domain features which are well related to the wear degree of the cutter; secondly, appropriate machine learning models are selected for different features, a support vector machine (SVM) is adopted to process time domain features, a Random Forest is adopted to process frequency domain features, and a deep neural network (DNN) is adopted to process time domain and frequency domain combined features; and finally, evaluating and predicting the performance of the model by adopting indexes such as a root mean square error (RMSE) and a mean absolute error (MAE) of the three models, and verifying the effectiveness of the experimental method. The tool wear state can be timely and accurately predicted and diagnosed, and real-time monitoring of the tool wear state under the normal production condition is facilitated.

Description

Technical Field

[0001] The invention belongs to the field of tool wear prediction, and in particular relates to a multi-feature fusion method for predicting tool wear of CNC machine tools. Background Art

[0002] As one of the core equipments of modern manufacturing industry, CNC machine tools are widely used in various production environments due to their high degree of automation and high processing precision. The tool is a key functional component in the processing of CNC machine tools, and its status directly affects the processing quality and production efficiency of products. Therefore, timely and accurate diagnosis and prediction of tool wear failures are crucial to maintaining the reliability of CNC machine tools. Accurate prediction of tool wear has far-reaching significance for improving the production efficiency and reducing the production cost of enterprises.

[0003] In recent years, signal processing technology has become increasingly mature, and sensor types have become more and more complete. It is possible to select a variety of sensors to monitor tool wear status. However, in the industrial production process, due to system errors, sensor errors, and human factors, the collected data is often accompanied by noise, missing values, and even data inconsistencies, which seriously affect the accuracy and reliability of the model. Data preprocessing is required before modeling, which helps to reduce the impact of noise, correct data inconsistencies, identify and delete outliers, thereby improving the robustness of the model and preventing the model from overfitting.

[0004] In traditional machining processes, sensor monitoring data is time series data without a fixed pattern, which is not conducive to identifying the relationship between data features and wear degree. Therefore, feature extraction is required to map low-dimensional time series data to high-dimensional representations, and discover and construct new input data features and output data features that are helpful for model fitting. The feature extraction process aims to map low-dimensional time series data to a higher-dimensional feature space, and use machine learning models to process industrial data in higher-dimensional feature spaces. Machine learning models can identify the correlation between different features and select the features that are most closely related to tool wear. Machine learning can also model nonlinear features and has good adaptability to the variability and complexity of industrial environments. Summary of the invention

[0005] In order to accurately predict the tool wear during machining, a multi-feature fusion method for tool wear prediction of CNC machine tools was proposed.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A multi-feature fusion CNC machine tool tool wear prediction method performs the following steps to obtain a multi-feature fusion tool wear prediction model, including the following steps:

[0008] 1) Collect the original data set, trim outliers, and replace data according to the median deviation;

[0009] 2) Extract multiple features of the data in the time and frequency domains, and map the data set from low dimension to high dimension to reflect the data information of the cutting process;

[0010] 3) Establish a fusion prediction model and use different machine learning models for iterative training according to different feature data to accurately predict tool wear.

[0011] The original data set is the tool wear data of CNC machine tools, including: reading the data of 7 channels collected by sensors in the PHM2010 Industrial Big Data Competition data set: vibration signals in the three directions of x-axis, y-axis and z-axis, cutting force signals in the three directions of x-axis, y-axis and z-axis, and sound signals.

[0012] The trimming of outliers adopts the third quartile method, including:

[0013] For data sets with outliers, the upper quartile Q3 of the collected tool wear data is calculated, and the upper quartile is used as the critical value of invalid data to cut off the outliers.

[0014] The data replacement according to the median deviation adopts the Hampel filtering method, including: calculating the median and MAD of the data by setting the number of samples, the number of windows and the upper and lower coefficients. If the sample value is greater than the upper and lower coefficients, it is regarded as an outlier and replaced by the median.

[0015] The time and frequency domain features include:

[0016] a. Extract time domain features based on statistics;

[0017] b. Convert the time domain signal into the frequency domain signal and extract the frequency domain features based on the amplitude;

[0018] c. Combine the time domain and frequency domain, and extract joint features of the time and frequency domains based on wavelet transform.

[0019] The time domain features include: root mean square, variance, maximum and minimum values, skewness, peak-to-peak value and kurtosis.

[0020] The frequency domain characteristics include: spectrum skewness coefficient, spectrum peak coefficient and spectrum power.

[0021] The three-way parallel network structure of the fusion prediction model includes:

[0022] a. Based on statistical time domain feature data, the support vector machine (SVM) model is used for training;

[0023] b. For frequency domain feature data, the random forest method is used for training;

[0024] c. For the time-frequency joint domain feature data, a deep neural network is used for training.

[0025] The method further includes evaluating the tool wear prediction model based on multi-feature fusion:

[0026] For different tool type data sets, time domain features are extracted respectively and input into three parallel network structures in the tool wear prediction model of multi-feature fusion for prediction, and RMSE and MAE values ​​are used as evaluation indicators.

[0027] The present invention has the following beneficial effects and advantages:

[0028] 1. The data preprocessing adopted the third quartile method and Hampel filtering method to ensure the quality and continuity of the data and provide a reliable data basis for feature extraction and prediction model construction.

[0029] 2. Use statistics to extract features in the time domain, amplitude spectrum to extract features in the frequency domain, and wavelet transform to extract features in the joint time-frequency domain. Map the data from low dimensions to high dimensions to reflect more cutting process data information.

[0030] 3. Time-frequency joint domain feature extraction shows the best effect in tool wear prediction. Feature extraction based on wavelet packet energy can more comprehensively capture the changes in time domain and frequency domain information during tool wear. This invention improves the quality of tool prediction data sets and accurately predicts and diagnoses tool wear status.

[0031] 4. Use multiple prediction models such as support vector machines, random forests and deep neural networks to adjust and combine multiple tool wear features to find the best fit, effectively improving the prediction performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of the present invention.

[0033] Figure 2 It is a time series signal of 7 channels.

[0034] Figure 3 The data after preprocessing.

[0035] Figure 4 is the relationship between time domain characteristics and wear amount.

[0036] Figure 5 is the amplitude spectrum.

[0037] Figure 6 is the relationship between frequency domain characteristics and wear amount.

[0038] Figure 7 is the relationship between wavelet packet energy characteristics and wear amount.

[0039] Figure 8a , Figure 8b , Figure 8c This is the prediction effect of SVM in time domain features.

[0040] Figure 9a , Figure 9b , Fig.9c This is the prediction effect of random forest in frequency domain features.

[0041] Fig.10a , Fig.10b , Fig.10c This is the prediction effect of DNN in the joint time-frequency domain features. DETAILED DESCRIPTION

[0042] The following is combined with Figure 1 The method of the present invention is described in further detail.

[0043] The present invention firstly uses the third quartile and Hampel filtering method to preprocess the raw data. Secondly, the preprocessed data is analyzed in the time domain, frequency domain, time-frequency domain, etc., and multiple transformation domain features with good correlation with the degree of tool wear are extracted. Then, a suitable machine learning model is selected for different features, and a support vector machine is used to process the time domain features, a random forest is used to process the frequency domain features, and a deep neural network is used to process the joint features of the time and frequency domains. Finally, the root mean square error (RMSE) and mean absolute error (MAE) of the above three models are selected to evaluate the performance of the prediction model, verifying the effectiveness of the experimental method.

[0044] A multi-feature fusion CNC machine tool tool wear prediction method includes the following steps:

[0045] Step 1: Read the PHM2010 Industrial Big Data Competition data set. The 7 channels of data collected by the sensor include vibration signals in the three directions of x-axis, y-axis, and z-axis, cutting force signals in the three directions of x-axis, y-axis, and z-axis, and sound signals. The training data sets are c1, c4, and c6.

[0046] like Figure 2 The time series signals of 7 channels are shown. From top to bottom, they are x-axis force signal, y-axis force signal, z-axis force signal, x-axis vibration signal, y-axis vibration signal, z-axis vibration signal and acoustic emission signal.

[0047] Step 2: Preprocess the raw data using the third quartile and Hampel filtering methods.

[0048] Step 3: Extract useful information from the preprocessed data for subsequent modeling and analysis. Map the data from low dimension to high dimension to reflect more cutting process data information.

[0049] Step 4: Use different machine learning models for different features to accurately predict tool wear.

[0050] The step 2 comprises:

[0051] Step 2.1: The third quartile method is applicable to data sets with outliers. The way to identify outliers is to calculate the upper quartile (Q3) of the collected tool wear data and use the upper quartile as the critical value of invalid data, thereby trimming outliers from the tool wear data and improving the robustness of the model.

[0052] Step 2.2: Hampel filtering method reduces the impact of outliers on the data while retaining the overall trend of the data. This method is based on the median and median absolute deviation. By setting the number of samples, the number of windows, and the upper and lower coefficients, the median and MAD of the data are calculated. If the sample value is greater than the upper and lower coefficients, it is regarded as an outlier and is replaced by the median.

[0053] like Figure 3 The data after preprocessing is shown in Figure 2. The relatively clean cutting process data is obtained, laying the foundation for the next step of data feature extraction and model prediction.

[0054] The step 3 comprises:

[0055] Step 3.1: Extract time domain features based on statistics. Extract statistical features from the time domain information of the time series data, and calculate the root mean square, variance, maximum and minimum values, skewness, peak-to-peak value and kurtosis to describe the basic statistical properties of the data.

[0056] like Figure 4 The figure shows the relationship between the time domain characteristics and the wear amount. : The relationship between the root mean square, variance, maximum value, skewness, kurtosis, peak-to-peak value and the wear amount.

[0057] Step 3.2: Extract frequency domain features based on amplitude, convert the time domain signal into a frequency domain signal, and extract the spectral skewness coefficient, spectral peak coefficient, and spectral power from the amplitude.

[0058] like Figure 5 The figure shows the amplitude spectrum. It can be seen from the figure that the amplitude spectrum can be processed to obtain the spectrum skewness coefficient, spectrum peak coefficient and spectrum power.

[0059] like Figure 6 The figure shows the relationship between frequency domain characteristics and wear loss. The figure shows the relationship between spectral skewness coefficient, spectral peak coefficient and spectral power and wear loss.

[0060] Step 3.3: Extract the joint features of time and frequency domains based on wavelet transform, combine the time domain and frequency domain, and extract the joint features of time and frequency domains from the wavelet packet energy.

[0061] like Figure 7 The figure shows the relationship between wavelet packet energy characteristics and wear. The wavelet function uses the Dobesi wavelet function, the decomposition layer number N = 8, and the number of wavelet packets in the last layer is 256.

[0062] The time domain feature calculation method in step 3.1 is:

[0063] Root mean square, the effective value of the signal in the reaction data:

[0064] Variance reflects the degree of dispersion of the data set:

[0065] Maximum and minimum values ​​reflect the highest and lowest points of the data within a specific time:

[0066] Skewness reflects the degree of skewness of the data, including the direction and magnitude of the skewness:

[0067] Peak-to-peak value, reflecting the range of data variation: PP = max(X)-min(X);

[0068] Kurtosis reflects the sharpness of the data relative to the normal distribution:

[0069] The frequency domain feature calculation method in step 3.2 is:

[0070] The formula for Fourier transform is:

[0071] The amplitude of the signal |F(ω)| can be expressed as:

[0072] Signal phase It can be expressed as:

[0073] Spectral skewness coefficient: the degree of skewness of the response amplitude distribution:

[0074] Spectral peak coefficient: the sharpness of the amplitude waveform:

[0075] Spectral power: reflects the signal power within a unit frequency band:

[0076] The step 4 comprises:

[0077] The different machine learning models include:

[0078] a. Support vector machine processes time domain features. Support vector machine is often used for classification and regression tasks. Its principle is to determine the hyperplane that can effectively separate data points of different categories based on statistical learning theory and convex optimization, while maximizing the interval between data points of different categories. SVM is usually used to process statistical time domain features.

[0079] b. Random forest processes frequency domain features. Random forest is commonly used for classification and regression tasks. It is an ensemble learning method that makes predictions based on multiple decision trees and integrates the results. Each decision tree is constructed in the same way, randomly selecting data subsets and feature subsets, and the prediction results are ultimately determined by voting or averaging. Random forest can be used to process frequency domain features.

[0080] c. Deep neural networks process time-frequency domain joint features. Deep neural networks are often used to simulate and solve complex nonlinear modeling tasks. They are a machine learning model based on an artificial neuron hierarchy. Deep neural networks contain multiple hidden layers, and control the transmission and calculation of information between neurons through weights and activation functions. The core idea of ​​deep neural networks is to adjust weights through forward propagation and back propagation to reduce prediction errors. Deep neural networks can be used to process time-frequency domain joint features.

[0081] Step 4.1: For each tool type (C1, C4, C6), extract a set of time domain features and use the support vector machine as a prediction model for experiments. Calculate the RMSE and MAE values ​​of the support vector machine model.

[0082] Step 4.2: For each tool type (C1, C4, C6), extract a set of frequency domain features and use random forest as a prediction model for experiments. Calculate the RMSE and MAE values ​​of the random forest model.

[0083] Step 4.3: For each tool type (C1, C4, C6), extract the joint features in the time and frequency domains and use the deep neural network as the prediction model for experiments. Calculate the RMSE and MAE values ​​of the deep neural network.

[0084] like Figure 8a , Figure 8b , Figure 8c The figure shows the prediction effect of SVM in time domain features. The experimental results show that the RMSE of C1 is 11.2137 and the MAE is 5.6425; the RMSE of C4 is 12.0537 and the MAE is 6.6083; the RMSE of C6 is 9.9263 and the MAE is 6.1389. This shows that the time domain features have a good prediction effect on C1 and C6 tools, but are slightly inferior in the prediction of C4, probably because the wear behavior of C4 is more complex and it is difficult to capture its changes only through time domain features.

[0085] like Figure 9a , Figure 9b , Fig.9c The figure shows the prediction effect of random forest in frequency domain features. The experimental results show that the RMSE of C1 is 8.1544 and the MAE is 5.4815; the RMSE of C4 is 11.9296 and the MAE is 6.7128; the RMSE of C6 is 8.1638 and the MAE is 4.4791. The frequency domain features perform well in the prediction of C1 and C6, especially the prediction error of C6 is relatively low, which may be because the frequency domain features can better capture the spectral characteristics of its wear behavior. However, the prediction effect of frequency domain features for C4 is poor, which indicates that more complex feature engineering methods may be needed to deal with its complex wear pattern.

[0086] like Fig.10a , Fig.10b , Fig.10c The figure shows the prediction effect of DNN in the joint time-frequency domain features. The experimental results show that the RMSE of C1 is 6.7089 and the MAE is 4.7856; the RMSE of C4 is 7.4001 and the MAE is 5.5078; the RMSE of C6 is 5.1465 and the MAE is 3.1343. These results show that the joint time-frequency domain features based on wavelet packet energy perform well in the prediction of C1 and C6, but there is still a certain error in the prediction of C4. The use of DNN models may help to better capture the complexity of the joint time-frequency domain features, but for different tool types, the network structure needs to be adjusted to further improve the performance.

Claims

1. A multi-feature fusion CNC machine tool tool wear prediction method, characterized in that: Perform the following steps to obtain a multi-feature fusion tool wear prediction model, including the following steps: 1) Collect the original data set, trim outliers, and replace data according to the median deviation; 2) Extract multiple features of the data in the time and frequency domains, and map the data set from low dimension to high dimension to reflect the data information of the cutting process; 3) Establish a fusion prediction model and use different machine learning models for iterative training according to different feature data to accurately predict tool wear.

2. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 1, characterized in that: The original data set is the tool wear data of CNC machine tools, including: reading the data of 7 channels collected by sensors in the PHM2010 Industrial Big Data Competition data set: vibration signals in the three directions of x-axis, y-axis and z-axis, cutting force signals in the three directions of x-axis, y-axis and z-axis, and sound signals.

3. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 1, characterized in that: The trimming of outliers adopts the third quartile method, including: For data sets with outliers, the upper quartile Q3 of the collected tool wear data is calculated, and the upper quartile is used as the critical value of invalid data to cut off the outliers.

4. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 1, characterized in that: The data replacement according to the median deviation adopts the Hampel filtering method, including: calculating the median and MAD of the data by setting the number of samples, the number of windows and the upper and lower coefficients. If the sample value is greater than the upper and lower coefficients, it is regarded as an outlier and replaced by the median.

5. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 1, characterized in that: The time and frequency domain features include: a. Extract time domain features based on statistics; b. Convert the time domain signal into the frequency domain signal and extract the frequency domain features based on the amplitude; c. Combine the time domain and frequency domain, and extract joint features of the time and frequency domains based on wavelet transform.

6. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 5, characterized in that: The time domain features include: root mean square, variance, maximum and minimum values, skewness, peak-to-peak value and kurtosis.

7. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 5, characterized in that: The frequency domain characteristics include: spectrum skewness coefficient, spectrum peak coefficient and spectrum power.

8. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 1, characterized in that: The three-way parallel network structure of the fusion prediction model includes: a. Based on statistical time domain feature data, the support vector machine (SVM) model is used for training; b. For frequency domain feature data, the random forest method is used for training; c. For the time-frequency joint domain feature data, a deep neural network is used for training.

9. The multi-feature fusion CNC machine tool tool wear prediction method according to claim 1, characterized in that: The method further includes evaluating the tool wear prediction model based on multi-feature fusion: For different tool type data sets, time domain features are extracted respectively and input into three parallel network structures in the tool wear prediction model of multi-feature fusion for prediction, and RMSE and MAE values ​​are used as evaluation indicators.

Images