Fault Diagnosis Method for Rotating Machinery Based on Interpretable Implicit Diffusion Model
By constructing an interpretable implicit diffusion model and combining it with knowledge from the field of fault diagnosis, we designed an interpretable convolutional kernel and a vector quantization variational autoencoder network. This solved the problems of difficult data generation and high computational resource consumption in the fault diagnosis of rotating machinery equipment in scenarios with scarce samples, and achieved efficient and interpretable fault diagnosis results.
Patent Information
- Application Number
- CN202311455613.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-03
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2043-11-03
AI Technical Summary
In scenarios where samples are scarce, existing intelligent fault diagnosis technologies face challenges such as difficulty in data generation, high consumption of computing resources, and lack of interpretability, resulting in insufficient diagnostic performance.
By adopting an interpretable implicit diffusion model, and by designing interpretable convolutional kernels and vector quantization variational autoencoders, combined with knowledge from the field of fault diagnosis, an interpretable implicit diffusion model is constructed to generate high-quality fault vibration signals, thereby improving computational efficiency and diagnostic performance.
It enables the generation of high-quality fault vibration signals under limited data conditions, improves the interpretability and computational efficiency of fault diagnosis of rotating machinery, and enhances the performance of the diagnostic model.
Smart Images

Figure CN117909800B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of mechanical fault diagnosis technology, and specifically to a fault diagnosis method for rotating machinery based on an interpretable implicit diffusion model. Background Technology
[0002] With the rapid development of industrial information technology and artificial intelligence technology, deep learning theory has provided new opportunities and challenges for the application of intelligent fault diagnosis technology in the condition monitoring and maintenance of mechanical equipment. Currently, the large amount of monitoring data accumulating in the industrial field has promoted the large-scale and systematic research and development of intelligent fault diagnosis technology. Since mechanical systems inevitably experience performance degradation and even failures after long-term operation, providing reasonable operation and maintenance strategies is essential for extending the service life of mechanical equipment and improving safety. Although intelligent fault diagnosis technology based on deep learning theory has achieved remarkable results, the data-hungry nature of deep neural networks restricts its further development in the industrial field; efficient diagnostic models often rely on the support of massive amounts of data. In many industrial scenarios, obtaining a large amount of usable fault data is impractical. On the one hand, mechanical equipment usually stops quickly for repair when a fault occurs; on the other hand, manual annotation relying on expert knowledge is time-consuming and labor-intensive. Therefore, how to complete fault diagnosis tasks with limited available data is a pressing problem to be solved in intelligent diagnosis.
[0003] For current intelligent fault diagnosis needs in scenarios with scarce samples, a direct and effective solution is to employ data generation technology. The mainstream data-generated intelligent diagnosis solution currently uses Generative Adversarial Networks (GANs). However, due to the adversarial game training mechanism between the generator and discriminator, GANs are difficult to train and prone to pattern collapse. Furthermore, this model has high requirements for the selection and parameter settings of regularizers. Currently, denoising diffusion probability models have become one of the more efficient data-generated solutions due to their excellent performance and training stability. However, diffusion models, due to their progressive destruction and reconstruction learning method, have high training costs, posing new challenges for intelligent fault diagnosis scenarios with limited computing resources. In addition, many data-generated solutions are limited by the black-box nature of their networks, lacking interpretability in their learning process. Furthermore, the lack of interpretable design for fault information makes it difficult to efficiently focus on fault-related information, thus affecting the performance of the generative model. Therefore, it is necessary to develop high-performance, computationally efficient, and interpretable data-generated intelligent fault diagnosis solutions for scenarios with scarce samples. Summary of the Invention
[0004] To overcome the shortcomings of existing technologies, the present invention aims to propose a fault diagnosis method for rotating machinery based on an interpretable latent diffusion model (ILDM). By leveraging the high performance and training stability of the diffusion model, the quality of data generation in the fault diagnosis field is improved. Furthermore, a vector quantization variational autoencoder network is constructed, and the diffusion model is embedded in its hidden layers to improve computational efficiency. Finally, interpretable convolutional kernels are designed by incorporating knowledge from the fault diagnosis domain and embedded into the input layer of the vector quantization variational autoencoder network, thereby enhancing both model interpretability and the network's ability to extract fault information. Through the above design, ILDM achieves improvements in data generation quality, computational efficiency, and interpretability, providing synthetic fault vibration signals for subsequent rotating machinery diagnosis tasks and improving the diagnostic performance of general fault diagnosis models in scenarios with scarce samples.
[0005] To address the above problems, this invention provides a fault diagnosis method for rotating machinery based on an interpretable implicit diffusion model, comprising the following steps:
[0006] S1. Fault vibration signal acquisition and preprocessing:
[0007] Collect fault vibration signals from rotating machinery and construct an original vibration dataset by normalizing the data;
[0008] S2. Construction of an interpretable implicit diffusion model:
[0009] To obtain the optimal resonant band center frequency of the fault vibration signal of rotating machinery, utilize the periodic attenuation of the fault impact signal of the rotating machinery, and combine convolution operations to extract similarity information, an interpretable convolution kernel for the attenuation impact of the fault in rotating machinery is designed. The expression S of the interpretable convolution kernel is as follows: conv for:
[0010] S conv =cos(2πf) c τ)exp(-δτ)+λN
[0011] In the formula: f c τ is the optimal resonant band center frequency of the fault vibration signal of rotating machinery; τ is the length of the impact attenuation sequence of the interpretable convolution kernel; δ is the attenuation coefficient of the fault impact signal under the influence of the transmission path; N is the noise sampled from the random normal distribution; λ is the balance coefficient to simulate the fault impact signal under noise interference.
[0012] An interpretable convolutional kernel for attenuated impact from faults in rotating machinery is embedded into a vector quantization variational autoencoder (VQAE) model, constructing a VQAEAE network with an interpretable kernel embedding. Through temporal convolution operations between the interpretable convolutional kernel and the input fault vibration signal, the VQAEAE model achieves targeted extraction of fault information, specifically the periodically attenuated fault impact signal related to the fault within the fault vibration signal. A denoising diffusion probability model is then embedded into the low-dimensional discrete latent space following the vector quantization operation of the VQAEAE network, constructing an interpretable implicit diffusion model. The optimization loss function L in the low-dimensional discrete latent space of the diffusion model is... VQ-LDM The expression is:
[0013]
[0014] Where: ε θ For a trainable U-Net network; ε θ (z q,t (t, y) represents the noise intensity at step t in the forward diffusion process predicted by the implicit diffusion model; y is the fault category label; ε t Z is the noise vector sampled from the normal distribution in step t; E is the expectation operator; z q The z represents the low-dimensional discrete vector resulting from the fault vibration signal after dimensionality reduction and vector quantization operations using a vector quantization variational autoencoder network. q,t is a low-dimensional discrete vector after t consecutive noise addition steps during the forward diffusion process; T is the maximum number of noise addition steps in the forward diffusion process;
[0015] S3. Generation of vibration signals for rotating machinery faults:
[0016] Based on existing fault vibration signals and fault categories, an extended dataset of fault vibration signals is obtained by augmenting the original vibration dataset through conditional generation using an interpretable implicit diffusion model.
[0017] S4. Identify the fault type of the faulty vibration signal and perform fault diagnosis:
[0018] A diagnostic model based on a convolutional neural network is constructed and trained using an expanded dataset of fault vibration signals to achieve fault diagnosis of rotating machinery.
[0019] In a preferred embodiment, step S1 specifically includes:
[0020] S11. Collect fault vibration signals of rotating machinery equipment through vibration sensors, and use a fixed-length sampling window to slide sampling to obtain a dataset {X,Y}, where X is the set of fault vibration signals obtained by window sliding sampling, and Y is the set of fault category labels corresponding to the fault vibration signals.
[0021] S12. By standardizing each fault vibration signal obtained in step S11, the signal amplitude is converted to the 0-1 range to obtain a normalized original vibration dataset.
[0022]
[0023] In the formula: x i This represents the i-th fault vibration signal; For x i The average value; σ i For x i Standard deviation; x i (j) is x i The j-th sample point.
[0024] Preferably, step S2 includes the following sub-steps:
[0025] S21. Construct a vector quantization variational autoencoder network to learn and reconstruct fault vibration signals in the original vibration dataset; the vector quantization variational autoencoder network includes a convolutional autoencoder, a vector quantization module, a vector quantization constraint loss, and a mean square error reconstruction loss; wherein, the training loss function L of the vector quantization variational autoencoder network... VQ The expression is:
[0026]
[0027] In the formula: x represents the fault vibration signal of the rotating machinery input to the network; E represents the encoder architecture; D represents the decoder architecture; sg[·] represents the gradient stopping operation; z q This is the discrete vector processed by the vector quantization layer; ζ is the balance coefficient.
[0028] S22. Design interpretable convolutional kernels;
[0029] S23. An interpretable convolutional kernel is embedded into the input layer of the vector quantization variational autoencoder network described in step S21, replacing the original randomly initialized convolutional kernel structure, and an interpretable vector quantization variational autoencoder network is constructed.
[0030] S24. The collected fault vibration signal of the rotating machinery is input into the interpretable vector quantization variational autoencoder network constructed in step S23. The interpretable vector quantization variational autoencoder network is pre-trained. The interpretable vector quantization variational autoencoder network is trained using the Adam gradient descent optimization method according to the loss function in step S21. The parameter update of the interpretable convolutional kernel follows the gradient descent optimization process of the vector quantization variational autoencoder. The update expression of the optimal resonance band center frequency and attenuation coefficient of the fault vibration signal of the rotating machinery in the interpretable convolutional kernel is as follows:
[0031]
[0032]
[0033] In the formula: η is the learning rate.
[0034] Preferably, step S3 includes the following sub-steps:
[0035] S31. After training in steps S24 and S25, an interpretable implicit diffusion model is obtained. The decoder architecture of the diffusion model in the hidden layer and the vector quantization variational autoencoder model is retained to generate a synthetic fault vibration signal that is close to the original fault vibration signal.
[0036] S32. Construct a normally distributed random noise vector based on the dimension of the hidden layer discrete vector data, combine it with the category label of the fault vibration signal to be generated, and send it into the implicit diffusion model in step S25 to generate a new discrete vector in the hidden layer.
[0037] S33. The discrete vector generated in step S32 is fed into the decoder architecture of the interpretable vector quantization variational autoencoder network in step S24. The discrete vector in the hidden layer is restored to the dimension of the original fault vibration signal, thereby generating a new synthetic fault vibration signal. This is to augment the original vibration dataset with few fault vibration signal samples to obtain an expanded dataset of fault vibration signals. Finally, the expanded fault vibration signal dataset is used to train the diagnostic model.
[0038] Preferably, step S4 includes the following sub-steps:
[0039] S41. Construct a diagnostic model based on a convolutional neural network; the diagnostic model includes a convolutional feature extraction module, a fully connected fault classification module, and a cross-entropy fault vibration signal classification loss; wherein, the cross-entropy classification loss is expressed as:
[0040]
[0041] In the formula: n is the number of body fault vibration signal samples; y i C represents the fault category of the i-th fault vibration signal predicted by the model. i The actual fault category of the i-th fault vibration signal;
[0042] S42. Input the expanded fault vibration signal dataset obtained in step S33 into the diagnostic model in step S41, train the diagnostic model using cross-entropy classification loss, and evaluate the fault diagnosis classification performance of the diagnostic model in the test set of fault vibration signals.
[0043] Furthermore, the rotating mechanical equipment in step S1 includes a gearbox, a planetary gearbox, and rolling bearings.
[0044] Furthermore, the sampling frequency of the rolling bearing is 10kHz.
[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0046] 1. This invention collects fault vibration signals from rotating machinery, obtains the optimal resonant band center frequency of the fault signal through spectral kurtosis analysis, utilizes the periodic attenuation impact fault information of the rotating machinery, and extracts similar information through convolution operations to design an interpretable convolution kernel for fault attenuation impact. This interpretable convolution kernel is constructed and embedded into the input layer of the fault signal generation model, replacing the traditional randomly initialized convolution kernel, thereby achieving interpretability in the generation of fault vibration signals from rotating machinery. This invention addresses the black-box nature of the generation model by combining the attenuation impact characterization of the fault signals from rotating machinery, designing and constructing an interpretable convolution kernel to achieve interpretability of the deep generation model, while simultaneously improving the generation quality of fault vibration signals.
[0047] 2. This invention addresses the computational resource consumption of traditional generative models by embedding a diffusion model in the hidden space, thereby saving computational resources while ensuring the quality of generated fault vibration signals.
[0048] 3. This invention can be applied to rotating machinery and equipment, and can be extended to rotating components such as wind turbine gearboxes, nuclear power planetary gearboxes, and aircraft engine bearings. Attached Figure Description
[0049] Figure 1 This is a flowchart of the fault diagnosis method for rotating machinery based on an interpretable implicit diffusion model according to the present invention.
[0050] Figure 2 This is a flowchart of training step S2 of the fault diagnosis method for rotating machinery based on an interpretable implicit diffusion model according to the present invention.
[0051] Figure 3 This is a schematic diagram of the interpretable vector quantization variational autoencoder model structure and parameter settings in an embodiment of the present invention;
[0052] Figure 4 This is a schematic diagram of the diffusion model structure and parameter settings of ILDM in an embodiment of the present invention;
[0053] Figure 5 This is a schematic diagram of the ILDM fault vibration signal generation process in an embodiment of the present invention;
[0054] Figures 6A1 to 6D2 This is a comparison diagram of the time-domain fault signal generated by ILDM and the original time-domain fault signal in an embodiment of the present invention;
[0055] Figure 7 This is a schematic diagram of the rolling bearing experimental apparatus in an embodiment of the present invention. Detailed Implementation
[0056] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.
[0057] This invention collects fault vibration signals from rotating machinery, obtains the optimal resonant band center frequency of the fault signal through spectral kurtosis analysis, utilizes the periodic attenuation of the fault impact signal from the rotating machinery, and combines the principle of extracting optimal similar information through convolution operations to design an interpretable convolution kernel similar to the attenuated impact shape of the rotating machinery fault. The expression S of the interpretable convolution kernel is... conv as follows:
[0058] S conv =cos(2πf) c τ)exp(-δτ)+λN
[0059] In the formula: f c τ is the optimal resonant band center frequency of the vibration signal from a fault in rotating machinery; τ is the sequence length of the interpretable convolution kernel; δ is the attenuation coefficient of the fault impact signal under the influence of the transmission path; N is the noise sampled from a random normal distribution; λ is the balance coefficient to simulate the fault impact signal under noise interference.
[0060] In this invention, an interpretable convolution kernel for attenuation impact of rotating machinery faults is embedded in the input layer of a vector quantization variational autoencoder model, replacing the traditional randomly initialized convolution kernel. Through the temporal convolution operation between the interpretable convolution kernel and the input fault vibration signal, the vector quantization autoencoder model can selectively extract fault information, thereby achieving interpretability in the generation of fault vibration signals of rotating machinery.
[0061] Furthermore, this invention embeds a denoising diffusion probability model (hereinafter referred to as the diffusion model) into the low-dimensional discrete latent space after the vector quantization operation of the vector quantization variational autoencoder network, constructing an interpretable implicit diffusion model. In this way, on the one hand, the interpretable convolutional feature extraction process increases the focus on fault information, allowing more computational resources to be used to learn and reconstruct fault information to augment the fault vibration data; on the other hand, compared to learning the probability distribution of the fault vibration signal using the diffusion model in the original fault vibration signal dimensional space, learning the probability distribution of small-sized low-dimensional discrete latent vectors in the low-dimensional space improves computational efficiency and saves computational resources.
[0062] This invention designs interpretable convolutional kernels by incorporating knowledge from the fault diagnosis domain. This allows for the adaptive extraction of fault impact signals from fault vibration signals in an interpretable manner, improving the network's focus on fault information during feature extraction. Specifically, this invention embeds interpretable convolutional kernels for attenuated fault impacts in rotating machinery into a vector quantization variational autoencoder (VQA) model, constructing a VQA network with interpretable kernel embedding. Through temporal convolution operations between the interpretable convolutional kernels and the input fault vibration signal, the VQA model achieves targeted extraction of fault information, specifically the periodically attenuated fault impact signal related to the fault within the fault vibration signal. A denoising diffusion probability model is embedded in the low-dimensional discrete latent space following the vector quantization operation of the VQA network, constructing an interpretable implicit diffusion model. The optimized loss function L in the low-dimensional discrete latent space of the diffusion model... VQ-LDM The expression is:
[0063]
[0064] Where: ε θ For a trainable U-Net network; ε θ (z q,t (t, y) represents the noise intensity at step t in the forward diffusion process predicted by the implicit diffusion model; y is the fault category label; ε t Z is the noise vector sampled from the normal distribution in step t; E is the expectation operator; z q The z represents the low-dimensional discrete vector resulting from the fault vibration signal after dimensionality reduction and vector quantization operations using a vector quantization variational autoencoder network. q,t is a low-dimensional discrete vector after t steps of continuous noise addition during the forward diffusion process; T is the maximum number of steps for adding noise during forward diffusion.
[0065] In a preferred embodiment of the present invention, a fault diagnosis method for rotating machinery based on an interpretable implicit diffusion model is provided, such as... Figures 1-6D2 As shown, the specific steps include:
[0066] S1. Fault vibration signal acquisition and preprocessing: Acquire fault vibration signals of rotating machinery and construct the original dataset by normalizing the data.
[0067] S2. Construction of an interpretable implicit diffusion model: Design an interpretable convolutional kernel and embed it into a vector quantization variational autoencoder model. Construct a vector quantization variational autoencoder network with an interpretable kernel embedding. After the vector quantization operation in the intermediate layer of the autoencoder network, embed a denoising diffusion probability model to construct an interpretable implicit diffusion probability model.
[0068] S3. Generation of fault vibration signals of rotating machinery equipment: Based on the existing fault vibration signals and fault categories, conditional generation is performed through an interpretable implicit diffusion model to augment the original vibration dataset with few fault vibration signal samples to obtain an expanded dataset of fault vibration signals.
[0069] S4. Identify the fault category of the fault vibration signal and perform fault diagnosis. Utilize the expanded dataset of the fault vibration signal to construct and train a diagnostic model based on a convolutional neural network to diagnose typical faults of rotating machinery, such as inner ring wear, outer ring cracks, and cage failures in rolling bearings.
[0070] Specifically, the implementation process of step S1 is as follows:
[0071] S11. The fault vibration signal collected by the vibration acceleration sensor is sampled by a fixed-length sampling window to obtain the original dataset {X,Y}, where X is the set of fault vibration signals obtained by window sliding sampling and Y is the fault category label of the corresponding signal.
[0072] S12. By standardizing each fault vibration signal obtained in step S11 for the formula, the signal amplitude is converted to the 0-1 range to obtain a normalized fault vibration signal dataset, so as to improve the training stability of the model.
[0073]
[0074] In the formula: x i This is the i-th vibration signal; For x i The average value; σ i For x i Standard deviation; x i (j) is x i The j-th sample point.
[0075] like Figure 2 As shown, the specific implementation process of step S2 is as follows:
[0076] S21. Construct a vector quantization variational autoencoder network to learn and reconstruct the fault vibration signals in the original dataset. The vector quantization variational autoencoder network includes a convolutional autoencoder, a vector quantization module, a vector quantization constraint loss, a regularization loss, and a mean squared error reconstruction loss. The training loss expression for the vector quantization variational autoencoder network is as follows:
[0077]
[0078] In the formula: x represents the fault vibration signal of the rotating machinery input to the network; E represents the encoder architecture; D represents the decoder architecture; sg[·] represents the gradient stopping operation; zq z is a discrete vector processed by the vector quantization layer. q It is also a low-dimensional discrete vector of the fault vibration signal after dimensionality reduction and vector quantization operations by a vector quantization variational autoencoder network; ζ is the balance coefficient;
[0079] S22. Design an interpretable convolution kernel. Obtain the optimal resonant band center frequency of the fault vibration signal of rotating machinery through spectral kurtosis analysis. Utilize the periodic attenuation of the fault impact signal of rotating machinery and combine the principle of optimal extraction of similar information through convolution operations to design an interpretable convolution kernel for the attenuated impact of faults in rotating machinery. The interpretable convolution kernel design is based on the fact that fault information in rotating machinery is mainly characterized as attenuated impact. The interpretable convolution kernel expression S... conv for:
[0080] S conv =cos(2πf) c τ)exp(-δτ)+λN
[0081] In the formula: f c τ is the optimal resonant band center frequency of the fault vibration signal of rotating machinery; τ is the length of the impact attenuation sequence of the interpretable convolution kernel; δ is the attenuation coefficient of the fault impact signal under the influence of the transmission path; N is the noise sampled from the random normal distribution; λ is the balance coefficient to simulate the fault impact signal under noise interference.
[0082] S23. An interpretable convolutional kernel is embedded into the input layer of the vector quantization variational autoencoder network in step S21, replacing the original randomly initialized convolutional kernel structure. An interpretable kernel-embedded vector quantization variational autoencoder network is constructed. Through the temporal convolution operation between the interpretable convolutional kernel and the input fault vibration signal, the vector quantization autoencoder model can selectively extract fault information. The fault information is the periodically decaying fault impact signal related to the fault in the fault vibration signal. This improves the ability of the vector quantization variational autoencoder network to pick up fault information during the encoding process.
[0083] S24. Input the collected vibration monitoring data of the rotating machinery into the interpretable vector quantization variational autoencoder network constructed in step S23, pre-train the interpretable vector quantization variational autoencoder network, and train the interpretable vector quantization variational autoencoder network using the Adam gradient descent optimization method according to the loss function in step S21; wherein, the parameter update of the interpretable convolutional kernel follows the gradient descent optimization process of the vector quantization variational autoencoder, and the update expression of the optimal resonance band center frequency and attenuation coefficient of the interpretable convolutional kernel is:
[0084]
[0085]
[0086] In the formula: η is the learning rate;
[0087] S25. Embed the diffusion model in the hidden layer of the interpretable vector quantization variational autoencoder network constructed in step S23; obtain the low-dimensional discrete vector output by the vector quantization module in the interpretable vector quantization variational autoencoder network optimized in step S24, and input it into the diffusion model along with the fault category label. By learning the input low-dimensional discrete vector, a reconstructed synthetic discrete vector is output; the diffusion model includes a forward Markov discrete vector noise-adding random diffusion process, a backward Markov discrete vector progressive sampling recovery inverse diffusion process, a U-Net main network, and a diffusion model optimization loss function; wherein, the non-Gaussian distributed discrete vector becomes normally distributed after the forward diffusion process, and the optimization loss function L of the diffusion model in the low-dimensional discrete latent space... VQ-LDM The expression is:
[0088]
[0089] Where: ε θ For a trainable U-Net network; ε θ (z q,t (t, y) represents the noise intensity at step t in the forward diffusion process predicted by the implicit diffusion model; y is the fault category label; ε t Z is the noise vector sampled from the normal distribution in step t; E is the expectation operator; z q The z represents the low-dimensional discrete vector resulting from the fault vibration signal after dimensionality reduction and vector quantization operations using a vector quantization variational autoencoder network. q,t is a low-dimensional discrete vector after t steps of continuous noise addition during the forward diffusion process; T is the maximum number of steps for adding noise during forward diffusion.
[0090] The specific implementation process of step S3 is as follows:
[0091] S31. After training in steps S24 and S25, an interpretable implicit diffusion model is obtained. The decoder architecture of the diffusion model in the hidden layer and the vector quantization variational autoencoder model is retained to generate a synthetic fault vibration signal that is close to the original monitoring fault vibration signal.
[0092] S32. Construct a normally distributed random noise vector based on the dimension of the hidden layer discrete vector data, combine it with the category label of the fault vibration signal to be generated, and send it into the implicit diffusion model in step S25 to generate a new discrete vector in the low-dimensional hidden space.
[0093] S33. The discrete vector generated in step S32 is fed into the decoder architecture of the interpretable vector quantization variational autoencoder network in step S24 to restore the discrete vector in the low-dimensional hidden space to the dimension of the original fault vibration signal, thereby generating a new synthetic fault vibration signal. This is to augment the vibration dataset under the original few fault vibration signal samples to obtain an expanded dataset of fault vibration signals. Finally, the expanded fault vibration signal dataset is used to train the diagnostic model.
[0094] The specific implementation process of step S4 is as follows:
[0095] S41. Construct a diagnostic model based on a convolutional neural network; the diagnostic model includes a convolutional feature extraction module, a fully connected fault classification module, and a cross-entropy fault vibration signal classification loss; wherein, the cross-entropy classification loss is expressed as:
[0096]
[0097] In the formula: n is the number of body fault vibration signal samples; y i C represents the fault category of the i-th fault vibration signal predicted by the model. i The actual fault category of the i-th fault vibration signal;
[0098] S42. Input the expanded fault vibration signal dataset obtained in step S33 into the diagnostic model in step S41. Train the diagnostic model using cross-entropy classification loss and evaluate the fault diagnosis classification performance of the diagnostic model in the fault vibration signal test set. The test set contains the same fault categories as the fault vibration signals in the training set. The diagnostic performance of the model is evaluated by correctly predicting the probability of the fault category of the fault vibration signal in the test set.
[0099] In one specific embodiment of the present invention, the rotating mechanical device is a rolling bearing, and the sampling frequency for collecting monitoring data of the rolling bearing is 10kHz.
[0100] In the verification process of the embodiments, traditional data generation methods were selected as references, including data generation methods based on Generative Adversarial Networks (GAN), Deep Convolutional Generative Adversarial Networks (DCGAN), Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP), and Denoising Diffusion Probability Models (DDPM). Among them, GAN is a typical representative of deep learning for data generation in the field of intelligent fault diagnosis. According to the standard intelligent fault diagnosis process based on data generation, firstly, a supervised network is trained using labeled training set samples to generate a generative model. Then, corresponding fault samples are generated according to fault category conditions to expand the original dataset. On this basis, a CNN-based intelligent fault diagnosis model is constructed and trained on both the original dataset and the expanded augmented dataset. Finally, the CNN is trained and tested on new data to determine the improvement of the CNN's diagnostic performance by the generated data. The CNN model consists of three convolutional layers and two fully connected layers, and is mainly used to verify the improvement of the model's diagnostic performance after the generative model expands the dataset.
[0101] This invention was applied to a rolling bearing fault diagnosis embodiment for verification. Specifically, in rolling bearing fault diagnosis, this invention used bearing data from a transmission test bench for verification. A schematic diagram of the test bench is shown below. Figure 7 As shown, the test bench includes a drive motor 1, a coupling 2, a speed and torque sensor 3, a support bearing 4, the bearing under test 5, a radial hydraulic loading unit 6, and a vibration sensor 7. The power of the test bench is provided by the drive motor 1. This invention uses vibration data with a sampling frequency of 10kHz for analysis. Typical failure categories or forms of rolling bearings include inner ring cracks, inner ring wear, outer ring cracks, outer ring wear, inner and outer ring cracks, inner and outer ring wear, and bearing failure. The test bench speed is set to 1200rpm, and the radial load is 1.5KN. Forty samples are collected for each category, and each sample contains 2048 data points. After the dataset is expanded, an additional 100 samples are selected for performance testing of the CNN-based diagnostic model.
[0102] Based on the collected fault vibration signal dataset, and after end-to-end training of the ILDM model and generation of fault vibration signals, in Figures 6A1 to 6D2 The results clearly show that the fault vibration signals generated by ILDM are quite similar to the original fault vibration signals, indicating that ILDM can augment and expand the original fault vibration signal dataset based on the original signals. Further, the similarity between the generated data and the original data was evaluated using three metrics: Spearman correlation coefficient (SCC), Fréchet distance (FD), and mean squared error (RMSE). The similarity within the original data was also calculated. The comparison results of different generation methods are shown in Table 1. Here, "↑" indicates a larger value is better, and "↓" indicates a smaller value is better. The table shows that the quality of the data generated by the diffusion model is superior to that of GAN-based generative solutions. Furthermore, through the design of a vector quantization variational autoencoder and an interpretable kernel, it can be found that ILDM not only significantly reduces computational costs compared to DDPM, but also significantly improves the quality of the generated data.
[0103] Furthermore, the generated fault vibration signal data was combined with the original fault vibration signal data to form an augmented dataset, which was then trained and tested using a constructed CNN. Table 2 shows the comparison results of the augmented datasets obtained from different data generation methods on the improvement of intelligent diagnostic performance. The test results are the average and standard deviation of the accuracy across five tests. Table 2 shows that the data quality generated by GAN and DCGAN was insufficient, resulting in lower test accuracy than the un-augmented original fault vibration signal dataset. The diffusion model showed a significant improvement in test results after generating fault vibration signals and augmenting the dataset compared to before the dataset augmentation. Based on this, the diagnostic performance of the ILDM model, which incorporates a vector quantization variational autoencoder and an interpretable kernel design, was further improved, confirming the effectiveness and superior performance of the ILDM method. Therefore, ILDM can provide a high-quality, computationally efficient, and interpretable solution for generating fault vibration signals for rotating machinery equipment even with limited available data.
[0104] Table 1. Evaluation results (%) of fault vibration signal generation results for various methods in bearing diagnostic embodiments.
[0105]
[0106] Table 2. Results of Intelligent Diagnostic Tests (%)
[0107]
[0108] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A fault diagnosis method for rotating machinery based on an interpretable implicit diffusion model, characterized in that, It includes the following steps: S1. Fault vibration signal acquisition and preprocessing: Collect fault vibration signals from rotating machinery and construct an original vibration dataset by normalizing the data; S2. Construction of an interpretable implicit diffusion model: To obtain the optimal resonant band center frequency of the fault vibration signal of rotating machinery, utilize the periodic attenuation of the fault impact signal of the rotating machinery, and combine convolution operations to extract similarity information, an interpretable convolution kernel for the attenuation impact of the fault in rotating machinery is designed. The expression S of the interpretable convolution kernel is as follows: conv for: S conv =cos(2πf c τ)exp(-δτ)+λN In the formula: f c τ is the optimal resonant band center frequency of the fault vibration signal of rotating machinery; τ is the length of the impact attenuation sequence of the interpretable convolution kernel; δ is the attenuation coefficient of the fault impact signal under the influence of the transmission path; N is the noise sampled from the random normal distribution; λ is the balance coefficient to simulate the fault impact signal under noise interference. An interpretable convolutional kernel for attenuated impact from faults in rotating machinery is embedded into a vector quantization variational autoencoder (VQAE) model, constructing a VQAEAE network with an interpretable kernel embedding. Through temporal convolution operations between the interpretable convolutional kernel and the input fault vibration signal, the VQAEAE model achieves targeted extraction of fault information, specifically the periodically attenuated fault impact signal related to the fault within the fault vibration signal. A denoising diffusion probability model is then embedded into the low-dimensional discrete latent space following the vector quantization operation of the VQAEAE network, constructing an interpretable implicit diffusion model. The optimization loss function of the diffusion model in the low-dimensional discrete latent space is... The expression is: Where: ε θ For a trainable U-Net network; ε θ (z q,t (t, y) represents the noise intensity at step t in the forward diffusion process predicted by the implicit diffusion model; y represents the fault category label; ε t Let be the noise vector sampled from the normal distribution in step t; z is the expectation operator; q The z represents the low-dimensional discrete vector resulting from the fault vibration signal after dimensionality reduction and vector quantization operations using a vector quantization variational autoencoder network. q,t is a low-dimensional discrete vector after t consecutive noise addition steps during the forward diffusion process; T is the maximum number of noise addition steps in the forward diffusion process; S3. Generation of vibration signals for rotating machinery faults: Based on existing fault vibration signals and fault categories, an extended dataset of fault vibration signals is obtained by augmenting the original vibration dataset through conditional generation using an interpretable implicit diffusion model. S4. Identify the fault type of the faulty vibration signal and perform fault diagnosis: A diagnostic model based on a convolutional neural network is constructed and trained using an expanded dataset of fault vibration signals to achieve fault diagnosis of rotating machinery.
2. The method for fault diagnosis of rotating machinery based on an interpretable implicit diffusion model according to claim 1, characterized in that, Step S1 specifically includes: S11. Collect fault vibration signals of rotating machinery equipment through vibration sensors, and use a fixed-length sampling window to slide sampling to obtain a dataset {X,Y}, where X is the set of fault vibration signals obtained by window sliding sampling, and Y is the set of fault category labels corresponding to the fault vibration signals. S12. By standardizing each fault vibration signal obtained in step S11, the signal amplitude is converted to the 0-1 range to obtain a normalized original vibration dataset. In the formula: x i This represents the i-th fault vibration signal; For x i The average value; σ i For x i Standard deviation; x i (j) is x i The j-th sample point.
3. The method for fault diagnosis of rotating machinery based on an interpretable implicit diffusion model according to claim 1, characterized in that, Step S2 includes the following sub-steps: S21. Construct a vector quantization variational autoencoder network to learn and reconstruct fault vibration signals in the original vibration dataset; the vector quantization variational autoencoder network includes a convolutional autoencoder, a vector quantization module, a vector quantization constraint loss, and a mean square error reconstruction loss; wherein, the training loss function of the vector quantization variational autoencoder network is... The expression is: In the formula: x represents the fault vibration signal of the rotating machinery input to the network; E represents the encoder architecture; D represents the decoder architecture; sg[·] represents the gradient stopping operation; z q This is the discrete vector processed by the vector quantization layer; ζ is the balance coefficient. S22. Design interpretable convolutional kernels; S23. An interpretable convolutional kernel is embedded into the input layer of the vector quantization variational autoencoder network described in step S21, replacing the original randomly initialized convolutional kernel structure, and an interpretable vector quantization variational autoencoder network is constructed. S24. The collected fault vibration signal of the rotating machinery is input into the interpretable vector quantization variational autoencoder network constructed in step S23. The interpretable vector quantization variational autoencoder network is pre-trained. The interpretable vector quantization variational autoencoder network is trained using the Adam gradient descent optimization method according to the loss function in step S21. The parameter update of the interpretable convolutional kernel follows the gradient descent optimization process of the vector quantization variational autoencoder. The update expression of the optimal resonance band center frequency and attenuation coefficient of the fault vibration signal of the rotating machinery in the interpretable convolutional kernel is as follows: In the formula: η is the learning rate.
4. The method for fault diagnosis of rotating machinery based on an interpretable implicit diffusion model according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S31. After training in steps S24 and S25, an interpretable implicit diffusion model is obtained. The decoder architecture of the diffusion model in the hidden layer and the vector quantization variational autoencoder model is retained to generate a synthetic fault vibration signal that is close to the original fault vibration signal. S32. Construct a normally distributed random noise vector based on the dimension of the hidden layer discrete vector data, combine it with the category label of the fault vibration signal to be generated, and send it into the implicit diffusion model in step S25 to generate a new discrete vector in the hidden layer. S33. The discrete vector generated in step S32 is fed into the decoder architecture of the interpretable vector quantization variational autoencoder network in step S24. The discrete vector in the hidden layer is restored to the dimension of the original fault vibration signal, thereby generating a new synthetic fault vibration signal. This is to augment the original vibration dataset with few fault vibration signal samples to obtain an expanded dataset of fault vibration signals. Finally, the expanded fault vibration signal dataset is used to train the diagnostic model.
5. The method for fault diagnosis of rotating machinery based on an interpretable implicit diffusion model according to claim 1, characterized in that, Step S4 specifically includes the following sub-steps: S41. Construct a diagnostic model based on a convolutional neural network; the diagnostic model includes a convolutional feature extraction module, a fully connected fault classification module, and a cross-entropy fault vibration signal classification loss; wherein, the cross-entropy classification loss is expressed as: In the formula: n is the number of body fault vibration signal samples; y i C represents the fault category of the i-th fault vibration signal predicted by the model. i The actual fault category of the i-th fault vibration signal; S42. Input the expanded fault vibration signal dataset obtained in step S33 into the diagnostic model in step S41, train the diagnostic model using cross-entropy classification loss, and evaluate the fault diagnosis classification performance of the diagnostic model in the test set of fault vibration signals.
6. The method for fault diagnosis of rotating machinery based on an interpretable implicit diffusion model according to claim 1, characterized in that, The rotating mechanical equipment in step S1 includes a gearbox, a planetary gearbox, and rolling bearings.
7. The method for fault diagnosis of rotating machinery based on an interpretable implicit diffusion model according to claim 6, characterized in that, The sampling frequency of the rolling bearing is 10kHz.
