Fault diagnosis data enhancement method based on diffusion model and generative adversarial network

By introducing ADGAN into the fault diagnosis model, using diffusion model and adaptive data enhancement method of generating adversarial networks, the stability problem of GAN model under unbalanced data and the problem of low diffusion model generation efficiency are solved, and higher quality is achieved and diverse sample generation improves the accuracy and stability of fault diagnosis.

CN119066403APending Publication Date: 2024-12-03NORTHEAST FORESTRY UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411148722.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

GAN-based fault diagnosis models are prone to pattern collapse or gradient disappearance under unbalanced data or data enhancement conditions, and the reverse generation process of traditional diffusion models is slow, affecting the fault diagnosis effect.

Method used

Adaptive data augmentation method (ADGAN) based on diffusion model and generative adversarial networks are adopted, and the diffusion walk length is dynamically adjusted by introducing a stepwise accumulation noise mechanism that simulates the diffusion process in the generator, and using a hybrid distribution training discriminator and generator to dynamically adjust the diffusion walk length to Improve the quality and diversity of sample generation.

Benefits of technology

It effectively avoids pattern collapse and gradient disappearance, significantly improves the quality and diversity of generated samples, improves the accuracy of fault diagnosis and model stability, and reduces the demand for computing resources.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066403A_ABST
    Figure CN119066403A_ABST
Patent Text Reader

Abstract

The invention discloses a new fault diagnosis data enhancement method, namely an ADGAN which is a generative adversarial network of an adaptive diffusion model, and the ADGAN generates Gaussian mixture distribution noise through a forward diffusion chain, so that the gradient is smoother and more stable, and the problems of gradient disappearance and the like in a traditional GAN model are avoided. The ADGAN learns and distinguishes diffusion real data and generated data on each diffusion time step length by using discriminators related to different noise data proportions and time steps, the generator learns from feedback of the discriminators through back propagation of a forward diffusion chain, the length of the forward diffusion chain can be adaptively adjusted, the diversity of the generator is improved, and the method is suitable for large-scale popularization and application. And the problems of mode collapse and slow reverse generation process of the traditional diffusion model are effectively avoided. According to the invention, by using a bearing data set of the Kassi University and a data set acquired by an NEFU data set acquisition test bed, it is verified that ADGAN is excellent in data generation effect and stability, the problem of fault samples in an unbalanced data set can be effectively solved, the fault diagnosis performance is significantly improved, and the fault diagnosis efficiency is improved. And a new solution is provided for data enhancement and fault diagnosis.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of fault diagnosis and data enhancement, and in particular to an adaptive data enhancement method based on a diffusion model and a generative adversarial network, referred to as ADGAN, which is used to solve the problem of sample data imbalance in industrial equipment fault diagnosis. Background Art

[0002] Traditional data augmentation methods such as oversampling and undersampling often fail to capture complex patterns in the data. Generative adversarial networks (GANs) and diffusion models have become powerful tools, but still suffer from stability issues and computational inefficiencies;

[0003] In actual production, it is easier to obtain samples in a healthy state, while the scarcity of fault samples leads to sample imbalance, that is, the number of normal data samples is far greater than that of abnormal or faulty data samples. This imbalance will affect the performance of the model, causing it to tend to classify samples into the majority class (i.e., normal class), making it difficult to accurately identify the minority class (i.e., faulty class). To address the problem of sample imbalance, current research focuses on innovative strategies for data enhancement;

[0004] In recent years, scholars have made many studies on data enhancement models, for example: S. Shao et al. proposed an ACGAN framework for data generation and enhancement; X. Gao et al. proposed a Wasserstein generative adversarial network (WGAN-GP) based on gradient penalty to generate data samples; J. Wang et al. proposed a new fault diagnosis method combining conditional generative adversarial network (CGAN) and stacked autoencoder, generating multi-category fault signals through spectrum normalization and category label conditional information, and using stacked autoencoder for feature extraction and classification; H. Wang et al. proposed a Fourier transform GAN (FTGAN) method, which improves the quality of synthetic data by approximating Fourier transform through autoencoder, and generates bearing fault signals close to real data in time-frequency domain. However, the above GAN-based related models are very prone to mode collapse or gradient disappearance under the conditions of unbalanced data or data enhancement. The reasons include unstable confrontation between the generator and the discriminator, gradient propagation problems in complex model structures, and the insufficiency of the model in generating diversified samples, which affect the effect of fault diagnosis;

[0005] The diffusion model uses a forward noise-adding, reverse-generated diffusion chain to train data samples. Since the entire reverse generation process requires multiple iterations of the same generator network based on the U-Net architecture, each step aims to gradually remove the noise introduced in the previous stage to restore the original data. This repeated traversal significantly prolongs the time required to generate data and places great demands on computing resources. Summary of the invention

[0006] The purpose of the present invention is to provide a fault diagnosis data enhancement method based on a diffusion model and a generative adversarial network to solve the problem that the GAN-based related model is very prone to mode collapse or gradient disappearance under the conditions of unbalanced data or data enhancement, and at the same time solve the problem that the reverse generation process of the traditional diffusion model is slow.

[0007] The technical solution of the present invention is as follows:

[0008] Step 1: Preprocess the original data to obtain training samples for the model;

[0009] Step 2: Use the training samples obtained in step 1 as a data set to train the ADGAN model;

[0010] The ADGAN framework consists of a generator and a discriminator;

[0011] Compared with the generator in the traditional GAN ​​model, the generator adds a mechanism for gradually accumulating noise during the simulated diffusion process;

[0012] The generator receives a random vector z~p(z) as input and generates a fault sample x g =G(z);

[0013] The generator gradually transforms random vectors into Gaussian noise-distributed data through a forward diffusion chain;

[0014] Through the inverse diffusion chain, the generator restores the noisy data into realistic fault samples;

[0015] The discriminator is used to learn to distinguish the real sample data and the generated sample data after diffusion at each time step t, and receives the generated sample and the real sample at each time step for discrimination;

[0016] The discriminator extracts sample features through a convolutional network and outputs the authenticity score of the sample;

[0017] The discriminator at each time step is learned based on the noisy data and the real data to distinguish between the generated data and the real data;

[0018] Step 3: Using the Case Western Reserve University (CWRU) dataset and the laboratory self-test dataset (ENFU dataset), a large number of experiments were conducted based on the convolutional neural network (CNN) and long short-term memory network (LSTM) models to verify the superiority of the ADGAN model;

[0019] Furthermore, the step 1 includes the following steps:

[0020] First, a sliding window is performed on the original vibration signals of different health states in the relevant dataset to obtain samples;

[0021] Then, continuous wavelet transform (CWT) is performed on the sample to obtain a two-dimensional image;

[0022] Furthermore, the step 2 includes the following steps:

[0023] First, the generator receives the random vector z~p(z) as input and generates the sample x g =G(z);

[0024] Then, ADGAN simulates the mechanism of gradually accumulating noise during the diffusion process. At each diffusion step t, it continuously compares the real sample x and the generated sample x g Perform diffusion noise processing to obtain diffusion data and

[0025] Next, the discriminator learns to distinguish the data at each time step t and

[0026] Finally, the generator learns from the feedback of the discriminator through back-propagation and adjusts its parameters to make the generated samples closer to the real data;

[0027] In order to ensure the quality and diversity of generated samples, ADGAN obtains corresponding noise samples in any diffusion step t by defining a mixed distribution:

[0028]

[0029]

[0030] Among them, π t represents the mixing weight at step t, Indicates that given the original image x and time step t, the noise sample The conditional probability distribution of Indicates that given the original image x g and time step t, the noise sample The conditional probability distribution of ;

[0031] The mixed distribution can be used to obtain real samples and generated samples with different noise levels;

[0032] As the number of diffusion steps increases, the accumulation of noise in the generated samples leads to the gradual loss of information in the original signal. The diffusion-induced mixed distribution is used to train the time-step-dependent discriminator and generator. The distribution of the generated noise samples matches the distribution of the real noise samples, which can improve the quality and diversity of sample generation.

[0033] Furthermore, the training objective of the model is defined as:

[0034]

[0035] in, Representation sample The parameterized function D of the discriminator at diffusion step t φ The logarithm of the probability of being judged as a true sample, Indicates the use of generator G θ Generate samples at diffusion step t The discriminator D φ The logarithm of the probability of generating a sample;

[0036] The objective function optimizes the adversarial training of the generator and the discriminator, so that the generator can generate samples close to the real data distribution, while the discriminator can effectively distinguish between real samples and generated samples, thereby promoting the model to generate more realistic and diverse samples;

[0037] Furthermore, in order to ensure that the complexity of the model is moderate and avoid computational redundancy, the present invention dynamically adjusts the maximum diffusion step length T to achieve adaptive diffusion step length adjustment. The specific steps are as follows:

[0038] In each training iteration, the loss of the current diffusion chain is calculated, and the maximum diffusion step size T is dynamically adjusted according to the change in loss;

[0039] If the loss decreases quickly, increase the step size T so that the model can explore the data distribution more deeply; if the loss decreases slowly or fluctuates greatly, reduce the step size T to avoid overfitting or over-calculation;

[0040] The implementation formula for adaptively modifying the diffusion step size T is as follows, which adaptively modifies the maximum diffusion step size T:

[0041]

[0042] where d target It is used to control the degree of fit of the discriminator to the data. C represents the amount of change each time T is updated. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The main structure of the diffusion model

[0044] Figure 2 The main structure of GAN

[0045] Figure 3 Figure 2. ADGAN data enhancement process diagram

[0046] Figure 4Data acquisition testbed for CWRU

[0047] Figure 5 The inner ring fault frequency domain of the CWRU dataset with different data enhancement methods

[0048] Figure 6 CNN and LSTM model structure diagram

[0049] Figure 7 Comparison of indicators of CNN models using different data augmentation methods for the CWRU dataset

[0050] Figure 8 Comparison of indicators of LSTM model using different data augmentation methods for CWRU dataset

[0051] Fig. 9 T-SNE diagram of the LSTM model for the CWRU dataset when using different data augmentation methods

[0052] Fig.10 Acquisition of a testbed for the NEFU dataset

[0053] Fig.11 Four types of faulty bearings for NEFU: (a) Outer raceway fault. (b) Inner raceway fault. (c) Rolling element failure. (d) Composite fault of inner and outer rings.

[0054] Fig.12 The inner circle fault frequency domain of the NEFU dataset with different data enhancement methods

[0055] Fig.13 Comparison of indicators of CNN models using different data enhancement methods for the NEFU dataset

[0056] Fig.14 Comparison of indicators of LSTM model using different data enhancement methods for NEFU dataset

[0057] Fig.15 Confusion matrix of CNN model using different data augmentation methods. DETAILED DESCRIPTION

[0058] like Figure 3 As shown in the figure, the implementation process of a fault diagnosis data enhancement method based on diffusion model and generative adversarial network is as follows: first, the data is preprocessed, and then the processed data is used as the input for training ADGAN model to realize full supervised training of the network. Finally, the effect of ADGAN model is verified by two sets of data sets.

[0059] Further, step one includes the following steps:

[0060] First, a sliding window is performed on the original vibration signals of different health states in the relevant dataset to obtain samples;

[0061] Then, continuous wavelet transform (CWT) is performed on the sample to obtain a two-dimensional image;

[0062] Finally, the two-dimensional data samples after wavelet transformation are used as data sets to train the ADGAN model and obtain a large number of generated samples;

[0063]

[0064] Furthermore, the process of step 2 is shown in the table above:

[0065] Before the set number of training times is reached, the following operations are performed in each cycle:

[0066] 1. Training initialization: set the number of iterations, learning rate, and batch size of the training model;

[0067] 2. Training loop: When 'epoch' is less than the set number of iterations, the following two steps are performed:

[0068] Step 1: Update the discriminator:

[0069] First, collect noise samples: sample n noise vectors {z1, z2, ..., z n}, that is, p(z);

[0070] Then generate samples: use the generator to receive the noise vector and generate samples {x g,1 , x g,2 , ..., x g,n}, that is, x g,i =G(z i );

[0071] Sample n real samples {x1, x2, ..., x n};

[0072] For each sample i (from 1 to n), the intermediate sample is sampled by the diffusion process and As samples for adversarial training;

[0073] The loss function is maximized to update the parameters of the discriminator so that it can distinguish between real samples and generated samples.

[0074] Step 2: Update the generator:

[0075] First, noise samples are sampled again in the pre-trained dataset, and n noise vectors {z1, z2, ..., z n};

[0076] Then generate samples, use the generator to receive the noise vector to generate samples {x g,1 , x g,2 , ..., x g,n}, that is, x g,i =G(z i );

[0077] Time step sampling, randomly select time steps {t1, t2, ..., t n}.

[0078] Step 3: Update the diffusion process;

[0079] Adjustment of diffusion step size:

[0080] Diffusion step size adjustment is performed once every k epochs;

[0081] Calculate the updated value of the current diffusion step size T, the formula is as follows:

[0082]

[0083] in, represents the expected value of the diffusion step length on the real data and the generated data, d target is the target step size, C is a constant used to control the speed of step size adjustment;

[0084] The training cycle ends when:

[0085] When the number of training iterations reaches the maximum set value eopch, the training cycle ends;

[0086] Summary: The training process of ADGAN alternately updates the parameters of the discriminator and the generator and dynamically adjusts the diffusion step size to ensure that the generated samples gradually approach the distribution of the real data, thereby improving the quality of the generated samples and the stability of the model.

[0087] Implementation verification example 1:

[0088] A public dataset provided by CWRU was used to evaluate the diagnostic effectiveness of the proposed method. Figure 4 The test bench shown is collected.

[0089] The bearing dataset contains three fault types: inner race fault (IF), outer race fault (OF) and ball fault (BF). Each fault type contains three fault sizes with diameters of 7 mils, 14 mils and 21 mils. Therefore, the bearing dataset contains one normal state and nine fault states. The state labels are shown in the following table:

[0090]

[0091] For the original vibration signals of ten different health states, 400 sliding window samplings with a length of 1024 were performed, and the vibration signals were converted into 400 two-dimensional images of 64×64 pixels through wavelet transform;

[0092] Randomly sample 400 fault samples with 9 different labels with label values ​​of 1-9, of which 32 samples are used for training the data enhancement model, and the remaining samples will be used for subsequent quality evaluation;

[0093] In order to more intuitively illustrate the data enhancement effect of the model, the most mainstream data enhancement models DDPM, VAE_GAN, WGAN_GP, and VAE are used to perform data enhancement comparative training with the same number of traversals to generate 9×120 sample data for a more fair quality evaluation;

[0094] Quality evaluation:

[0095] The frequency domain analysis of the samples was performed using Fourier transform, and the performance of different generative models was intuitively quantified by comparing the subtle features of the data in the frequency dimension;

[0096] The experimental results are as follows Figure 5 As shown in Figure (a), the blue line plots the real frequency domain signal, and Figures (b) to (f) respectively plot the frequency domain signals generated by the DDPM, VAE_GAN, WGAN_GP, VAE and ADGAN model data enhancement. The red line represents the frequency domain signal generated by each model. It can be clearly seen that the red line generated by the ADGAN model has a higher degree of overlap with the blue line in terms of shape, amplitude and frequency distribution, that is, the data generated by it is closer to the real sample;

[0097] This method also uses the feature quantitative evaluation index J to quantify the quality of generated data by comparing the intra-class covariance and inter-class covariance. The specific method is as follows:

[0098]

[0099]

[0100]

[0101]

[0102] Among them, S ω represents the within-class covariance matrix, S b represents the inter-class covariance matrix, C represents the number of categories, N C represents the number of samples in the cth class, f i c represents the feature vector of the i-th sample in the c-th class, represents the average eigenvector of the cth class, represents the average feature vector of all samples;

[0103] Within-class covariance S ω It reflects the closeness of data points in the same category, and the inter-class covariance S b It reflects the degree of separation between data points of different categories;

[0104] Therefore, when the J value is large, it means that the inter-class covariance is larger than the intra-class covariance, that is, the data points of different categories have a high degree of separation, while the data points in the same category are relatively close, indicating that the generated data set can form a good distinction between different categories while maintaining consistency within the category. When the J value is small, it means that the ratio of the inter-class covariance to the intra-class covariance is small, that is, the separation between the data points of different categories is not obvious enough, or the data points in the same category are too scattered, indicating that the generated data set is not clear enough in category distinction, or there are more noise and outliers;

[0105] Case 1 uses different data augmentation methods to generate data. The J value of the proposed method is the highest among all the methods in the five models.

[0106]

[0107] In order to evaluate the effect of generated data in actual fault diagnosis, this method uses two deep learning models, CNN and LSTM, to achieve accurate evaluation of the quality of generated data;

[0108] In order to simulate data imbalance and training after data enhancement in fault diagnosis, the remaining 400 samples in normal health state after data partitioning, 9×105 real samples after random sampling, and 9×120 generated samples are divided as follows:

[0109] 300 samples are randomly selected from the 400 samples in normal health status as training sets, and 100 samples are used as test sets; 5 samples from the 105 real samples of 9 different faults are used as training sets for data imbalance training, and 100 samples are used as test sets; all 120 generated samples of 9 different faults are used as training sets after data enhancement;

[0110] After completing the data division, we use CNN and LSTM deep learning models to verify the model structure. Figure 6 As shown in the figure, the training parameters are unified as follows: batch size is 128, learning rate is 0.0001, and epoch is 100;

[0111] The average accuracy of the last 50 epochs after the model converges is selected as the single accuracy record, and then the same experiment is performed multiple times to obtain the average value;

[0112] In addition to the accuracy index, this paper also uses three evaluation indicators: precision, recall and F1 value to reflect the actual performance of the model;

[0113] The experimental results obtained by training on CNN and LSTM models are as follows Figure 7 and Figure 8 As shown, 'None' represents the experimental setting with a data imbalance ratio of 300:5;

[0114] pass Figure 7 and Figure 8 It can be seen that the five different data augmentation methods generate more fault samples, enabling the model to learn richer and more comprehensive fault features, thereby improving the generalization ability and stability of the model;

[0115] Among the five different data augmentation methods, ADGAN achieved the best performance in all four indicators of the two models, indicating that ADGAN is significantly superior to other data generation methods in terms of the quality of generated data, with higher diversity and balance;

[0116] In order to more intuitively demonstrate the ability of different data augmentation methods to distinguish different categories in fault diagnosis, the t-distributed stochastic neighbor embedding (t-SNE) method is used to visualize the target domain output data processed by the Convformer-SENet model, as shown in the figure. Fig. 9 ;

[0117] Observe the T-SNE graph ( Fig. 9 ) It can be clearly found that the clustering diagram without data enhancement shows that more than half of the data is loosely distributed, and the boundaries are not clear and are in a chaotic state. After using different data enhancement methods, the clustering effect of each category is effectively improved, and the classification ability of the LSTM model is enhanced. Among them, after using ADGAN for data enhancement, the clusters formed in the same category are the most compact, and the boundaries between different categories are also the clearest, which shows that the data enhancement method of ADGAN is the most effective;

[0118] Implementation verification example 2:

[0119] In order to further verify the effectiveness of the ADGA model, a gear fault data set was collected using the comprehensive test platform for rotating machinery fault implantation at Northeast Forestry University;

[0120]

[0121] For the original vibration signals in the five different health states shown above, 400 sliding window samplings with a length of 1024 were performed respectively, and the vibration signals were converted into 400 two-dimensional images of 64×64 pixels through wavelet transform;

[0122] Randomly sample 400 fault samples with four different labels with label values ​​of 1-4, of which 32 samples are used for training the data enhancement model, and the remaining samples will be used for subsequent quality evaluation;

[0123] In order to more intuitively illustrate the data enhancement effect of the model, the most mainstream data enhancement models DDPM, VAE_GAN, WGAN_GP, and VAE are used to perform data enhancement comparative training with the same number of traversals to generate 9×120 sample data for a more fair quality evaluation;

[0124] Quality evaluation:

[0125] The frequency domain analysis of the samples was performed using Fourier transform, and the performance of different generative models was intuitively quantified by comparing the subtle features of the data in the frequency dimension;

[0126] The experimental results are as follows Fig.12 As shown in Figure (a), the blue line plots the real frequency domain signal, and Figures (b) to (f) respectively plot the frequency domain signals generated by DDPM, VAE_GAN, WGAN_GP, VAE and ADGAN model data enhancement, and the red lines represent the frequency domain signals generated by each model;

[0127] By observing the overlap between the red and blue lines in Figures (b) to (f), it can be clearly seen that the red and blue lines generated by the ADGAN model have a higher overlap in shape, amplitude, and frequency distribution, that is, the data it generates is closer to the real samples;

[0128] By calculating the feature quantitative evaluation index J for the generated data, the quality of the generated data is quantified by comparing the intra-class covariance and inter-class covariance. The calculation results are shown in the following table;

[0129]

[0130] Among the five models, the proposed method ADGAN has the highest performance among all methods, indicating that the inter-class covariance is larger than the intra-class covariance, that is, the data points of different categories have a high degree of separation, and the data points in the same category are relatively close. The generated dataset can form a good distinction between different categories while maintaining the consistency within the category;

[0131] The remaining 400 samples in normal health status after data division, 4×110 real samples after random sampling, and 9×120 generated samples are divided as follows:

[0132] Among them, 300 samples of 400 normal healthy state samples were randomly sampled as training sets, and 100 samples were used as test sets; 10 samples of 110 real samples of 4 different faults were used as training sets for data imbalance training, 100 samples were used as test sets, and all 120 generated samples of 9 different faults were used as training sets after data enhancement;

[0133] After completing the data division, the CNN and LSTM deep learning models were used for verification using the same training parameters as in Case 1. The average accuracy of the last 50 epochs after the model converged was selected as the single accuracy record, and then the same experiment was performed multiple times to obtain the average value.

[0134] The experimental results obtained by training on CNN and LSTM models are as follows Fig.13 and Fig.14 As shown, 'None' represents the experimental setting with a data imbalance ratio of 300:10;

[0135] The experimental results show that in case 2, ADGAN performs best in all evaluation indicators on CNN and LSTM, which indicates that the data generated by ADGAN is significantly better than other methods in terms of quality, diversity, and balance. This once again confirms the superiority of ADGAN in the quality of generated data, and it has wide applicability. It is a very effective data enhancement method with broad application prospects and research value.

[0136] like Fig.15 As shown, the confusion matrix when different data augmentation methods are applied to the CNN model is shown;

[0137] By observing the confusion matrix, it can be clearly found that the confusion matrix without data enhancement shows a more vague classification effect;

[0138] After using data enhancement, the diagonal of the confusion matrix becomes clearer, indicating that different data enhancement methods can effectively improve the classification ability of the CNN model. Among them, after using ADGAN for data enhancement, the samples are classified into the real category with the best effect, which shows that the data enhancement method of ADGAN is the most effective.

Claims

1. The specific process of a fault diagnosis data enhancement method based on diffusion model and generative adversarial network is as follows: collect the vibration signal of the mechanical system, input the training data into the model, and realize the complete training process of the model. Comparative experiments using two data sets and multiple data enhancement methods are used to verify the superiority of the new model for data enhancement; A fault diagnosis data enhancement method based on diffusion model and generative adversarial network (ADGAN model) is obtained by the following method: Step 1: Obtain the original vibration signal of the bearing, and divide the original vibration signal of the bearing into a training set, a verification set and a test set; Step 2: Build the ADGAN framework and train the model; The ADGAN framework described mainly consists of two parts: the generator and the discriminator; Compared with the generator in the traditional GAN ​​model, the generator adds a mechanism for gradually accumulating noise during the simulated diffusion process; The generator receives a random vector z~p(z) as input and generates a fault sample x g =G(z); The generator gradually transforms random vectors into Gaussian noise-distributed data through a forward diffusion chain; Through the inverse diffusion chain, the generator restores the noisy data into realistic fault samples; The discriminator is used to learn to distinguish the real sample data and the generated sample data after diffusion at each time step t, and receives the generated sample and the real sample at each time step for discrimination; The discriminator extracts sample features through a convolutional network and outputs the authenticity score of the sample; The discriminator at each time step is learned based on the noisy data and the real data to distinguish between the generated data and the real data; Step 3: Using the Case Western Reserve University (CWRU) dataset and the laboratory self-test dataset, a large number of experiments were conducted based on the convolutional neural network (CNN) and long short-term memory network (LSTM) models to verify the superiority of the ADGAN model.

2. The method according to claim 1, characterized in that: The step 1 comprises the following steps: First, the sensor collects the vibration signal of the mechanical system and obtains some sample data; Then, sliding window sampling with non-overlapping sample windows is used to divide the data into training set, validation set and test set.

3. The method according to claim 1, characterized in that: ADGAN obtains the corresponding noise samples in any diffusion step t by defining a mixed distribution to ensure the quality of the generated samples: where π t represents the mixing weight at step t, Indicates that given the original image x and time step t, the noise sample The conditional probability distribution of Indicates that given the original image x g and time step t, the noise sample The conditional probability distribution of .

4. The method according to claim 1, characterized in that: The generator and discriminator in the ADGAN model compete with each other by optimizing the loss functions shown in Formulas 1 and 2, respectively: min G V(G,D)=E z~pz(z) [log(1-D(G(z)))] (3) max D V(G,D)=E x~p(x) [logD(x)]+Ε z~p(z) [log(1-D(G(z)))] (4) Among them, x is a sample from the true data distribution p(x); z is a noise sample from the prior distribution p(z); G(z) is the sample generated by the generator; D(x) is the probability estimate of the discriminator that sample x is the real data.

5. The method according to claim 1, characterized in that: The ADGAN model uses adaptive diffusion steps and perturbation functions to dynamically adjust the diffusion steps to ensure that the complexity of the model is moderate and adaptively modify the maximum diffusion step T: where d target It is used to control the degree of fit of the discriminator to the data, and C represents the amount of change each time T is updated.

6. The method according to claim 1, characterized in that: The experimental verification step uses the CWRU dataset and the NEFU gear dataset to perform unsupervised learning of the ADGAN model, and conducts data enhancement effect comparison experiments with other mainstream data enhancement models (including DDPM, VAE_GAN, WGAN_GP, VAE). The experimental verification further includes: fixing hyperparameters and repeating the experiment to verify the effectiveness and repeatability of the method; comparing from multiple angles, including confusion matrix, clustering graph and F1 score graph.

7. The method according to claim 6, characterized in that: The evaluation of data enhancement effect lies in quality evaluation, and the quality evaluation method is: The frequency domain graphs of the real data and generated data of each data augmentation method are plotted respectively, and the effect of data augmentation is judged by the degree of overlap between the real data and the generated data. The quality of the generated data is quantified by calculating the feature quantitative evaluation index J for the generated data and comparing the intra-class covariance and inter-class covariance. With the help of two deep learning models, CNN and LSTM, an accurate evaluation of the quality of the generated data is achieved.

Citation Information

Cited By

  • Fracturing truck intelligent diagnosis data generation method and diagnosis method

    CN120372294A

  • Electronic nose data generation method based on auto-encoder and generative adversarial network

    CN120579042A