A tool wear state monitoring method based on DDPM diffusion model

Abstract

The present invention discloses a tool wear state monitoring method based on DDPM diffusion model, which specifically includes the following steps: step 1: collecting cutting force signals and corresponding tool wear states during milling; step 2: determining the optimal parameters for cutting force signal processing by minimizing multi-scale permutation entropy principle, and then obtaining the phase space matrix by using the optimal parameters; step 3: converting the cutting force signal into a grayscale image according to the phase space matrix, and superimposing a color distance map; step 4: using the color distance map as the original input to the DDPM diffusion model to generate a pseudo sample of each type of wear state; step 5: dividing the color distance map into a training set of several unbalanced situations and a balanced test set, and balancing the training set; step 6: using the training set and the test set to train and test the Resnet18 network respectively, and classifying and identifying the tool wear state. The present invention can achieve high-precision tool state monitoring and accurate classification.

Description

Technical Field

[0001] The present invention relates to the field of metal cutting processing, and specifically to a tool wear state monitoring method based on a DDPM diffusion model. Background Art

[0002] As the most sensitive and fragile part of the machining process, the gradual wear of the tool has become an important factor affecting the dimensional accuracy, surface roughness and machining cost of the workpiece. Its wear value directly affects the machining efficiency and the quality and accuracy of the workpiece. Therefore, early detection and monitoring of the wear process is necessary for cost optimization and safe maintenance of the machining process. With the increasing complexity of mechanical systems and the latest application of intelligent technology, tool wear condition monitoring based on deep learning has received increasing attention because it avoids manual feature extraction and improves monitoring performance compared with traditional machine learning methods. However, the performance of tool condition monitoring depends largely on the quality and quantity of collected data to achieve good classification accuracy, which may have the following problems in real environments: (1) Due to the complexity of the machining process, some samples of wear conditions are often lost. (2) At present, most classification models based on deep learning methods are supervised learning, which requires a large amount of labeled data to train the model. However, collecting a large amount of labeled data in a real machining environment is expensive and time-consuming. (3) The tool life is different at different wear stages, which easily leads to imbalanced data collected in the real environment. Using unbalanced samples for fault diagnosis increases the sensitivity of the model to the majority class samples and makes the classification boundary of the model classifier biased towards the majority class samples. Minority class samples are often considered as noise of majority class samples, making them difficult to identify and affecting the accuracy of fault diagnosis, thus causing overfitting or misdiagnosis.

[0003] At present, the common methods to solve the problem of model recognition accuracy under sample imbalance conditions are mainly divided into two categories. One is to improve the classification algorithm and improve the classification accuracy of the minority class. This is often performed by changing the loss function of the algorithm to assign appropriate weights to minority class samples, such as the self-organizing map (SOM) model to handle imbalanced data, and the multi-index generative adversarial network (MI-GAN) to monitor minority class samples by learning the characteristics of majority class samples and using L2-norm, CORT and RISK score metrics. However, the performance of these methods is limited by their ability to understand and capture the distribution of minority group data, and may ignore other important features and information, which will cause the model to perform poorly when facing unknown data or changes in data distribution. The other is to reduce data imbalance and generate more minority class samples through sample generation methods to achieve data balance and effectively alleviate the problem of sample imbalance. Such as variational autoencoder (VAE), generative adversarial network (GAN), deep convolutional generative adversarial network (DCGAN), self-attention generative adversarial network (SAGAN). However, GAN and its various variants are prone to problems such as training instability, gradient disappearance, and mode collapse. Summary of the invention

[0004] The purpose of the present invention is to provide a tool wear state monitoring method based on the DDPM diffusion model. The present invention can achieve high-precision tool state monitoring and accurate classification.

[0005] The technical solution of the present invention is a tool wear state monitoring method based on the DDPM diffusion model, which specifically includes the following steps:

[0006] Step 1: Collect cutting force signals and corresponding tool wear status during milling;

[0007] Step 2: Determine the optimal parameters for cutting force signal processing by minimizing the multi-scale permutation entropy principle, and then use the optimal parameters to obtain the phase space matrix;

[0008] Step 3: Convert the cutting force signal into a grayscale image according to the phase space matrix, and superimpose the grayscale image to generate a color distance map;

[0009] Step 4: Use the color distance map as the original input to the DDPM diffusion model to generate pseudo samples for each type of wear state;

[0010] Step 5: Divide the color distance map into several unbalanced training sets and balanced test sets, and supplement the training set with the generated pseudo samples to balance the training set;

[0011] Step 6: Use the training set and test set to train and test the Resnet18 network respectively, and then use the trained Resnet18 network to classify and identify the tool wear status.

[0012] In the aforementioned tool wear state monitoring method based on the DDPM diffusion model, in step 2, the optimal parameters include the optimal embedding dimension m*, the optimal time delay t* and the optimal coarse-graining coefficient S*; the optimal parameters are determined by the principle of minimizing multi-scale permutation entropy by first coarsening the given time series to obtain a new coarse-grained time series, then reconstructing the coarse-grained time series, arranging the reconstructed coarse-grained time series in ascending order, calculating the probability of each symbol sequence, normalizing and calculating the permutation entropy of each coarse-grained time series, and finally performing minimization processing.

[0013] In the above-mentioned tool wear state monitoring method based on DDPM diffusion model, the time series coarse-graining process is to calculate the time series X={x1, x 2, ..., x N}Perform coarse-graining processing to obtain a coarse-grained time series Y (S) ={y1 (S) , y2 (S) , ..., y T (S)},Right now:

[0014]

[0015]

[0016] Where: N represents the length of time series data, S represents the coarse-grained coefficient, and T represents the coarse-grained length of the sequence.

[0017] In the tool wear state monitoring method based on the DDPM diffusion model, the time series reconstruction is to obtain the coarse-grained time series y i(s) To refactor:

[0018]

[0019] Where: k is the kth reconstruction component, m is the embedding dimension, and t is the delay time.

[0020] In the aforementioned tool wear state monitoring method based on the DDPM diffusion model, the probability calculation is to arrange the reconstructed coarse-grained time series in ascending order. The symbol sequence S(r) = {j1j2, ..., j m}(r=1,2,...,k,k≤m!), calculate the probability p of each symbol sequence r The permutation entropy calculation is to calculate the permutation entropy value of each coarse-grained time series by normalization to obtain the multi-scale permutation entropy H p :

[0021]

[0022] In the aforementioned tool wear condition monitoring method based on the DDPM diffusion model, the minimization process is to construct an optimization model that minimizes the square of the MPE skewness:

[0023]

[0024] Where: SKEW represents the square of the skewness of the MPE of all real samples, q represents the real sample, Q is the size of all real samples q, SK q represents the MPE skewness of the real sample q, H q represents the MPE of the real sample q, H avg represents the average MPE sample of all true samples q.

[0025] In the tool wear condition monitoring method based on the DDPM diffusion model, the optimal parameters m*, t* and S* are obtained by minimizing the MPE skewness square optimization model in step 6, and then the phase space matrix Y * :

[0026]

[0027] in, make Represents Y * The i-th state in .

[0028] In the aforementioned tool wear state monitoring method based on the DDPM diffusion model, in step 3, the Euclidean distance between various wear states in the phase space matrix is ​​calculated. The phase space matrix is ​​encoded as a grayscale image, and then the grayscale distance map is superimposed to generate a color distance map.

[0029] In the aforementioned tool wear state monitoring method based on the DDPM diffusion model, in step 4, the color distance map of different wear states of all tools obtained in step 3 is used as the original input x0 of the DDPM diffusion model; in the forward process of the DDPM diffusion model, Gaussian noise is sampled from the standard normal distribution and gradually added to the sample with the diffusion step number tt until the sample becomes a pure Gaussian noise. The process is as follows:

[0030]

[0031] And α tt =1-β tt , β tt As the time step tt increases from 0.01 to 0.2, ε ttis the noise sampled from a standard normal distribution;

[0032] In the reverse process, a U-Net neural network is used to predict the noise added by each step tt in the forward process, so as to gradually denoise the pure Gaussian noise. The process is as follows:

[0033]

[0034] Where: Q(x tt-1 |x tt ) represents the distribution of the reverse process, which obeys the normal distribution; Represents the mean of the normal distribution; generates pseudo samples of each type of wear state that are similar to the distribution of the original input x0.

[0035] In the aforementioned tool wear state monitoring method based on the DDPM diffusion model, the tool wear state includes five states: initial wear, slight wear, stable wear, severe wear and failure; the training set includes five imbalance conditions in which the ratios between the five states of initial wear, slight wear, stable wear, severe wear and failure are 200:1:1:1:1:1, 100:1:1:1:1:1, 50:1:1:1:1:1, 20:1:1:1:1:1 and 10:1:1:1:1:1 respectively.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] The present invention uses a deep learning model and image processing technology. The method can more accurately monitor the tool wear state, issue a timely warning, and avoid product quality degradation or processing accidents caused by tool wear. The present invention optimizes signal processing parameters, improves the efficiency and accuracy of data processing, and reduces unnecessary computing resource consumption by using the multi-scale permutation entropy principle. The present invention generates pseudo samples through the DDPM diffusion model, solves the problem of unbalanced data sets, enhances the recognition ability of the model on minority class samples, and improves the overall generalization performance, especially when processing unbalanced data sets. It is excellent and can effectively avoid model deviations. The present invention inputs the balanced training set and test set into the Resnet18 network, and uses its powerful feature extraction ability to classify and identify the tool wear state. The Resnet18 network can learn deeper feature representations through multiple layers of convolution and fully connected layers, thereby improving the accuracy of classification. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0039] Figure 2 It is the MPE optimal parameter calculation process of the present invention;

[0040] Figure 3 This is a schematic diagram of the dimensional expansion of a one-dimensional time series signal according to the present invention;

[0041] Figure 4 This is a comparison diagram of the wear state images of the present invention;

[0042] Figure 5 A schematic diagram for generating a sample of the DDPM diffusion model of the present invention;

[0043] Figure 6 This is a comparison diagram of the confusion matrix before and after the DDPM diffusion model of the present invention is balanced;

[0044] Figure 7 It is a line graph of the accuracy of different methods under different imbalance rates of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but they are not intended to limit the present invention.

[0046] Embodiment: A tool wear state monitoring method based on DDPM diffusion model, such as Figure 1-7 This example takes the end mill as an example, and the overall flow chart is as follows Figure 1 As shown, the specific steps are as follows:

[0047] Step 1: Collect cutting force signals and corresponding tool wear status during milling;

[0048] First, a tool milling experiment was carried out. Each tool started the milling experiment from a brand new milling cutter until the tool failed. After each milling stage, the tool was removed and the wear condition of the tool was measured under a GP-300C optical microscope. The wear image of the tool was retained and the tool wear status was divided into the following five categories according to the wear trend: initial wear, slight wear, stable wear, severe wear, and failure.

[0049] The signal acquisition system, three-component dynamometer, three-axis accelerometer and other equipment are used to collect the cutting force signals of the workpiece on three channels (X, Y, Z) under five tool states (which can be divided into initial wear, slight wear, stable wear, severe wear and failure according to the wear area) and the corresponding real samples of tool wear states. The sampling frequency is 12kHz.

[0050] Step 2: Determine the optimal parameters for cutting force signal processing by minimizing the multi-scale permutation entropy principle, and then use the optimal parameters to obtain the phase space matrix;

[0051] A suitable length is selected for the collected one-dimensional cutting force signals of each channel, and the optimal embedding dimension m*, optimal time delay t* and optimal coarse-graining coefficient S* corresponding to each segment of the signal are found by minimizing the multi-scale permutation entropy (MPE) principle. The optimal parameters are determined by minimizing the multi-scale permutation entropy principle by first coarsening the given time series to obtain a new coarse-grained time series, then reconstructing the coarse-grained time series, arranging the reconstructed coarse-grained time series in ascending order, calculating the probability of each symbol sequence, normalizing and calculating the permutation entropy of each coarse-grained time series, and finally minimizing the process.

[0052] Specifically, the time series coarse-graining process is to give a time series X = {x1, x2, ..., x N}Perform coarse-graining processing to obtain a coarse-grained time series Y (S) ={y1 (S) ,y2 (S) ...,y T (S)},Right now:

[0053]

[0054]

[0055] Where: N represents the length of time series data, S represents the coarse-grained coefficient, and T represents the coarse-grained length of the sequence.

[0056] Specifically, time series reconstruction is to obtain the coarse-grained time series y i (s) Refactoring:

[0057]

[0058] Where: k is the kth reconstruction component, m is the embedding dimension, and t is the delay time; it can be expanded to an m-dimensional phase space matrix with embedding dimension m and time delay t:

[0059]

[0060] Specifically, the probability calculation is to sort the reconstructed coarse-grained time series in ascending order, and each The symbol sequence S(r) = {j1, j2, ..., j m}(r=1,2,...,k,k≤m!), calculate the probability p of each symbol sequence r The permutation entropy calculation is to calculate the permutation entropy value of each coarse-grained time series by normalization to obtain the multi-scale permutation entropy H p :

[0061]

[0062] Specifically, the minimization process is to construct an optimization model that minimizes the square of the MPE skewness:

[0063]

[0064] Where: SKEW represents the square of the skewness of the MPE of all real samples, q represents the real sample, Q is the size of all real samples q, SK q represents the MPE skewness of the real sample q, H q represents the MPE of the real sample q, H avg represents the average MPE sample of all real samples q; the mathematical optimization model of Min SKEW described in step 6 obtains the optimal parameters m*, t* and S* through intelligent evolutionary algorithms (such as genetic algorithms and particle swarm algorithms, etc.), and then transforms the phase space matrix Y * :

[0065]

[0066] in, make Represents Y * The i-th state in .

[0067] The collected channel signals are expanded, and the optimal parameters m*, t*, S* are obtained by minimizing the MPE skewness square optimization model, as shown in Table 1:

[0068] Table 1 MPE parameters of three-channel signals

[0069]

[0070] Step 3: Convert the cutting force signal into a grayscale image according to the phase space matrix, and superimpose the grayscale image to generate a color distance map;

[0071] Calculate the Euclidean distance between various wear states in the phase space matrix The phase space matrix is ​​encoded into a grayscale image, and then the grayscale distance maps of each channel are superimposed to generate a color distance map. Figure 3 A schematic diagram showing the conversion of the entire one-dimensional time series signal into an image.

[0072] Step 4: Use the color distance map as the original input to the DDPM diffusion model to generate pseudo samples for each type of wear state;

[0073] DDPM (Denoising Diffusion Probabilistic Models) is a generative model based on the diffusion process, which simulates the data generation process by gradually adding and removing noise. The core idea of ​​DDPM is to learn a noise predictor that can predict and remove noise in the reverse diffusion process, thereby recovering the original data from random noise. This model has shown powerful capabilities in tasks such as image generation, image restoration, and image super-resolution.

[0074] According to the color distance map of all tool wear states obtained in step 3, it is used as the original input x0 of the DDPM diffusion model, and the diffusion step number tt=10000, β tt As the time step tt increases, it increases linearly from 0.01 to 0.2. The last four types of wear state samples are generated through the forward denoising process and the reverse denoising process of DDPM to generate pseudo samples of each type of wear state. The relationship between the five types of tool wear states and their corresponding color distance maps and the last four types of wear state pseudo samples generated by the DDPM diffusion model are shown in the attached figure. Figure 4 shown.

[0075] In the forward process of the DDPM diffusion model, Gaussian noise is sampled from the standard normal distribution and gradually added to the sample with the diffusion step number tt until the sample becomes a pure Gaussian noise. The process is as follows:

[0076]

[0077] And α tt =1-β tt , β tt As the time step tt increases from 0.01 to 0.2, ε tt is the noise sampled from a standard normal distribution;

[0078] In the reverse process, a U-Net neural network is used to predict the noise added by each step tt in the forward process, so as to gradually denoise the pure Gaussian noise. The process is as follows:

[0079]

[0080] Where: Q(t tt-1 |x tt ) represents the distribution of the reverse process, which obeys the normal distribution; represents the mean of the normal distribution; generates pseudo samples with a distribution similar to the original input x0, and the schematic diagram of the DDPM diffusion model of the entire color distance map is as follows Figure 5 shown.

[0081] Step 5: Divide the color distance map into several unbalanced training sets and balanced test sets, and supplement the training set with the generated pseudo samples to balance the training set;

[0082] The color distance map of the real sample is set as the training set of 5 types of imbalanced conditions, initial wear is used as the majority class sample, slight wear, stable wear, severe wear and failure are used as minority class samples, and the imbalance ratios of initial wear, slight wear, stable wear, severe wear and failure are 1:200, 1:100, 1:50, 1:20, 1:10 respectively. The number of samples of each wear state in the test set is the same. After generating pseudo samples of the color distance map of all tool wear states in step 4, the pseudo samples generated by the DDPM diffusion model are supplemented to the minority class samples of the training set under the imbalance ratio for data enhancement, and a balanced training set with mixed true and false samples is obtained. Specifically, the details of the training set and test set samples are shown in Table 2:

[0083] Table 2 Number of samples in training set and test set

[0084]

[0085]

[0086] Step 6: Use the training set and test set to train and test the Resnet18 network respectively, and then use the trained Resnet18 network to classify and identify the tool wear status.

[0087] ResNet18 is a deep convolutional neural network (CNN) model that achieves this by introducing so-called residual blocks and skip connections, allowing the network to be trained deeper while maintaining good performance. Due to its excellent performance and shallow network depth, the ResNet18 network is widely used in computer vision tasks, including image classification, object detection, and semantic segmentation.

[0088] The divided training set and test set are input into the Resnet18 network for training and testing respectively, and then go through the convolution layer, batch normalization, and RELU activation function. After extracting the features of all samples in the training set, they enter the fully-connected layer for classification and recognition.

[0089] Comparative Example: Figure 6The confusion matrix comparison of the classification of five tool wear states (initial wear, slight wear, stable wear, severe wear and failure) without and with the DDPM diffusion model data enhancement of this method is shown respectively when the imbalance rate is 1:200. The average classification accuracy of the five categories is 64.4% and 91.1%, and the accuracy is significantly improved after using this method. DCGAN is a deep convolutional generative adversarial network, which is improved on the basis of the traditional generative adversarial network (GAN) to improve the stability of the model and the quality of the generated image. The key improvements of DCGAN include the use of a fully convolutional network structure, the removal of the maximum pooling layer in the traditional GAN, the use of stride convolution and deconvolution to achieve up and down sampling, and the application of batch normalization in the generator and discriminator. SAGAN is a further developed generative adversarial network, which introduces the self-attention mechanism on the basis of DCGAN. The self-attention mechanism allows the network to consider the long-range dependencies of different positions in the image when generating images, which helps the model capture the global contextual information of the image, thereby generating more consistent and refined images. The self-attention module of SAGAN is complementary to the traditional convolution operation. It can directly model the long-range dependencies in the feature map, improving the ability of image generation. Comparing the classification accuracy after balancing using DCGAN and SAGAN networks under different imbalanced conditions, it can be seen from Table 3 that the classification accuracy of the method proposed in the present invention is 9.6% to 26.7% higher than that of the imbalanced samples and the use of DCGAN and SAGAN for balancing.

[0090] Table 3 Comparison of the accuracy of different methods in the experiment

[0091] Unbalanced situation 200:1:1:1:1 100:1:1:1:1 50:1:1:1:1 20:1:1:1:1 10:1:1:1:1 Sample imbalance 64.4％ 67.5％ 72.4％ 76.4％ 79.9％ DCGAN 72.3％ 74.2％ 79.6％ 80.5％ 82.4％ SAGAN 74.8％ 75.7％ 79.8％ 83.6％ 84.5％ DDPM Diffusion Model 91.1％ 92.4％ 92.9％ 93.3％ 94.1％

[0092] In order to more intuitively demonstrate the diagnostic effect of this method, the test accuracy inflection point diagram of ResNet18 under different imbalance conditions and different generation methods is drawn, as shown in the figure: Figure 7 As shown in the figure, we can see that as the proportion of real samples increases, no matter which sample generation method is used, its accuracy is significantly improved, and the sample generation method using the DDPM diffusion model has a much higher accuracy than other generation methods regardless of the imbalance ratio.

[0093] The present invention uses a deep learning model and image processing technology. The method can more accurately monitor the tool wear state, issue a timely warning, and avoid product quality degradation or processing accidents caused by tool wear. The present invention optimizes signal processing parameters, improves the efficiency and accuracy of data processing, and reduces unnecessary computing resource consumption by using the multi-scale permutation entropy principle. The present invention generates pseudo samples through the DDPM diffusion model, solves the problem of unbalanced data sets, enhances the recognition ability of the model on minority class samples, and improves the overall generalization performance, especially when processing unbalanced data sets. It is excellent and can effectively avoid model deviations. The present invention inputs the balanced training set and test set into the Resnet18 network, and uses its powerful feature extraction ability to classify and identify the tool wear state. The Resnet18 network can learn deeper feature representations through multiple layers of convolution and fully connected layers, thereby improving the accuracy of classification.

[0094] In summary, the present invention can achieve high-precision tool status monitoring and accurate classification.

Claims

1. A tool wear state monitoring method based on DDPM diffusion model, characterized in that: The specific steps include: Step 1: Collect cutting force signals and corresponding tool wear status during milling; Step 2: Determine the optimal parameters for cutting force signal processing by minimizing the multi-scale permutation entropy principle, and then use the optimal parameters to obtain the phase space matrix; the optimal parameters include the optimal embedding dimension m*, the optimal time delay t* and the optimal coarse-graining coefficient S*; the optimal parameters determined by minimizing the multi-scale permutation entropy principle are to first perform a time series coarse-graining process on a given time series to obtain a new coarse-grained time series, and then reconstruct the coarse-grained time series, arrange the reconstructed coarse-grained time series in ascending order, calculate the probability of each symbol sequence, normalize and calculate the permutation entropy of each coarse-grained time series, and finally perform minimization processing; the time series coarse-graining process is to perform a time series coarse-graining process on a given time series X={x1, x2, ..., x N }Perform coarse-graining processing to obtain a coarse-grained time series Y (S) =Py1 (S) , y2 (S) , ..., y T (S) },Right now: Where: N represents the length of the time series data, S represents the coarse-grained coefficient, and T represents the coarse-grained length of the sequence; the time series reconstruction is to obtain the coarse-grained time series y i (s) To refactor: Where: k is the kth reconstruction component, m is the embedding dimension, and t is the delay time; Step 3: Convert the cutting force signal into a grayscale image according to the phase space matrix, and superimpose the grayscale image to generate a color distance map; Step 4: Use the color distance map as the original input to the DDPM diffusion model to generate pseudo samples for each type of wear state; Step 5: Divide the color distance map into several unbalanced training sets and balanced test sets, and supplement the training set with the generated pseudo samples to balance the training set; Step 6: Use the training set and test set to train and test the Resnet18 network respectively, and then use the trained Resnet18 network to classify and identify the tool wear status.

2. The tool wear state monitoring method based on the DDPM diffusion model according to claim 1 is characterized in that: The probability calculation is to arrange the reconstructed coarse-grained time series in ascending order, and each The symbol sequence S(r) = {j1, j2, ..., j m }(r=1,2,...,k,k≤m!), calculate the probability p of each symbol sequence r The permutation entropy calculation is to calculate the permutation entropy value of each coarse-grained time series by normalization to obtain the multi-scale permutation entropy H p :

3. The tool wear state monitoring method based on the DDPM diffusion model according to claim 2 is characterized in that: The minimization process is to construct an optimization model that minimizes the square of the MPE skewness: Where: SKEW represents the square of the skewness of the MPE of all real samples, q represents the real sample, Q is the size of all real samples q, SK q represents the MPE skewness of the real sample q, H q represents the MPE of the real sample q, H avg represents the average MPE sample of all true samples q.

4. The tool wear state monitoring method based on the DDPM diffusion model according to claim 3 is characterized in that: The optimal parameters m*, t* and s* are obtained by minimizing the optimization model of the MPE skewness square; based on the obtained optimal parameters m*, t* and s*, the phase space matrix is ​​further expressed as: in, make Represents Y * The i-th state in .

5. The tool wear state monitoring method based on the DDPM diffusion model according to claim 4 is characterized in that: In step 3, the Euclidean distance between various wear states in the phase space matrix is ​​calculated The phase space matrix is ​​encoded as a grayscale image, and then the grayscale distance map is superimposed to generate a color distance map.

6. The tool wear state monitoring method based on the DDPM diffusion model according to claim 1 is characterized in that: In step 4, the color distance map of all tool wear states obtained in step 3 is used as the original input x0 of the DDPM diffusion model. In the forward process of the DDPM diffusion model, Gaussian noise is sampled from the standard normal distribution and gradually added to the sample with the diffusion step number tt until the sample becomes a pure Gaussian noise. The process is as follows: And α tt =1-β tt , β tt As the time step tt increases from 0.01 to 0.2, ε tt is the noise sampled from a standard normal distribution; In the reverse process, a U-Net neural network is used to predict the noise added by each step tt in the forward process, so as to gradually denoise the pure Gaussian noise. The reverse process is derived according to the Bayesian formula and reparameterization technique as follows: Where: Q(x tt-1 |x tt ) represents the distribution of the reverse process, which obeys the normal distribution; Represents the mean of the normal distribution; generates pseudo samples of each type of wear state that are similar to the distribution of the original input x0.

7. The tool wear state monitoring method based on the DDPM diffusion model according to claim 1 is characterized in that: The tool wear states include five states: initial wear, slight wear, stable wear, severe wear and failure; the training set includes five imbalance conditions in which the ratios between the five states are 200:1:1:1:1:1, 100:1:1:1:1:1, 50:1:1:1:1:1, 20:1:1:1:1:1 and 10:1:1:1:1:1 respectively.

Images