Multi-tooth cutter damage monitoring method based on ResNet-GAN and wavelet neural network

Abstract

The invention discloses a multi-tooth cutter damage monitoring method based on ResNet-GAN and a wavelet neural network, and belongs to the technical field of cutter monitoring. The method and the device are used for improving the quality of signals generated in sample expansion, reducing the imbalance ratio of sample training and realizing visualization of depth features. According to the method, sample expansion is performed on the damaged samples, data enhancement of minority class samples is realized, so that the algorithm can more fully learn features of damaged signals, the imbalance ratio of sample training is reduced, and the training accuracy is improved; in addition, through combination of a ResNet residual block and a gradient penalty term, normal gradient propagation can still be ensured in a deep neural network, smooth and stable training can be kept, the gradient explosion phenomenon is reduced, and the quality of generated signals is improved; signal deep features are extracted from different dimensions through a multi-wavelet function kernel convolution module, feature visualization is achieved, and different damage states can be better distinguished.

Description

Technical Field

[0001] The present invention belongs to the technical field of tool monitoring, and in particular relates to a multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network. Background Art

[0002] The tool will wear during normal cutting, and the evolution process follows certain rules, usually including initial wear, stable wear and severe wear stages. However, under intermittent cutting, high-speed processing and high-load conditions, the tool may suddenly break and break due to excessive mechanical stress, thermal stress or impact while wearing. In order to avoid safety accidents and economic losses during the processing process, the research and application of tool damage identification technology is crucial to avoid catastrophic accidents in mechanical processing.

[0003] Smart tools are widely used in the field of tool monitoring technology. Different from ordinary tools, smart tools have multi-dimensional signal sensors that can characterize the tool status. In the field of tool breakage detection, force signals and vibration signals are often used to indirectly determine whether the tool is damaged.

[0004] In milling, multi-tooth milling cutters are often used for cutting. The wear status of different teeth during the machining process is not necessarily the same, and they may break at different times. Therefore, tool breakage monitoring is not just a binary classification problem, it also involves the classification of different breakage states. In the classification process, convolutional neural networks are often used. This network can extract deep features from the signal, but the convolution kernel is random, and people cannot intuitively observe the feature basis of the algorithm classification, and cannot clearly realize the mapping between signal features and tool breakage.

[0005] In addition, in milling, normal tool wear is common, while tool breakage is rare. Therefore, when collecting data sets for training, there will be an imbalance in samples, and the algorithm cannot learn the abnormal features represented by a few samples well, and thus cannot achieve a high classification accuracy. The existing sample expansion network is not very effective in generating samples, and cannot always achieve the effect of "fake is real". One of the reasons is that the gradient explosion phenomenon occurs in deep network training, so it is necessary to add relevant modules to the network to improve the quality of generated signals. Summary of the invention

[0006] Based on the unclear mapping mechanism between the tool damage status and signal characteristic parameters and the sample imbalance phenomenon, the present invention proposes a multi-tooth tool damage monitoring method based on ResNet-GAN and wavelet neural network, which is used to improve the quality of generated signals in sample expansion, reduce the imbalance ratio of sample training, and realize the visualization of deep features.

[0007] The technical solution adopted by the present invention is:

[0008] A multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network includes the following steps:

[0009] S1. Build a tool damage status monitoring system to collect cutting processing sensor signal sample data corresponding to different tool damage states;

[0010] S2. Perform sample expansion for the collected unbalanced real samples;

[0011] S3. The data set after sample expansion is divided into a training set and a test set, the training set is input into the tool damage monitoring model based on wavelet neural network for training, and the test set is used for testing;

[0012] S4. Collect new real-time signals through the tool breakage status monitoring system, input the signals into the trained tool breakage monitoring model based on wavelet neural network, and output the tool breakage status and feature visualization results at this moment.

[0013] Compared with the prior art, the present invention has the following beneficial effects:

[0014] The present invention expands the damaged samples and realizes data enhancement of minority samples, so that the algorithm can learn the characteristics of the damaged signals more fully, reduce the imbalance ratio of sample training, and improve the training accuracy; in addition, the present invention combines the ResNet residual block and the gradient penalty term to ensure the normal propagation of gradients in deep neural networks, maintain smooth and stable training, reduce the gradient explosion phenomenon, and improve the quality of generated signals; through the multi-wavelet function kernel convolution module, the deep features of the signal are extracted from different dimensions, feature visualization is realized, and different damage states can be better distinguished. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is the overall flow chart of the monitoring method of the present invention;

[0016] Figure 2 is a training flow chart of the sample expansion module of the present invention;

[0017] Figure 3 It is a sample expansion network based on ResNet-GAN of the present invention;

[0018] Figure 4 It is the sample expansion result and similarity evaluation result of the present invention;

[0019] Figure 5 The present invention is a tool breakage monitoring model based on wavelet neural network;

[0020] Figure 6 It is a multi-wavelet kernel convolution weight tensor training flow chart of the present invention;

[0021] Figure 7 It is a schematic diagram of the training accuracy and sample visualization results of the present invention. DETAILED DESCRIPTION

[0022] In order to better understand the purpose, structure and function of the present invention, the present invention is described in further detail.

[0023] like Figure 1 As shown, a tool breakage status monitoring system includes: a multi-channel force sensor, a vibration sensor, a data acquisition card, and a computer terminal; the vibration sensor is fixed on a workpiece, the workpiece is fixed on a CNC machine tool, the multi-channel force sensor is fixed on the spindle of the milling cutter, and the vibration sensor and the multi-channel force sensor are both connected to the data acquisition card for signal transmission; the computer terminal includes a memory, a processor, a computer program, and a sensor signal acquisition module, the computer program is stored in the memory and runs on the processor, the sensor signal acquisition module is connected to the data acquisition card for signal transmission, and is used to obtain milling processing sensor signal sample data, and the milling processing sensor signal includes the force signal of the multi-channel force sensor and the vibration signal of the vibration sensor.

[0024] The multi-channel force sensor uses the Swiss Kistler 9170B force measuring tool holder, which can measure the force and torque signals in the X-axis, Y-axis and Z-axis directions.

[0025] The vibration sensor uses Kistler 8762A5 acceleration sensor, which can measure vibration signals in the X-axis, Y-axis and Z-axis directions.

[0026] The machining method is that the milling cutter cuts on the upper surface of the workpiece along multiple transverse paths.

[0027] The CNC machine tool uses a three-axis milling machine VMC855, the milling cutter used is a TAP400R-63-22-3 three-tooth disc milling cutter, the blade model is APMT1135, the material is polycrystalline diamond, and the workpiece material is aluminum-based silicon carbide.

[0028] The data acquisition card uses the domestic Jianyi Technology 5324 and 69529 series acquisition boards, which monitor and record the force signals of multi-channel force sensors and vibration signals of vibration sensors in real time. The signal acquisition frequency is 2500Hz and the number of channels is six.

[0029] The monitoring method of the present invention is executed when the processor runs the computer program.

[0030] Subsequent processing and analysis of milling cutter sample data are carried out on the computing platform of NVIDIA RTX4060Ti 16GB GPU.

[0031] A multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network of the present invention comprises the following steps:

[0032] S1: Build a tool damage status monitoring system to collect cutting processing sensor signal sample data corresponding to different tool damage status; cutting processing sensor signal sample data includes sensor signal data collected by the sensor in each cutting processing experiment and the corresponding damage label category. Figure 1 As shown,

[0033] S101: Build a milling cutter damage status monitoring device;

[0034] S102: Install three intact blades on the blade holder;

[0035] S103: Turn on the machine tool and the multi-channel force sensor and vibration sensor to collect the milling processing sensor signal corresponding to the tool damage state (this represents the three teeth intact state, 3H);

[0036] S104: Replace the chipped blade or disassemble the blade, and obtain four tool damage states in sequence: single-tooth chipped blade (2 intact tools and 1 chipped blade tool, 2H1B), single-tooth complete damage (2 intact tools, 2H1C), two-tooth chipped blade (1 intact tool and 2 chipped blade tools, 1H2B), and two-tooth complete damage (2 chipped blade tools, 1H2C), and collect the corresponding milling processing sensor signals of different damage labels in sequence;

[0037] S105: Constructing milling processing sensor signal sample data through the collected sensor signals and damaged labels.

[0038] S2: for the collected unbalanced real samples, use the cutting processing sensor signal sample data to perform data segmentation and standardization processing on the real samples of each label, and then input them into the sample expansion network based on ResNet-GAN for training to generate the required number of new samples, that is, perform sample expansion work to reduce the data imbalance ratio;

[0039] The sample expansion network based on ResNet-GAN, the full name of the network is ResNet-ACWGAN-GP network, that is, a generative adversarial network based on Wasserstein distance and gradient penalty for residual network assisted classification. The network architecture includes residual blocks (ResNet), which are used to more effectively learn the characteristics of real samples; auxiliary classifiers (discriminators, Auxiliary Classifier, AC) help the generator to learn data distribution more effectively; the loss function of the auxiliary classifier uses Wasserstein distance as an indicator to calculate the difference between real samples and generated samples, and adds a gradient penalty term (GP) to improve training stability. The combination of residual blocks (ResNet) and gradient penalty terms (GP) ensures that in deep deep learning models, while maintaining smooth and stable training, the gradient explosion phenomenon will not occur due to too many network layers. Figure 2 , Figure 3 As shown,

[0040] S201: establishing a data set, each sample of the data set includes the signal after the data segmentation and standardization processing and its corresponding damaged label category, and then dividing the data set into a training set and a test set;

[0041] S20101: Segment the processing signal samples using a fixed-size time window to remove the redundant components of the advance and retract tool to obtain a stable cutting signal;

[0042] S20102: installing a certain number of sample points (here 1000) for the stable cutting signal to perform data segmentation to obtain a number of original signal samples;

[0043] S20103: performing standardization processing on each of the above samples according to formula (1) so that the number of sample points is distributed between [-1, 1], thereby obtaining the sensor signal sample data sequence;

[0044]

[0045] Where x is the original signal, x1 is the standardized signal, max{x} represents the maximum value in the sample, and min{x} represents the minimum value in the sample.

[0046] S20104: The data set consisting of the standardized real samples is randomly divided into a training set and a test set in a ratio of 8:2.

[0047] The unbalanced real samples include normal samples and damaged samples. The damaged samples are further divided into different categories according to the damage form. The number of normal samples is often greater than the number of damaged samples. The samples with normal damaged label categories are used as the majority class sample set, and the damaged samples with different damaged label categories are used as the minority class sample set.

[0048] S202: Considering the minority sample set in the training set, the minority sample set is input into the sample expansion network based on ResNet-GAN for training, and the generator and the discriminator are trained at the same time until the discriminator cannot distinguish between the real samples and the generated samples, that is, the Nash equilibrium state is reached. Please refer to the attached figure of the specification for the network structure;

[0049] S20201: Inputting a random noise signal into the sample expansion network based on ResNet-GAN to obtain a generated signal;

[0050] S20202: Inputting the standardized corresponding label minority samples and the generated signal into the auxiliary classifier at the same time to obtain a classification result;

[0051] S20203: Compare the above classification results with the true label corresponding to the sample to obtain the scores of the generator and the auxiliary classifier;

[0052] In scoring, the auxiliary classifier will judge the score of real data as 1 and the score of generated data (fake data) as 0, and the score is expressed as the mathematical expectation of a batch of data.

[0053] The loss functions L(G) and L(D) of the generator and auxiliary classifier are shown in formulas (2) and (3):

[0054]

[0055] Where E z～Pz(z) [D(G(z))] is the expected value of the score of the generated sample in the auxiliary classifier. The generator G generates data through random noise z. The generator expects this value to be as large as possible, that is, the generator "cheats" the discriminator and achieves the result of "false and true". Adding a negative sign to this item can make the entire loss function decrease during the training process;

[0056] There are three terms in the loss function of the auxiliary classifier. The first term Represents the score of the real data x in the auxiliary classifier. This item is the auxiliary classifier's hope to give a high score to the real data; the second item It represents the score of the generated data G(z) in the auxiliary classifier. The auxiliary classifier hopes that the fake data will be scored low. The third item It is the gradient penalty term. By calculating the gradient norm of the interpolated sample between the generated sample and the real sample, the gradient of the auxiliary classifier is penalized to make it smooth for the gradient of the data input. The norm of the gradient is expected to be close to 1. Finally, the positive and negative signs are added to make the entire loss function decrease during the training process.

[0057] S20204: Determine whether the generator and the auxiliary classifier have reached Nash equilibrium (i.e., their scores are both around 0.5);

[0058] S20205: If Nash equilibrium is not reached, the loss function is calculated and back propagation is performed, and the generator and auxiliary classifier are iteratively trained. If Nash equilibrium is reached, the next step is entered;

[0059] S20206: When the Nash equilibrium is reached, the discriminator cannot distinguish between real samples and generated samples, and a trained generator network is obtained.

[0060] S203: Generate new samples using the generator trained in step 202. Each time a new sample is generated, a similarity evaluation is performed with real samples of the minority class sample set of the corresponding category, and samples that pass the evaluation are used as new samples for expansion.

[0061] S20301: Based on actual experience, set a hard threshold range for the maximum mean difference, Euclidean distance, and Pearson correlation coefficient;

[0062] The expressions of maximum mean difference (MMD), Euclidean distance (ED), and Pearson correlation coefficient (PCC) are as follows, where x represents the real data and y represents the generated data.

[0063]

[0064] Where m = n is the number of real data sample points and the number of generated data sample points, There are two sets of data points (x i ,y j ) between Gaussian kernel parameters;

[0065] S20302: Calculate the maximum mean difference, Euclidean distance, and Pearson correlation coefficient between the new sample and the real sample, and determine whether each indicator exceeds the threshold in turn. If at least one indicator exceeds the threshold, the sample is considered unqualified. If the three indicators do not exceed the threshold, the sample passes the hard threshold test, and then enter S20303;

[0066] In this embodiment, the MMD threshold is [0, 0.005], the ED threshold is [0, 6], and the PCC threshold is [0.8, 1];

[0067] S20303: First, generate a certain number of samples that pass the hard threshold test according to the previous steps. When the total number of samples generated reaches N (N is 3 in this example), calculate the mean and standard deviation of each indicator of these N samples, filter out outliers using the 3σ principle, mark this sample as unqualified, and mark the remaining samples as qualified.

[0068] S20304: Then, for each sample that passes the hard threshold test, determine whether each of its indicators is within the 3σ range of the qualified sample set. If at least one indicator is not within the 3σ range, the sample is considered unqualified. If all three indicators are within the 3σ range of the qualified sample set, the sample passes the soft threshold test, and then enters S20305;

[0069] S20305: Update the mean, standard deviation and 3σ range of the qualified sample set to facilitate the generation of subsequent samples;

[0070] S20306: When the number of qualified samples reaches the preset value, stop sample generation and proceed to subsequent operations. Figure 4 The generated samples are shown, as well as a comparison of the generated indicators of different networks with this network.

[0071] S3: The data set after sample expansion is divided into a training set and a test set. The training set is input into the tool breakage monitoring model based on wavelet neural network for training, and the test set is used for testing to evaluate the quality of the model.

[0072] The dataset after sample expansion includes the majority sample set consisting of real data and the minority sample set consisting of real data and expanded fake samples. Figure 3 to Figure 6 As shown,

[0073] S301: Establish a data set, convert the signal values ​​of each batch of samples in the training set into tensors and input them into the network. In the training set, for normal labels, the data set contains real samples, and for abnormal labels, the data set contains real samples and generated samples. In the test set, the data set contains real samples.

[0074] S302: performing convolution operations on the samples with different wavelet function kernels in sequence to obtain respective feature tensor matrices, training weight tensors corresponding to different wavelet function kernels, and performing weighted addition on respective feature tensor matrices;

[0075] The wavelet function kernels used in this embodiment are Laplace wavelet, Morlet wavelet and Mexhat wavelet, and their expressions are as follows:

[0076]

[0077] Where: ψ is the wavelet basis function, t is time, s is the scale factor, u is the translation factor, f is the signal frequency, usually 50 Hz, ξ is the viscous damping ratio, usually 0.03, A is the wavelet normalization function, which is 0.08;

[0078] S30201: Perform convolution operations on the signal samples with different wavelet function kernels to obtain their respective coefficient tensor matrices;

[0079] S30202: Combining tensors after compressing the time domain dimension through the global maximum pooling layer (GAP) and the global average pooling layer (GMP);

[0080] S30203: Input the merged tensor into the multi-layer perceptron layer (MLP) and output the weight tensor.

[0081] S30204: Perform convolution operations on signal samples and different wavelet functions to obtain their respective feature tensor matrices. Feature visualization can be achieved through weight tensors and feature tensor matrices.

[0082] S30205: Perform weighted addition on the respective feature tensor matrices and output a tensor;

[0083] S303: After passing through the residual block, normalization layer, activation layer and other modules, and then through the softmax layer, the prediction result is obtained, and the prediction result is compared with the true label to obtain the loss value to change the network weight.

[0084] S30301: The tensor after passing through the feature extraction module based on the wavelet neural network passes through the ResNet residual block module in sequence, and then passes through the global average pooling layer and the softmax layer to obtain a damage prediction result;

[0085] S30302: Compare the predicted label with the true label, use the cross entropy loss function to calculate the loss function, and then perform error back propagation to change the network weights;

[0086] S304: When a sufficient number of training rounds is reached, the training is stopped, and a tool damage monitoring model based on a wavelet neural network is output;

[0087] S305: The feature tensor matrix is ​​passed through the T-SNE feature dimension reduction module to obtain a two-dimensional matrix for result visualization.

[0088] S4: The tool damage status monitoring system collects new real-time signals, inputs the signals into the tool damage monitoring model based on the trained wavelet neural network, and outputs the tool damage status and feature visualization results at this moment, such as Figure 7 The comparison between different networks and this network in terms of damage identification accuracy and the visualization of signal features by this network (MKWC) are shown.

[0089] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A multi-tooth tool damage monitoring method based on ResNet-GAN and wavelet neural network, characterized by: The following steps are involved: S1. Build a tool damage status monitoring system to collect cutting processing sensor signal sample data corresponding to different tool damage states; S2. Perform sample expansion for the collected unbalanced real samples; S3. The data set after sample expansion is divided into a training set and a test set, the training set is input into the tool damage monitoring model based on wavelet neural network for training, and the test set is used for testing; S4. Collect new real-time signals, input the signals into the trained tool damage monitoring model based on wavelet neural network, and output the tool damage status and feature visualization results at this moment.

2. According to claim 1, a multi-tooth tool damage monitoring method based on ResNet-GAN and wavelet neural network is characterized in that: In S1, the tool breakage status monitoring system includes: a multi-channel force sensor, a vibration sensor, a data acquisition card and a computer terminal; the vibration sensor is fixed on a workpiece, the workpiece is fixed on a CNC machine tool, the multi-channel force sensor is fixed on the spindle of the milling cutter, and the vibration sensor and the multi-channel force sensor are both connected to the data acquisition card for signal transmission; the computer terminal includes a memory, a processor, a computer program and a sensor signal acquisition module, the computer program is stored in the memory and runs on the processor, the sensor signal acquisition module is connected to the data acquisition card for signal transmission, and is used to obtain milling processing sensor signal sample data, and the milling processing sensor signal includes the force signal of the multi-channel force sensor and the vibration signal of the vibration sensor.

3. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 2 is characterized in that: The specific steps of S1 are: S101. Build a milling cutter damage status monitoring device; S102. Install three intact blades on the blade holder; S103. Turn on the machine tool and the multi-channel force sensor and vibration sensor to collect the milling sensor signal corresponding to the tool damage state at this time; S104. Replace the chipped blade or remove the blade, obtain four tool damage states: single-tooth chipped blade, single-tooth complete damage, two-tooth chipped blade, two-tooth complete damage, and collect the corresponding milling processing sensor signals of different damage labels in turn; S105. Construct cutting processing signal sample data through the collected sensor signals and damage labels.

4. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 1 is characterized in that: S2 is, for the collected unbalanced real samples, using the cutting processing sensor signal sample data, performing data segmentation and standardization processing on the real samples of each label, and then inputting them into the sample expansion network based on ResNet-GAN for training to generate the required number of new samples, that is, performing sample expansion work to reduce the data imbalance ratio; The unbalanced real samples include normal samples and damaged samples. The damaged samples are further divided into different categories according to the damage form. The samples with normal damage label categories are used as the majority class sample set, and the damaged samples with different damage label categories are used as the minority class sample set.

5. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 4 is characterized in that: In S2, the network architecture of the sample expansion network based on ResNet-GAN includes Residual blocks are used to learn the features of real samples more effectively; Auxiliary classifier helps the generator learn data distribution more effectively. The loss function of the auxiliary classifier uses Wasserstein distance as an indicator to calculate the difference between real samples and generated samples, and adds a gradient penalty term. The loss function L(D) of the auxiliary classifier is: in: Represents the score of the real data x in the auxiliary classifier, It represents the score of the generated data G(z) in the auxiliary classifier. is the gradient penalty term.

6. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 5 is characterized in that: In S2, the training steps of the sample expansion process are as follows: S201. Establish a data set, each sample of the data set includes a signal after data segmentation and normalization and its corresponding damaged label category, and then divide the data set into a training set and a test set; S202. Input a small number of sample sets into a sample expansion network based on ResNet-GAN for training, and train the generator and discriminator at the same time until the discriminator cannot distinguish between real samples and generated samples, that is, a Nash equilibrium state is reached; S203. Generate new samples using the generator trained in step 202. Each time a new sample is generated, a similarity evaluation is performed with the real samples of the minority class sample set of the corresponding category, and the qualified samples are used as the expanded new samples.

7. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 6 is characterized in that: The similarity evaluation in S203 includes the following steps: S20301. According to actual experience, set a hard threshold range for the maximum mean difference, Euclidean distance, and Pearson correlation coefficient; S20302. Calculate the maximum mean difference, Euclidean distance, and Pearson correlation coefficient between the new sample and the real sample, and determine whether each indicator exceeds the threshold in turn. If at least one indicator exceeds the threshold, the sample is considered unqualified. If the three indicators do not exceed the threshold, the sample passes the hard threshold test, and then proceed to step 20303; S20303. First, generate samples that must pass the hard threshold test according to the previous steps. When the total number of generated samples reaches N, calculate the mean and standard deviation of each indicator of these N samples, filter out outliers using the 3σ principle, mark this sample as unqualified, and mark the remaining samples as qualified; S20304. Then, for each sample that passes the hard threshold test, determine whether each of its indicators is within the 3σ range of the qualified sample set. If at least one indicator is not within the 3σ range, the sample is determined to be unqualified. If all three indicators are within the 3σ range of the qualified sample set, the sample passes the soft threshold test, and then proceed to step 20305; S20305. Update the mean, standard deviation and 3σ range of the qualified sample set to facilitate the generation of subsequent samples; S20306. When the number of qualified samples reaches the preset value, stop sample generation and proceed to subsequent operations.

8. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 1 is characterized in that: In S3, the data set after sample expansion includes a majority sample set consisting of real data and a minority sample set consisting of real data and expanded false samples.

9. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 8, characterized in that: In S3, training is performed in a tool breakage monitoring model based on a wavelet neural network, including the following steps: S301. Establish a data set, convert the signal value of each batch of samples in the training set into a tensor input to a feature extraction module based on a wavelet neural network; S302. Convolve the samples with different wavelet function kernels in turn to obtain their respective feature tensor matrices, train the weight tensors corresponding to the different wavelet function kernels, and perform weighted addition on their respective feature tensor matrices; S303. After passing through the residual block, normalization layer, activation layer and other modules, and then passing through the softmax layer, the prediction result is obtained, and the prediction result is compared with the true label to obtain the loss value to change the network weight; S304. When sufficient training rounds are reached, the training is stopped and a tool breakage monitoring model based on a wavelet neural network is output; S305. The feature tensor matrix is ​​passed through the T-SNE feature dimensionality reduction module to obtain a two-dimensional matrix for result visualization.

10. The multi-tooth tool breakage monitoring method based on ResNet-GAN and wavelet neural network according to claim 9, characterized in that: In S302, training weight tensors corresponding to different wavelet function kernels includes the following steps: S30201. Perform convolution operations on the signal samples with different wavelet function kernels to obtain their respective feature tensor matrices; S30202. Combining the tensors after compressing the time domain dimension through the global maximum pooling layer and the global average pooling layer; S30203. Input the merged tensor into the multi-layer perceptron layer and output the weight tensor.

Images