Cutting flutter identification method based on multi-feature fusion convolutional neural network

Abstract

The invention discloses a cutting flutter identification method based on a multi-feature fusion convolutional neural network. The method is characterized by comprising the following steps: S1, acquiring multiple groups of vibration signal data samples in a cutting process; s2, constructing an explicit feature extraction module of the vibration signal based on a time domain analysis method, a frequency domain analysis method or a time-frequency domain analysis method so as to obtain various flutter features of the vibration signal in a time domain and a frequency domain; s3, establishing an implicit feature extraction module of the vibration signal based on a convolutional neural network, and further constructing a multi-feature fusion convolutional neural network model taking the vibration signal as input and the processing state as output; s4, training the constructed multi-feature fusion convolutional neural network model through the obtained multiple groups of vibration signal data samples; and S5, identifying the cutting state based on the trained multi-feature fusion convolutional neural network model. According to the method, the processing flutter recognition model is jointly constructed by combining the display features which reflect the flutter information and the implicit features which are automatically learned by the convolutional neural network and are extracted from the vibration signals, so that the accuracy of processing flutter recognition is improved.

Description

Technical Field

[0001] The present invention relates to a machining technology, in particular to a chatter suppression technology for machine tool cutting, and specifically to a cutting chatter recognition method based on a multi-feature fusion convolutional neural network. Background Art

[0002] Unreasonable selection of process parameters will cause the tool to vibrate violently and unstably during machining. This phenomenon is called machining chatter, which seriously affects the surface quality of parts, limits production efficiency, increases tool wear, and even damages the machine tool components. It is the main reason restricting the productivity of machine tools. Therefore, studying the cutting chatter recognition method based on vibration signals is of great significance for realizing the online monitoring of machining chatter and then ensuring machining stability.

[0003] Traditional chatter recognition methods extract chatter features of vibration signals based on time-domain and frequency-domain analysis, and then use threshold determination or shallow machine learning models for classification. Such methods have high computational efficiency and the extracted features are highly interpretable. However, they rely on expert experience to define feature extraction rules, cannot capture complex patterns in signals, and have limited capabilities in recognizing machining chatter under complex working conditions. In recent years, chatter recognition methods based on deep learning have gradually developed, which automatically extract implicit features of vibration signals through deep learning models such as convolutional neural networks and perform chatter recognition. Such methods do not rely on expert experience, have strong feature extraction capabilities, and can process complex vibration signal data. However, the extracted features depend on training data, have no specific physical meaning, and have weak generalization capabilities for different working conditions. Summary of the Invention

[0004] The purpose of the present invention is to provide a cutting chatter recognition method based on a multi-feature fusion convolutional neural network, comprehensively consider the time-domain and frequency-domain feature information of vibration signals to construct a multi-feature fusion convolutional neural network model, fuse explicit features and implicit features to characterize the machining process state from multiple perspectives, and then accurately recognize the chatter phenomenon during cutting.

[0005] The technical solution of the present invention is as follows:

[0006] A cutting chatter recognition method based on a multi-feature fusion convolutional neural network, the steps are as follows:

[0007] S1. Obtain multiple groups of vibration signal data samples during the cutting process, and each group of data samples consists of a vibration signal and its corresponding machining state;

[0008] S2. Based on time-domain analysis, frequency-domain analysis or time-frequency domain analysis methods, construct an explicit feature extraction module for vibration signals to obtain various chatter features of vibration signals in the time domain and frequency domain;

[0009] S3. Establish an implicit feature extraction module for vibration signals based on a convolutional neural network, and then construct a multi-feature fusion convolutional neural network model with vibration signals as the input and machining states as the output;

[0010] S4. Train the constructed multi-feature fusion convolutional neural network model with the obtained multi-group vibration signal data samples.

[0011] S5. Identify the cutting machining state based on the trained multi-feature fusion convolutional neural network model.

[0012] Preferably, the vibration signal in S1 is a cutting force signal, a sound signal, an electric current signal or an acceleration signal; the machining state is one of a finite number of chatter levels divided based on the vibration intensity during the machining process, where different chatter levels are represented by different numbers.

[0013] Preferably, the methods used for time-domain analysis, frequency-domain analysis or time-frequency domain analysis in S2 include statistical parameter method, dimensionless index analysis, envelope spectrum analysis, amplitude spectrum analysis, power spectrum analysis, Fourier transform, short-time Fourier transform, wavelet transform, empirical mode decomposition, variational mode decomposition. Among the multiple chatter features extracted, the time-domain features include waveform index, peak index, pulse index, margin index, skewness index and kurtosis index; the frequency-domain features include center frequency index, mean square frequency index and frequency variance index.

[0014] Preferably, establishing the implicit feature extraction module for vibration signals based on a convolutional neural network in S3 specifically means processing the vibration time-domain signal, frequency-domain signal or time-frequency domain signal input to the implicit feature extraction module through multiple convolutional units, where each convolutional unit consists of a convolutional layer, a pooling layer and an activation function, and realizing the feature extraction of vibration signals by gradually reducing the convolutional kernel size.

[0015] Preferably, the multi-feature fusion convolutional neural network model in S3 consists of an explicit feature extraction module, an implicit feature extraction module, a multi-feature fusion module and a chatter recognition module. The multi-feature fusion module fuses the two groups of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module to obtain a new group of feature vectors, and inputs them to the chatter recognition module, so as to output the corresponding machining state.

[0016] Preferably, the specific method for fusing the two sets of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module through the multi-feature fusion module includes: (1) directly concatenating the two sets of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module; (2) constructing a learnable weight vector to perform weighted fusion on the elements in the two sets of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module; (3) performing attention weighted fusion on the two sets of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module based on the multi-head attention mechanism.

[0017] Preferably, the flutter identification module is specifically a fully connected neural network, a support vector machine, or a hybrid model weighted by multiple fully connected neural networks or support vector machines.

[0018] Preferably, the training of the constructed multi-feature fusion convolutional neural network model in S4 is specifically to measure the prediction error of the multi-feature fusion convolutional neural network model using the cross-entropy loss function, Kullback-Leibler divergence, mean square error loss function, or mean absolute error, and complete the model training by optimizing the network parameters through the backpropagation algorithm.

[0019] Preferably, the identification of the cutting machining state in S5 is specifically to input the collected cutting machining vibration signal into the trained multi-feature fusion convolutional neural network model and output the corresponding cutting machining state.

[0020] The beneficial effects of the present invention are:

[0021] By fusing the explicit features extracted by the time-frequency domain analysis method and the implicit features extracted by the convolutional neural network, the present invention realizes the complementary of multi-angle feature information of the vibration signal, and further enhances the machining flutter identification ability of the model.

[0022] By combining the explicit features reflecting flutter information extracted from the vibration signal and the implicit features automatically learned by the convolutional neural network, the present invention jointly constructs a machining flutter identification model, which improves the accuracy of machining flutter identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a schematic diagram of the cutting machining flutter identification process of the present invention.

[0024] Figure 2 It is a schematic diagram of the multi-feature fusion convolutional neural network framework of the embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0025] The technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0026] As Figure 1-2 shown.

[0027] A method for identifying chatter in machining based on a multi - feature fusion convolutional neural network, as Figure 1 shown, it includes the following steps:

[0028] S1. Obtain multiple groups of vibration signal data samples during the machining process. Each group of data samples consists of a vibration signal and its corresponding machining state.

[0029] A milling experiment was carried out on a DMG DMU80P machining center. The workpiece material was high - performance 7 - series aluminum alloy. The acceleration signal was used as the machining chatter monitoring signal. Two PCB sensors were selected as the acceleration sensors, which were respectively attached to the X and Y directions of the aluminum alloy workpiece. The DEWEsoft acquisition card was used as the acquisition module. The machining state was determined according to the surface machining quality of the part and divided into two levels: chatter and stable, which were represented by 1 and 0 respectively. A total of 1020 groups of data were collected in the experiment, of which 970 groups constituted the training set and the remaining 50 groups of data constituted the test set. For the training set, the signal slices obtained with a set of machining parameters were made into one piece of data for model training. After the model training was completed with the training set data, the performance of the model was evaluated with the 50 groups of data in the test set.

[0030] S2. Conduct time - domain analysis, frequency - domain analysis or time - frequency domain analysis on the obtained vibration signals, and extract various signal features in the time domain and frequency domain of the vibration signals.

[0031] After obtaining the initial vibration signal through the acceleration sensor, the waveform index, peak index, pulse index, and margin index in the time - domain signal are obtained by combining the statistical parameter method and the dimensionless index analysis method; the centroid frequency index and mean - square frequency index in the frequency - domain signal are obtained through Fourier transform.

[0032] S3. Based on the convolutional neural network, establish an implicit feature extraction module for vibration signals, and then construct a multi - feature fusion convolutional neural network model with vibration signals as the input and machining states as the output;

[0033] such as Figure 2As shown in the figure, the implicit feature extraction module for vibration signals based on a convolutional neural network is specifically as follows: The time-domain vibration signal, frequency-domain signal, or time-frequency domain signal input to the implicit feature extraction module is processed by three consecutive convolutional units. Each convolutional unit consists of a one-dimensional convolutional layer, a max-pooling layer, and a ReLU activation function. The convolutional kernel sizes are 64×1, 32×1, and 16×1 in sequence. The feature extraction of the vibration signal is achieved by gradually reducing the convolutional kernel size. The pooling layer of each convolutional module uses one-dimensional max-pooling operation with a stride of 2, thereby realizing the extraction of implicit features in the vibration signal. Then, the multi-feature fusion module fuses the two sets of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module to obtain a new set of feature vectors, and inputs them into the flutter recognition module, thereby outputting the corresponding processing state.

[0034] S4: Train the constructed multi-feature fusion convolutional neural network model with the obtained multiple sets of vibration signal data samples.

[0035] For the binary classification characteristic of the flutter recognition task, use the Adam optimizer to minimize the cross-entropy loss. The initial learning rate is set to 10 -3 , the batch size is 50, and the number of iterations is 500. Train the parameters in the multi-feature fusion convolutional neural network with the multiple sets of vibration signal data samples obtained in S1, thereby completing the training of the model.

[0036] S5. Identify the cutting processing state based on the trained multi-feature fusion convolutional neural network model.

[0037] The specific identification of the cutting processing state is to input the collected cutting processing vibration signal into the trained multi-feature fusion convolutional neural network model and output the corresponding cutting processing state.

[0038] To further illustrate the effectiveness of the cutting processing flutter recognition method based on a multi-feature fusion convolutional neural network described in the present invention, use 50 test data to evaluate the prediction performance of the constructed multi-feature fusion convolutional neural network model, and compare it with a convolutional neural network model without fusing explicit features. The architecture of this convolutional neural network model is the same as the above. The comparison results are as follows:

[0039] Method Amount of training data Amount of test data Training error Test accuracy Convolutional neural network 970 50 0.135 92.0％ The present invention 970 50 0.001 96.0％

[0040] The results show that the cutting processing flutter recognition method based on a multi-feature fusion convolutional neural network proposed in the present invention can achieve better prediction performance.

[0041] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than 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 they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

[0042] The parts not involved in the present invention are the same as or can be implemented by the prior art.

Claims

1. A method for identifying chatter in cutting processes based on a multi-feature fusion convolutional neural network, characterized in that It includes the following steps: S1. Obtain multiple groups of vibration signal data samples during the cutting process, where each group of data samples consists of a vibration signal and its corresponding machining state; S2. Based on time-domain analysis, frequency-domain analysis, or time-frequency domain analysis methods, construct an explicit feature extraction module for vibration signals to obtain various chatter features of vibration signals in the time domain and frequency domain; S3. Based on a convolutional neural network, establish an implicit feature extraction module for vibration signals, and then construct a multi-feature fusion convolutional neural network model with vibration signals as the input and machining states as the output; S4. Train the constructed multi-feature fusion convolutional neural network model with the obtained multiple groups of vibration signal data samples; S5. Identify the cutting machining state based on the trained multi-feature fusion convolutional neural network model.

2. The method according to claim 1, wherein The vibration signal in S1 is a cutting force signal, a sound signal, a current signal, or an acceleration signal; the machining state is one of a finite number of chatter levels divided based on the vibration intensity during the machining process, where different chatter levels are represented by different numbers.

3. The method according to claim 1, wherein The methods used in time-domain analysis, frequency-domain analysis, or time-frequency domain analysis in S2 include statistical parameter method, dimensionless index analysis, envelope spectrum analysis, amplitude spectrum analysis, power spectrum analysis, Fourier transform, short-time Fourier transform, wavelet transform, empirical mode decomposition, variational mode decomposition.

4. The method according to claim 1, wherein Specifically, establishing an implicit feature extraction module for vibration signals based on a convolutional neural network in S3 is to process the vibration signal input to the implicit feature extraction module through multiple convolutional units, where each convolutional unit consists of a convolutional layer, a pooling layer, and an activation function, and the feature extraction of the vibration signal is achieved by gradually reducing the convolutional kernel size.

5. The method according to claim 1, wherein The multi-feature fusion convolutional neural network model in S3 consists of an explicit feature extraction module, an implicit feature extraction module, a multi-feature fusion module, and a chatter recognition module. The two groups of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module are fused through the multi-feature fusion module to obtain a new group of feature vectors, which are input to the chatter recognition module, thereby outputting the corresponding machining state.

6. The method according to claim 5, characterized in that, The specific methods for fusing the two groups of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module through the multi-feature fusion module include: (1) directly splicing the two groups of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module; (2) constructing a learnable weight vector to perform weighted fusion on the elements in the two groups of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module; (3) performing attention weighted fusion on the two groups of feature vectors obtained by the explicit feature extraction module and the implicit feature extraction module based on the multi-head attention mechanism.

7. The method according to claim 5, wherein The chatter recognition module is specifically a fully connected neural network, a support vector machine, or a hybrid model weighted by multiple fully connected neural networks or support vector machines.

8. The method according to claim 1, wherein Specifically, the training of the constructed multi-feature fusion convolutional neural network model in S4 is as follows: the cross-entropy loss function, Kullback-Leibler divergence, mean squared error loss function or mean absolute error is used to measure the prediction error of the multi-feature fusion convolutional neural network model, and the model training is completed by optimizing the network parameters through the backpropagation algorithm.

9. The method according to claim 1, characterized in that Specifically, the recognition of the cutting machining state in S5 is as follows: the collected cutting machining vibration signal is input into the trained multi-feature fusion convolutional neural network model, and the corresponding cutting machining state is output.

Images