Data imbalance fault diagnosis method based on diffusion model and depth separable convolution

Abstract

The invention provides a data imbalance fault diagnosis method based on a diffusion model and depth separable convolution, and relates to the field of electric digital signal processing, and the method comprises the steps: collecting a vibration time domain signal of a gear under a fault condition through an acceleration sensor, taking the vibration time domain signal as a gear fault signal, and constructing a data imbalance training set; establishing a gear fault signal generation model by adopting a diffusion model formed by an improved waveform signal generation network, inputting the gear fault signal into the gear fault signal generation model for training, and generating a virtual gear fault signal; performing quality evaluation on the virtual gear fault signal by adopting a signal filter, and merging the evaluated signal with an original training set to obtain an expanded data set; and constructing a gear fault classification model by adopting a deep separable convolution structure, training the gear fault classification model by using the expanded training set, finally obtaining a classification model for identifying gear fault types, and accurately classifying the gear fault types.

Description

Technical Field

[0001] The present invention relates to the technical field of data imbalance fault diagnosis, and more particularly, to a data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution. Background Art

[0002] As a core component in rotating machinery, gears play a crucial role in transmitting power and motion. During actual operation, gears are subjected to complex loadings, and their performance and condition are directly related to the reliability and stability of the entire rotating machinery. Once a gear fails, it not only leads to a decline in mechanical performance but may also trigger a shutdown failure, thereby affecting the safety and operating efficiency of the equipment. Therefore, effective fault diagnosis of gears to timely detect and address potential problems is of great significance for ensuring the normal operation of rotating machinery. With the continuous development of rotating machinery technology, the working environment and operating conditions of gears have become increasingly complex, which has increased the likelihood and risk of failures. Against this background, developing advanced gear fault diagnosis models to achieve effective condition monitoring and fault warning has become an important research direction.

[0003] Currently, most data-driven fault diagnosis frameworks for rotating machinery are mainly designed based on the assumption of data balance. These frameworks typically rely on a large amount of normal operation data and fault data to train classification models for accurate fault identification. To address the data imbalance problem, some methods adopt data generation models to augment the fault data set by generating virtual fault signals to increase the number of training samples, thereby improving the performance of the classification model. These methods have alleviated the impact of data imbalance on fault diagnosis to a certain extent, but there are still some limitations in practical applications.

[0004] However, the existing data imbalance fault diagnosis methods still have the following two main problems: Problem 1: When traditional data generation models generate virtual fault signals, they may not accurately capture the characteristics of vibration time-domain signals. This results in a significant difference between the generated virtual signals and actual fault signals, thereby affecting the performance of the classification model. Especially when the number of fault samples used to train the generation model is limited, the fidelity of the generated samples will be significantly affected. If these low-fidelity signals are used to augment the training set, it will cause the classification model to learn incorrect fault information, thereby reducing the accuracy of fault diagnosis.

[0005] Problem 2: Traditional classification models often perform poorly when dealing with imbalanced datasets, tending to bias towards majority-class samples, resulting in a decline in the ability to identify fault categories. Even when training in a mixed training set composed of real data and virtual data, it is difficult to fully extract effective fault features from the mixed dataset. This leads to a low accuracy rate of fault identification, unable to meet the requirements of high-precision fault diagnosis in practical applications. Summary of the Invention

[0006] To solve the problem of low accuracy in fault diagnosis of gears under data imbalance, a data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution is provided. The classification model finally obtained by the present invention has the characteristics of light weight, and at the same time can more effectively develop multi-level features in the augmented dataset, can extract more effective features from real data and generated virtual data, and has excellent fault classification ability.

[0007] The technical means adopted by the present invention are as follows: A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution, comprising: S1. Use an acceleration sensor to collect the vibration time-domain signals of the gear under various fault conditions as gear fault signals, and construct an imbalanced training set; S2. Adopt a diffusion model composed of an improved waveform signal generation network (Idfwave-cm network) to establish a gear fault signal generation model, input the collected gear fault signals into the gear fault signal generation model for training, and generate virtual gear fault signals; S3. Use a signal filter to evaluate the quality of the virtual gear fault signals, and merge the evaluated signals with the original training set to obtain an augmented dataset; S4. Adopt a depthwise separable convolution structure to construct a gear fault classification model, and use the training set in the augmented dataset to train the gear fault classification model, and finally obtain a classification model for identifying the types of gear faults.

[0008] Further, step S2 specifically includes: S21. The diffusion model gradually adds Gaussian noise to the gear fault signal through the diffusion process, completely converting the gear fault signal into Gaussian noise, and reconstructing the data distribution of the gear fault signal from the Gaussian noise through the inverse diffusion process. The diffusion process and the inverse diffusion process are as follows: ; ; Wherein, represents the diffusion process, represents the signal when the diffusion step size is ​ Indicates the signal when the diffusion step size is , Indicates the variance of the added Gaussian noise, Indicates the identity matrix, Indicates the step size in the diffusion process, Indicates the inverse diffusion process, Indicates in the inverse diffusion process and The related mean, Indicates in the inverse diffusion process and The related variance; S22. In the inverse diffusion process, it is quite challenging to accurately calculate the inverse distribution through mathematical operations. Therefore, the Idfwave-cm network structure is adopted to estimate this process. In addition, to improve the model training efficiency and reduce the computational complexity, the variational lower bound in the network training process is simplified, and the expression is as follows: ; Among them, Indicates the simplified result of the variational lower bound, Indicates taking the expectation of the joint distribution of multiple random variables, Indicates Gaussian noise, Indicates the predicted Gaussian noise, Indicates Cumulative multiplication, Indicates any Under , Indicates the initial gear fault signal; S23. The improved waveform signal generation network (Idfwave-cm) is a powerful network architecture designed to handle short-term and long-term dependencies in data, effectively capturing sequence information and global features. This structure is constructed using a simplified self-attention mechanism and an efficient multi-scale attention mechanism, and these two mechanisms work together to modulate the Diffwave network structure. The parameter Indicates the number of residual layers, set to 30. The input of the Idfwave-cm network Is a multi-dimensional encoded vector, formed by converting the diffusion step into multiple trigonometric functions, specifically as follows: ; S24. The training process adopts a linear noise addition mode, where the variance Of the added Gaussian noise linearly increases from 0.0001 to 0.05. After the model training is completed, a virtual gear fault signal is generated through the inverse diffusion process, and this process is expressed as follows: ; Among them, Indicates and Related intermediate variables, denote to cumulative multiplication, denote hyperparameters, denote Gaussian noise; Furthermore, the signal filter in step S3 is a signal filter based on three statistical indicators, which is used to evaluate the similarity between the generated signal and the real signal. The indicators include cosine similarity, Pearson correlation coefficient, and maximum mean discrepancy.

[0009] Furthermore, the calculation formulas for the cosine similarity and the Pearson correlation coefficient are as follows: ; ; where, denotes cosine similarity, denotes Pearson correlation coefficient. Both are used to calculate the similarity between the frequency domain characteristics of the generated signal and the actual signal. The higher the value, the higher the similarity. In the formula, denotes the generated signal frequency domain vector, denotes the real signal frequency domain vector; denotes the number of signal vector elements, denotes the average value of each element of the generated signal, denotes the average value of each element of the real signal.

[0010] Furthermore, the calculation formula for the maximum mean discrepancy is as follows: ; where, denotes the maximum mean discrepancy, which is used to measure the similarity of the data distribution characteristics between the generated signal and the actual signal in the time domain. The smaller the value, the higher the quality of the generated sample. In the formula, denotes the number of elements of the generated signal time domain vector, denotes the number of elements of the real signal time domain vector, denotes the mapping of the generated time domain signal in the Hilbert space, denotes the mapping of the real time domain signal in the Hilbert space, denotes the Hilbert space.

[0011] Furthermore, in step S3, a signal filter is used to evaluate the quality of the virtual gear fault signal, specifically including: Set the constraint conditions as follows: ; where, , and The thresholds corresponding to the three constraint conditions determine the diversity and similarity of the generated virtual gear fault signals; during the generation process, the thresholds of the three constraint conditions , and are set to 0.3, 0.75, and 0.8 respectively; The generated virtual gear fault signals are sent to a signal filter for quality assessment. If the three conditions specified in the constraint conditions are met, the signal is input into the original training set to expand the training set; otherwise, it is regarded as unqualified and discarded. Furthermore, in step S4, a gear fault classification model is constructed using a depthwise separable convolution structure, specifically including: A backbone network is constructed using depthwise separable convolution blocks, and an iterative attention feature fusion module is used to adaptively fuse deep features to improve the fault diagnosis performance of the classification model; The depthwise separable convolution blocks are stacked in a ratio of 1:1:3:1 to obtain deep features of the gear fault signals; To balance the high-level abstract information and low-level texture information of the features, the constructed network structure comprehensively utilizes features of different sizes in three stages. The features of the three stages are adaptively fused at the feature level through the iterative attention feature fusion module. The feature fusion process is as follows: ; ; Among them, and respectively represent the input features, represents the preliminary feature fusion result, represents the final output after feature fusion. The symbol represents broadcast addition, and the symbol represents element-wise multiplication operation, and respectively represent multi-scale channel attention modules at different stages.

[0012] Furthermore, the depthwise separable convolution block linearizes the conventional two-dimensional depthwise separable convolution block and consists of a 1×7 depth convolution, a pointwise convolution, layer normalization (LN), and a GELU activation function. The depthwise separable convolution block uses a large-size convolution to obtain a larger receptive field and uses pointwise convolution for information interaction and dimension transformation between channels, having excellent feature extraction ability and strong robustness.

[0013] Furthermore, the backbone of the feature fusion module is a multi-scale channel attention module, which uses local attention branches and global attention branches to obtain important attention positions, obtains weights through the attention mechanism, and realizes feature adaptive fusion. The multi-scale channel attention module realizes channel attention at multiple scales by changing the pooling size, aiming to effectively combine local and global features.

[0014] Compared with the prior art, the present invention has the following advantages: 1. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution provided by the present invention can expand the training set by generating virtual data when dealing with an imbalanced training set, thereby effectively alleviating the data imbalance problem and improving the accuracy and stability of gear fault diagnosis.

[0015] 2. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution provided by the present invention constructs a diffusion model for generating gear fault signals by using an improved waveform signal generation network (Idfwave-cm). And the generated gear fault signals are screened by a signal filter to ensure that the generated gear fault signals all meet the consistent quality standards. This gear fault signal generation model can learn the complex data distribution information of the input gear vibration signals and generate gear fault signals highly similar to the input signals.

[0016] 3. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution provided by the present invention constructs convolution blocks by using depthwise separable convolution, and constructs the backbone network of the classification model by stacking the convolution blocks to obtain deep features of the fault signals. And an iterative attention feature fusion module is used to adaptively fuse the features of each stage to balance the high-level abstract information and low-level texture information of the features. The integrated classification model has the ability to accurately analyze and distinguish various features from a data set mixed with generated and real signals.

[0017] For the above reasons, the present invention can be widely promoted in the fields of data imbalance fault diagnosis and the like. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 It is a flow chart of the method of the present invention.

[0020] Figure 2Structural diagram of the improved waveform signal generation network (Idfwave-cm) provided by the embodiments of the present invention.

[0021] Figure 3 Structural diagram of the classification model network provided by the embodiments of the present invention.

[0022] Figure 4 Comparison diagram of the generated signal and the real signal provided by the embodiments of the present invention.

[0023] Figure 5 Visualization diagram of the generated signal after dimensionality reduction and the real signal provided by the embodiments of the present invention.

[0024] Figure 6 Fault diagnosis accuracy rate diagram under different imbalance ratios provided by the embodiments of the present invention. Detailed implementation manners

[0025] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0026] It should be noted that the terms "including" and "having" and any variations thereof in the specification and claims of the present invention and the above-mentioned drawings are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those clearly listed steps or units, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0027] As Figure 1 shown, the present invention provides a data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution, including: S1. Using an acceleration sensor to collect the vibration time-domain signals of the gear under various fault conditions as gear fault signals, and constructing an imbalanced training set; S2. Using a diffusion model composed of an improved waveform signal generation network (Idfwave-cm) to establish a gear fault signal generation model, inputting the collected gear fault signals into the gear fault signal generation model for training, and generating virtual gear fault signals; S3. Using a signal filter to evaluate the quality of the virtual gear fault signals, and combining the evaluated signals with the original training set to obtain an expanded data set; S4. Adopt a depthwise separable convolution structure to construct a gear fault classification model, and use the training set in the augmented dataset to train the gear fault classification model, finally obtaining a classification model for identifying gear fault types.

[0028] In this embodiment, the dataset in step S1 is from a faulty gearbox test platform, which consists of a gearbox, a data acquisition system, an acceleration sensor, and a vibration input module. In specific operations, under the conditions that the rotational speed of the gearbox is set to 1450 r / min and the sampling frequency is set to 20 kHz, the time-domain vibration signals are collected through an accelerometer. To simulate gearbox faults, the healthy gears are preprocessed and various types of faults are introduced. By systematically replacing these faulty gears, vibration signals in various healthy states are collected, including time-domain vibration signals under gear wear (W), pitting corrosion (P), gear crack (C), broken teeth (BT), and normal (N) conditions. Based on the above dataset, 8 cases of data imbalance are constructed, as shown in Table 1 for details. These cases cover 2 single-class data imbalance scenarios and 6 multi-class data imbalance scenarios. In the subsequent fault classification experiments, these 8 cases will be used as the training set of the proposed fault diagnosis framework. At the same time, 200 vibration signals are selected from each type to form a test set for evaluating the performance of the diagnosis framework, and the signals in these test sets have the same length as the signals in the training set.

[0029] Table 1 Data imbalance cases

[0030] When specifically implemented, as a preferred implementation manner of the present invention, step S2 specifically includes: S21. The diffusion model gradually adds Gaussian noise to the gear fault signal through the diffusion process, completely converting the gear fault signal into Gaussian noise, and reconstructs the data distribution of the gear fault signal from the Gaussian noise through the inverse diffusion process. The diffusion process and the inverse diffusion process are as follows: ; ; Among them, represents the diffusion process, represents the signal when the diffusion step size is , represents the signal when the diffusion step size is , represents the variance of the added Gaussian noise, represents the identity matrix, represents the step size in the diffusion process, represents the inverse diffusion process, represents in the inverse diffusion process and The relevant mean value, represents the variance related to the inverse diffusion process and the relevant variance; S22. Simplify the variational lower bound during the network training process. The expression is as follows: ; Among them, represents the simplified result of the variational lower bound, represents taking the expectation of the joint distribution of multiple random variables, represents Gaussian noise, represents the predicted Gaussian noise, represents cumulative multiplication, represents any under , represents the initial gear fault signal; S23. The input of the improved waveform signal generation network (Idfwave-cm) is a multi-dimensional coding vector formed by converting the diffusion step into multiple trigonometric functions, specifically as follows: ; S24. The training process adopts a linear noise addition mode, where the variance of the added Gaussian noise linearly increases from 0.0001 to 0.05. After the model training is completed, a virtual gear fault signal is generated through the inverse diffusion process. This process is expressed as follows: ; Among them, represents the intermediate variable related to and represents from to cumulative multiplication, represents the hyperparameter,

[0031] In this embodiment, as Figure 2 shown, it is the network structure diagram of Idfwave-cm. Idfwave-cm is a powerful network architecture designed to handle short-term and long-term dependencies in data, effectively capturing sequence information and global features. This structure is constructed using a simplified self-attention mechanism and an efficient multi-scale attention mechanism, and these two mechanisms work together to modulate the Diffwave network structure. The parameter k represents the number of residual layers, which is set to 30.

[0032] In specific implementation, as a preferred implementation manner of the present invention, the signal filter in step S3 is a signal filter based on three statistical indicators, which is used to evaluate the similarity between the generated signal and the true signal. The indicators include cosine similarity, Pearson correlation coefficient, and maximum mean discrepancy. Among them: The calculation formulas of the cosine similarity and the Pearson correlation coefficient are as follows: ; ; Wherein, represents the cosine similarity, represents the Pearson correlation coefficient, and both are used to calculate the similarity between the frequency domain characteristics of the generated signal and the actual signal. The higher the value, the higher the similarity. In the formula, represents the generated signal frequency domain vector, represents the true signal frequency domain vector; represents the number of signal vector elements, represents the average value of each element of the generated signal, represents the average value of each element of the true signal.

[0033] The calculation formula of the maximum mean discrepancy is as follows: ; Wherein, represents the maximum mean discrepancy, which is used to measure the similarity of the data distribution characteristics of the generated signal and the actual signal in the time domain. The smaller the value, the higher the quality of the generated sample. In the formula, represents the number of elements of the generated signal time domain vector, represents the number of elements of the true signal time domain vector, represents the mapping of the generated time domain signal in the Hilbert space, represents the mapping of the true time domain signal in the Hilbert space, represents the Hilbert space.

[0034] In specific implementation, as a preferred implementation manner of the present invention, in step S3, a signal filter is used to evaluate the quality of the virtual gear fault signal, which specifically includes: Set the constraint conditions as follows: ; Wherein, , and respectively correspond to the thresholds of the three constraint conditions, which determine the diversity and similarity of the generated virtual gear fault signal; corresponding to the thresholds of the three constraint conditions, which determine the diversity and similarity of the generated virtual gear fault signal; during the generation process, the thresholds of the three constraint conditions , and are set to 0.3, 0.75, and 0.8 respectively; In specific implementation, as a preferred implementation manner of the present invention, in step S4, a depthwise separable convolution structure is adopted to construct a gear fault classification model as shown in Figure 3 as follows: A backbone network is constructed by using depthwise separable convolution blocks, and an iterative attention feature fusion module is used to adaptively fuse deep features to improve the fault diagnosis performance of the classification model; The depthwise separable convolution blocks are stacked in a ratio of 1:1:3:1 to obtain deep features of gear fault signals; In order to balance the high-level abstract information and low-level texture information of features, the constructed network structure comprehensively utilizes features of different sizes in three stages, and the features of the three stages are adaptively fused at the feature level through an iterative attention feature fusion module. The feature fusion process is as follows: ; ; wherein, and respectively represent the input features, represents the preliminary fusion result of the features, represents the final output after feature fusion, the symbol represents broadcast addition, and the symbol represents an element-wise multiplication operation, and respectively represent multi-scale channel attention modules at different stages.

[0035] In specific implementation, as a preferred implementation manner of the present invention, the depthwise separable convolution block linearizes the conventional two-dimensional depthwise separable convolution block, which consists of a 1×7 depth convolution, a pointwise convolution, layer normalization, and a GELU activation function. The depthwise separable convolution block uses a large-size convolution to obtain a larger receptive field, and the pointwise convolution is used for information interaction and dimension transformation between channels.

[0036] In specific implementation, as a preferred implementation manner of the present invention, the backbone of the feature fusion module is a multi-scale channel attention module, which uses local attention branches and global attention branches to obtain important attention positions, obtains weights through an attention mechanism, and realizes feature adaptive fusion. The multi-scale channel attention module realizes channel attention at multiple scales by changing the pooling size, aiming to effectively combine local and global features.

[0037] Embodiment Generate gear signal evaluation: Figure 4Presents the comparison graphs of the generated signals and the corresponding true signals in two fault categories. The time-domain graph shows that the waveforms of the virtual signals are highly similar to the true signals, but there are still slight differences, indicating that while the generated virtual samples and true samples maintain a high degree of similarity, they also have a certain degree of diversity. The frequency-domain graphs are presented in a stacked manner, more clearly revealing the similarity of the generated virtual signals and true signals in terms of frequency-domain characteristics. In addition, the Pearson correlation coefficient (PCC) and cosine similarity (CS) in the frequency domain are calculated to evaluate the similarity between the signals. The PCC values are 0.85 and 0.79 respectively, while the CS values are 0.89 and 0.86 respectively. The high levels of the PCC and CS values further verify the strong correlation between the two groups of signals.

[0038] To deeply analyze the spatial distribution characteristics of the generated signals, 500 virtual generated signals are extracted from each of the two fault categories, and 500 true signals are randomly selected as the reference benchmark. Since both the generated virtual signals and true signals have a large length, it is difficult to directly visualize the distribution characteristics of multiple signals. Therefore, principal component analysis is used for data dimensionality reduction to visualize the spatial distribution of the signals. Figure 6 Clearly shows the visualization of the spatial distribution of the generated signals and true signals after dimensionality reduction processing. From Figure 5 it can be seen that the generated virtual signals have been successfully embedded in the spatial distribution domain of the true signals, presenting a highly consistent distribution morphological characteristic with the true signals. The introduction of the virtual generated signals significantly expands the feature space of the training signals, and this expansion of the feature space can provide new learnable features for the fault classification model.

[0039] Data imbalance fault diagnosis analysis: To evaluate the practical value of the generated time-domain vibration signals, training sets with different imbalance ratios are first constructed. The initial training sets are Case 2 and Case 8. Then, the generated virtual signals are gradually injected into the minority-class samples in each case, thus adjusting the imbalance ratio of the training set from 1:100 to 1:50, 1:20, 1:5, and 1:1. These adjusted training sets are used to train the classification model to explore the improvement effect of the virtual signals on the fault classification performance. To prevent potential biases in the results caused by high-performance classification networks, a one-dimensional convolutional neural network is also selected as the fault classification network in this embodiment. Each experiment is repeated ten times, and finally the average fault classification accuracy is used as the evaluation index. The specific experimental results are as Figure 6 shown. Figure 6 Clearly shows that when using the training set with data imbalance, it is difficult for the classification model to achieve accurate fault classification. However, by adding more generated virtual signals to the original training set, the accuracy of fault classification can be significantly improved.

[0040] To verify the stability and effectiveness of the designed fault diagnosis framework, fault diagnosis was carried out on 8 data imbalance cases. In addition, two statistical metrics, namely accuracy and F1-score, were adopted to comprehensively evaluate the designed fault diagnosis framework. Table 2 shows the fault diagnosis results for all cases, and the results are presented as the average values of the evaluation metrics. Among them, the "no augmentation method" refers to directly inputting the data-imbalanced training dataset into a one-dimensional convolutional neural network for fault diagnosis. The results show that the designed diagnosis framework significantly improves the accuracy and F1-score of fault classification in all cases and successfully alleviates the problem of low classification accuracy under data imbalance conditions. In the case of single-category data imbalance, the fault classification accuracy of this framework exceeds 99%, and it can accurately identify the types of fault signals. Although the accuracy of fault classification will decrease when the number of imbalanced categories in the original training set increases, this framework still demonstrates strong fault recognition ability.

[0041] Table 2 Fault Diagnosis Results for Each Case

[0042] Result Analysis: The above data imbalance fault diagnosis results and the evaluation results of the generated gear signals effectively prove the effectiveness of the designed fault diagnosis framework in data imbalance scenarios.

[0043] The constructed gear vibration signal generation model adopts the Idfwave-cm network structure, effectively capturing the short-term and long-term dependencies in the vibration signals, thus realizing the generation of vibration signals. On this basis, through the application of a signal filter, the generated signals that deviate significantly from the real signals are screened out, further improving the reliability of the generated virtual signals. Adding these generated signals to the training set to expand the data-imbalanced training set significantly improves the performance of the fault diagnosis method.

[0044] The constructed gear fault classification model uses a depthwise separable convolution structure to extract features from the augmented training set, combines a multi-stage iterative attention feature fusion module, adaptively fuses the deep information, and enhances the model's feature extraction ability for the mixed dataset. It demonstrates superior fault diagnosis performance in various data imbalance situations.

[0045] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A data imbalance fault diagnosis method based on diffusion models and depthwise separable convolutions, characterized in that, Including: S1. Collect the vibration time-domain signals of the gear under various fault conditions by using an acceleration sensor as the gear fault signals, and construct an imbalanced training set. S2. Use a diffusion model composed of an improved waveform signal generation network to establish a gear fault signal generation model. Input the collected gear fault signals into the gear fault signal generation model for training, and generate virtual gear fault signals. S3. Use a signal filter to evaluate the quality of the virtual gear fault signals, and merge the evaluated signals with the original training set to obtain an expanded data set. S4. Adopt a depthwise separable convolution structure to construct a gear fault classification model. Use the training set in the expanded data set to train the gear fault classification model, and finally obtain a classification model for identifying gear fault types.

2. The data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 1, wherein Step S2 specifically includes: S21. The diffusion model gradually adds Gaussian noise to the gear fault signals through the diffusion process, completely converting the gear fault signals into Gaussian noise, and reconstructing the data distribution of the gear fault signals from the Gaussian noise through the inverse diffusion process. The diffusion process and the inverse diffusion process are as follows: ； ； Among them, represents the diffusion process, represents the signal when the diffusion step size is ; represents the signal when the diffusion step size is ; represents the variance of the added Gaussian noise, represents the identity matrix, represents the step size in the diffusion process, represents the reverse diffusion process, represents the mean related to in the reverse diffusion process, represents the variance related to in the reverse diffusion process; S22. Simplify the variational lower bound in the network training process, and the expression is as follows: ； Among them, represents the simplified result of the variational lower bound, represents taking the expectation of the joint distribution of multiple random variables, represents Gaussian noise, represents the predicted Gaussian noise, represents cumulative multiplication, represents arbitrary under , represents the initial gear fault signal; Input of the improved waveform signal generation network (Idfwave-cm) S23 is a multi-dimensional encoded vector formed by converting the diffusion step into multiple trigonometric functions as follows: ； S24. The training process adopts a linear noise addition mode, where the variance of the added Gaussian noise increases linearly from 0.0001 to 0.

05. After the model training is completed, a virtual gear fault signal is generated through the inverse diffusion process, and this process is expressed as follows: ； Among them, represents the intermediate variable related to represents from cumulative multiplication, represents the hyperparameter, represents Gaussian noise.

3. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 1, characterized in that, The signal filter in step S3 is a signal filter based on three statistical indicators, which is used to evaluate the similarity between the generated signal and the real signal. The indicators include cosine similarity, Pearson correlation coefficient, and maximum mean discrepancy.

4. The data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 3, wherein, The calculation formulas of the cosine similarity and the Pearson correlation coefficient are as follows: ； ； Among them, represents the cosine similarity, represents the Pearson correlation coefficient, and both are used to calculate the similarity between the frequency domain features of the generated signal and the actual signal. The higher the value, the higher the similarity. In the formula, represents the frequency domain vector of the generated signal, represents the frequency domain vector of the true signal; represents the number of elements of the signal vector, represents the average value of each element of the generated signal, represents the average value of each element of the true signal.

5. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 3, characterized in that The calculation formula of the maximum mean discrepancy is as follows: ； Among them, represents the maximum mean difference, which is used to measure the similarity of the data distribution characteristics between the generated signal and the actual signal in the time domain. The smaller the value, the higher the quality of the generated samples. In the formula, represents the number of elements of the generated signal time-domain vector, represents the number of elements of the real signal time-domain vector, represents the mapping of the generated time-domain signal in the Hilbert space, represents the mapping of the real time-domain signal in the Hilbert space, represents the Hilbert space.

6. The data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 1, wherein In step S3, using the signal filter to evaluate the quality of the virtual gear fault signals specifically includes: Set the following constraint conditions: ； Among them, , and correspond to the thresholds of three constraint conditions respectively, determining the diversity and similarity of the generated virtual gear fault signals; Send the generated virtual gear fault signals into the signal filter for quality evaluation. If the three conditions specified in the constraint conditions are met, input the signals into the original training set to expand the training set; otherwise, they are regarded as unqualified and discarded.

7. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 1, characterized in that In step S4, adopting a depthwise separable convolution structure to construct a gear fault classification model specifically includes: Use depthwise separable convolution blocks to construct a backbone network, and use an iterative attention feature fusion module to adaptively fuse deep features to improve the fault diagnosis performance of the classification model. The depthwise separable convolution blocks are stacked in a ratio of 1:1:3:1 to obtain the deep features of the gear fault signals. In order to balance the high-level abstract information and low-level texture information of the features, the constructed network structure comprehensively uses the features of different sizes in three stages. The features of the three stages are adaptively fused at the feature level through the iterative attention feature fusion module. The feature fusion process is as follows: ； ； Among them, and respectively represent the input features, represents the preliminary feature fusion result, represents the final output after feature fusion. The symbol represents broadcast addition, and the symbol represents an element-wise multiplication operation, and respectively represent the multi-scale channel attention modules at different stages.

8. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 7, characterized in that, The depthwise separable convolution block one-dimensionalizes the conventional two-dimensional depthwise separable convolution block, which consists of a 1×7 depth convolution, a pointwise convolution, a layer normalization, and a GELU activation function. The depthwise separable convolution block uses a large-size convolution to obtain a larger receptive field, and uses the pointwise convolution for information interaction and dimension transformation between channels.

9. A data imbalance fault diagnosis method based on a diffusion model and depthwise separable convolution according to claim 7, characterized in that, The backbone of the feature fusion module is the multi-scale channel attention module, which uses local attention branches and global attention branches to obtain important attention positions, obtains weights through the attention mechanism, and realizes feature adaptive fusion. The multi-scale channel attention module realizes channel attention at multiple scales by changing the pooling size, aiming to effectively combine local and global features.

Images