Hydraulic system intelligent fault diagnosis method based on diffusion model data enhancement framework

Abstract

The invention discloses a hydraulic system intelligent fault diagnosis method based on a diffusion model data enhancement framework, and the method comprises the following steps: firstly providing a diffusion model-based data enhancement framework, carrying out the fault simulation of a hydraulic system through a simulation model, converting a time domain signal into a gray image, and achieving the data enhancement through a diffusion model; carrying out quantitative evaluation on the quality of the generated image, and constructing an enhanced hydraulic system fault data set; based on the proposed framework, a dual-domain adversarial neural network model fused with an attention mechanism is proposed, the weight of a loss function of two domain discriminators is adjusted so as to realize global and local feature adaptive fusion optimization, adversarial training is realized through a gradient inversion layer, and the adaptive fusion optimization of global and local features is realized. Experimental fault state data serve as a test set to be input into a model for fault diagnosis, and the superiority of the method on indexes such as accuracy is verified through comparison of traditional models such as DANN and CNN. In addition, the method is simple and easy to implement, and is suitable for intelligent fault diagnosis of the hydraulic system under the zero sample condition.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent fault diagnosis in hydraulic systems, and specifically relates to an intelligent fault diagnosis method for hydraulic systems based on a diffusion model data augmentation framework. Background Art

[0002] As a key actuating element in modern industrial automation and construction machinery, hydraulic systems have been widely used in fields such as aviation, automotive, metallurgy, and petrochemical industries. Hydraulic systems have advantages such as smooth transmission, rapid response, and high power density. However, during their long-term operation, they are susceptible to wear, seal failure, and environmental factors, resulting in common faults such as hydraulic cylinder leakage and pipeline blockage. These faults not only reduce the working efficiency of the system but may also cause equipment shutdown or even safety accidents, thus posing higher requirements for the fault diagnosis of hydraulic systems.

[0003] Currently, hydraulic system fault diagnosis technologies are mainly divided into model-based methods and data-driven methods. Model-based diagnosis methods achieve fault detection by constructing mathematical or physical models of hydraulic systems and using the differences in system response parameters before and after faults, with good interpretability and theoretical basis. Data-driven methods, on the other hand, rely on a large amount of historical operation data and use machine learning or deep learning algorithms to automatically extract fault features and classify them, which can improve the diagnosis accuracy to a certain extent. However, data-driven methods have high requirements for the quantity and quality of fault data. In practical applications, due to the rarity of hydraulic system fault events and the limitations of data collection conditions, real fault data is often extremely scarce, becoming the key bottleneck restricting the widespread application of data-driven methods. Therefore, there is an urgent need for a hybrid-driven fault diagnosis method that can make full use of the combination of model simulation and data collection to make up for the limitations brought by insufficient fault samples.

[0004] In addressing the problem of scarce fault samples, some scholars have attempted to generate simulated fault data using simulation models and combine transfer learning techniques to achieve knowledge transfer between the source domain and the target domain. Such methods have expanded the training data to a certain extent and improved the generalization ability of the fault diagnosis model under different conditions. However, existing research still inevitably relies on a small amount of real fault data during the training process, failing to fundamentally solve the key bottleneck of scarce hydraulic system fault data, thus restricting the further improvement of fault diagnosis performance.

[0005] In view of the above problems, the present invention proposes an intelligent fault diagnosis method for hydraulic systems based on a diffusion model data augmentation framework. This method generates high-quality simulated fault data through simulation and constructing a diffusion model, quantitatively evaluates the data augmentation effect using image generation evaluation metrics, and constructs an enhanced hydraulic system fault dataset. At the same time, a dual-domain adversarial neural network model with attention mechanism (FADDAN) is adopted, which is trained only using the enhanced hydraulic system fault dataset and experimental normal state data, realizes the extraction of domain-invariant features from multi-sensor data, and promotes the adversarial training between the feature extractor and the domain discriminator through the gradient reversal layer, further optimizing the adaptability of the model between different data domains. Compared with the prior art, the present invention does not require the provision of fault samples during model training, alleviates the problem of insufficient fault samples, improves the accuracy and robustness of fault diagnosis, and has significant theoretical and practical application values. Summary of the Invention

[0006] The present invention designs an intelligent fault diagnosis method for hydraulic systems based on a diffusion model data augmentation framework to achieve intelligent fault diagnosis of hydraulic systems under zero-sample conditions. The aim is to enhance the hydraulic system fault simulation dataset using the diffusion model and complete the migration from simulation data to experimental data through the FADDAN model, realizing efficient and robust intelligent fault diagnosis of hydraulic systems.

[0007] The technical solution for realizing the present invention is as follows: An intelligent fault diagnosis method for hydraulic systems based on a diffusion model data augmentation framework, comprising the following steps:

[0008] Step 1: According to the fault types of the hydraulic system, conduct fault simulation, process the multi-sensor related signals in the normal state and fault state obtained by simulation, and obtain a single-channel simulation grayscale image with 128*128 pixels.

[0009] Step 2: Construct a diffusion model, and use the diffusion model to augment the simulation grayscale image data to obtain an enhanced hydraulic system fault dataset.

[0010] Step 3: Construct FADDAN, where FADDAN includes a feature extractor, a label predictor, a gradient reversal layer, and two domain discriminators.

[0011] Step 4: According to the fault types of the hydraulic system, conduct a fault simulation experiment, process the multi-sensor related signals in the normal state and fault state collected in the experiment, obtain a single-channel experimental grayscale image with 128*128 pixels, and use the experimental grayscale image corresponding to the normal state and all the images in the enhanced hydraulic system fault dataset as the training set.

[0012] Step 5: Train FADDAN using the training set to obtain a trained fault diagnosis model.

[0013] Step 6: Use the experimental grayscale images corresponding to the fault states as the test set, and input the test set into the trained fault diagnosis model for intelligent fault diagnosis of the hydraulic system.

[0014] Compared with the prior art, the remarkable advantages of the present invention are as follows:

[0015] (1) The present invention proposes a data augmentation framework based on the diffusion model. After generating fault data using the simulation model, the diffusion model is used to perform high-quality augmentation on the data. The diffusion model gradually adds Gaussian noise in the forward process and gradually removes the noise in the reverse process, thereby generating high-quality images approaching the true image distribution. Even when the original fault data is limited, it can effectively expand the diversity and representativeness of the data set.

[0016] (2) Based on the proposed framework, the FADDAN model is proposed. Among them, the global domain discriminator distinguishes the data corresponding to the normal test state from all augmented data (focusing on global features), and the local domain discriminator distinguishes the data corresponding to the normal test state from the augmented data corresponding to the normal simulation state (emphasizing local features). The weights of the loss functions of the two domain discriminators are flexibly adjusted to make the model reach a balanced state in the learning of global and local features. During the training phase, it is not necessary to introduce test fault state data, thus solving the problem of zero-shot fault diagnosis of the hydraulic system. Relying on the simulation data and the augmented data generated by the diffusion model, even under the condition of extremely scarce actual fault samples, a reliable training data set can be constructed to realize the intelligent diagnosis of hydraulic system faults. Description of the Drawings

[0017] Figure 1 is the flowchart of the intelligent fault diagnosis method for the hydraulic system based on the data augmentation framework of the diffusion model of the present invention.

[0018] Figure 2 is the schematic diagram of the proposed FADDAN model.

[0019] Figure 3 are the evaluation indicators of the images generated by the diffusion model.

[0020] Figure 4 are the fault diagnosis results of the FADDAN model.

[0021] Figure 5 is the fault diagnosis confusion matrix of the FADDAN model. Detailed Embodiments

[0022] To make the objectives, technical solutions, and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0023] The intelligent fault diagnosis method for hydraulic systems based on the diffusion model data augmentation framework proposed in the present invention first proposes a data augmentation framework based on the diffusion model, uses a simulation model to simulate faults in the hydraulic system, converts time-domain signals into grayscale images, realizes data augmentation through the diffusion model, and quantitatively evaluates the quality of the generated images to construct an augmented hydraulic system fault dataset. Based on the proposed framework, the FADDAN model is proposed, adversarial training is realized through a gradient reversal layer, the experimental fault state data is used as the test set to input into the model for hydraulic system fault diagnosis, and through comparison with traditional models such as DANN and CNN, the effectiveness of the present method in terms of indicators such as accuracy, recall rate, F1 score, and accuracy rate is verified. The schematic diagram of the FADDAN model is as Figure 2 , including a feature extractor, a label predictor, a gradient reversal layer, and two domain discriminators. In addition, the present invention is simple and easy to implement and is applicable to intelligent fault diagnosis of hydraulic systems under zero-sample conditions.

[0024] Combined with Figure 1 , the intelligent fault diagnosis method for hydraulic systems based on the diffusion model data augmentation framework proposed in the present invention is as follows:

[0025] Step 1: According to the fault types of the hydraulic system, conduct fault simulation and simulation, process the multi-sensor related signals of the normal state and the fault state obtained from the simulation, and obtain a single-channel simulation grayscale image of 128*128 pixels, specifically as follows:

[0026] The fault types of the hydraulic system include at least one of the cylinder leakage fault and the pipeline blockage fault. The related signals of the hydraulic system include the flow rate of the high-pressure chamber of the cylinder, the leakage flow rate of the cylinder, the pressure of the high-pressure chamber of the cylinder, the pressure of the low-pressure chamber of the cylinder, and the magnitude of the control signal. The data processing method is to stack the 5-way sensor signals in the time domain, eliminate redundant data, normalize and organize them within the range of 0 to 1, and then map them into a single-channel grayscale image of 128*128 pixels.

[0027] Step 2: Construct a diffusion model, use the diffusion model to perform data augmentation on the simulation grayscale image, and obtain an augmented hydraulic system fault dataset, specifically as follows:

[0028] The diffusion model includes a forward process and a reverse process. In the forward process, Gaussian noise is gradually added to the input data. After several steps of gradually adding noise, the original image will eventually turn into a pure noise image that conforms to the standard normal distribution. The mathematical form of the diffusion process can be described as:

[0029]

[0030] Among them, x t represents the image generated after t-step diffusion, x0 represents the input image, and the hyperparameter constant α i = 1 - β i , β i is a constant term, a constant less than 1, and ∈ is Gaussian distribution noise.

[0031] The reverse process of the diffusion model is a Markov chain that gradually reverses the forward diffusion process through parametric learning. Its core goal is to achieve a generative mapping from standard normal distribution noise to the original data distribution. The reverse process first samples the initial noise from , sets the time from T to 1, calculates the mean and variance of the denoising operation at this time, and samples x t-1 , and finally outputs it as the generated image x0. The mean can be expressed as:

[0032]

[0033] where ∈ θ (x t , t) is the predicted noise, and usually a U-Net is used for prediction.

[0034] The variance σ t can be expressed as:

[0035]

[0036] β t is a constant term.

[0037] The mathematical expression of the reverse process of the diffusion model is:

[0038]

[0039] Among them, x t-1 represents the denoised image obtained at the (t - 1)-th step, α t is a constant, ∈ θ (x t , t) represents the noise estimated by the U-Net, σ t represents the variance of the denoising operation at the t-th step, and z represents Gaussian noise.

[0040] The data augmentation process includes diffusion model training and data augmentation.

[0041] When training the diffusion model, all simulated grayscale images are sampled twice to obtain a uniformly distributed diffusion step size t and the corresponding Gaussian distribution noise ∈. Next, U-Net is used to estimate the noise added in the forward diffusion process, and the loss function is calculated based on the difference between the noise estimated by U-Net and the actual noise. Finally, the network parameters of the diffusion model are updated through back propagation according to the loss function value to obtain a trained diffusion model.

[0042] When augmenting data, the noise image x is sampled from a probability distribution that follows a standard normal distribution T , use the inverse process of the trained diffusion model to reduce the noise of the image, and perform at most T steps until the enhanced simulated grayscale image x0 is generated. Repeat the data enhancement process multiple times to obtain multiple different enhanced simulated grayscale images and obtain the enhanced hydraulic system fault data set.

[0043] The fault simulation data under different fault modes are enhanced through the diffusion model. The quality of the images generated by the diffusion model is evaluated by image generation evaluation indicators to construct an enhanced hydraulic system fault dataset.

[0044] The MMD, FID, KID, and IS indicators are used to calculate the five data enhancement methods DDPM, VAE, GAN, SNGAN, and WGAN-GP respectively. Due to the different evaluation methods of the indicators, the calculation results are normalized. The closer the value is to 0, the worse the effect is, and the closer the value is to 1, the better the effect is, so as to achieve an accurate evaluation of the data enhancement effect. The diffusion model generates image evaluation indicators such as Figure 2 .

[0045] Step 3: Construct FADDAN, which includes a feature extractor, a label predictor, a gradient reversal layer, and two domain discriminators, as follows:

[0046] Traditional domain adversarial neural networks require all data of the target domain state during training. However, it is difficult to obtain hydraulic system fault data. Most of the time, only normal operating state data can be obtained. This situation is called the zero-sample problem. When using traditional domain adversarial neural networks to solve the zero-sample problem, it is impossible to fully extract the abnormal features hidden in the fault data, which affects the model's discrimination effect and generalization performance.

[0047] To this end, the present invention proposes the FADDAN model for the first time. The model innovatively proposes a local domain discriminator, deeply mines local domain-irrelevant features, and flexibly adjusts the role of the two domain discriminators in the model training process by designing the weights of the loss functions of the two domain discriminators. This allows the model to achieve a balanced state in the learning of global features and local features, thereby effectively improving the model's fault diagnosis capability in zero-sample conditions.

[0048] The feature extractor includes a 3-layer convolutional neural network, a CBAM attention mechanism, and a 1-layer fully connected layer. The label predictor includes 2 layers of fully connected layers; the two domain discriminators are a global domain discriminator and a local domain discriminator respectively, and the above two domain discriminators are both composed of 2 layers of fully connected layers; during the forward propagation process, the gradient reversal layer acts as an identity mapping, and during the backward propagation process, the gradient reversal layer multiplies the gradient by the negative identity matrix -I to achieve the reversal of the gradient.

[0049] The mathematical description of the forward process of the gradient reversal layer is:

[0050] GRL(x) = x

[0051] where x is the input image of the gradient reversal layer, and GRL represents the gradient reversal layer.

[0052] The mathematical description of the forward process of the gradient reversal layer is:

[0053]

[0054] The gradient reversal layer is located between the feature extractor and the two domain discriminators, and the two domain discriminators share one gradient reversal layer.

[0055] Step 4: According to the fault types of the hydraulic system, conduct fault simulation experiments, process the multi-sensor related signals in the normal state and fault state collected in the experiments to obtain single-channel experimental grayscale images of 128*128 pixels, and use the experimental grayscale images corresponding to the normal state and all the images in the enhanced hydraulic system fault dataset as the training set.

[0056] Step 5: Use the training set to train FADDAN to obtain a trained fault diagnosis model, specifically as follows:

[0057] For the training set x i , the gray image serial number i = 1,..., a,..., n,..., N. Assume that the first n are all simulated gray images with classification labels, and the last n' are experimental gray images corresponding to the normal state without classification labels. The total number of gray images in the training set N = n + n'. Among the first n simulated gray images with classification labels, the first a are simulated gray images corresponding to the normal state with classification labels.

[0058] The training set x i enters FADDAN. First, the first two layers of the convolutional neural network in the feature extractor are used to extract low-level features to generate a preliminary feature map F, then the CBAM attention mechanism in the feature extractor is used to adaptively weight the feature map, and then the third layer of the convolutional neural network in the feature extractor is used to extract high-level features, and the core discriminant information is retained through the fully connected layer to obtain the final feature F out , and its mathematical description is:

[0059]

[0060] Among them, G f (·; θ f ) is a feature extractor, and θ f is the network parameter of the feature extractor.

[0061] The final feature F out outputs the class prediction probability through the label predictor. The final feature F out outputs the global domain prediction probability through the gradient reversal layer and the global domain discriminator. The final feature F out The feature F o ′ ut extracted from the simulation grayscale image corresponding to the normal state and the experimental grayscale image corresponding to the normal state in the training set through the feature extractor outputs the local domain prediction probability through the gradient reversal layer and the local domain discriminator.

[0062] During training, the loss function L cls of the label classifier is:

[0063]

[0064] Among them, is the cross-entropy loss function CrossEntropyLoss, G y (·; θ y ) is the label classifier, θ y is the network parameter of the label predictor, y is the classification label, and y i is the classification label corresponding to x i .

[0065] During training, the loss function L dom1 of the global domain discriminator is:

[0066]

[0067] Among them, is the binary cross-entropy loss function BCEWithLogitsLoss, G d1 (·; θ d1 ) is the global domain discriminator, θ d1 is the network parameter of the global domain discriminator, is the gradient reversal layer, d1 is the global domain discriminator domain label, and d 1i is the global domain discriminator domain label corresponding to x i .

[0068] During training, the loss function L dom2 of the local domain discriminator is:

[0069]

[0070] Among them, G d2 (·; θ d2 ) is a local domain discriminator, and θ d2 are the network parameters of the local domain discriminator. d2 is the domain label of the local domain discriminator, and d 2i is the domain label of the local domain discriminator corresponding to x i .

[0071] The total training loss function L total is as follows:

[0072] L total = L cls + λ1L dom1 + λ2L dom2

[0073] Among them, λ1 is the loss weight of the global domain discriminator, and λ2 is the loss weight of the local domain discriminator. After performing the above operations on all the images in the training set, a trained fault diagnosis model is obtained.

[0074] Step 6: Use the experimental grayscale image corresponding to the fault state as the test set, and input the test set into the trained fault diagnosis model for intelligent fault diagnosis of the hydraulic system.

[0075] Compare the fault diagnosis method of the present invention with DANN, CNN-CBAM, and CNN through comparative experiments, and verify the effectiveness of the method by calculating the precision, recall rate, F1 score, and accuracy of the model.

[0076] Example 1: Data Verification of the Hydraulic System Fault Simulation Test Bench

[0077] The data used in this experiment comes from the hydraulic system fault simulation model and the corresponding test bench. The collected signals mainly include 5 types: the flow rate of the high-pressure chamber of the hydraulic cylinder, the leakage flow rate of the hydraulic cylinder, the pressure of the high-pressure chamber of the hydraulic cylinder, the pressure of the low-pressure chamber of the hydraulic cylinder, and the magnitude of the control signal. A total of 5 health states are set in this experiment: normal state, slight leakage fault of the hydraulic cylinder, severe leakage fault of the hydraulic cylinder, slight blockage fault of the pipeline, and severe blockage fault of the pipeline. The data acquisition time for each health state is about 200s.

[0078] Stack and reconstruct the collected continuous data, eliminate redundant data, normalize and organize it within the range of 0 to 1, and then map it into a single-channel grayscale image of 128 * 128 pixels. Use the diffusion model to perform data augmentation on the hydraulic system fault simulation and simulation data, generating 100 images for each healthy state. Use the augmented data and the normal state data of the test bench to form the training dataset of the network, and use the corresponding fault data of the test bench to form the test set of the network. It should be noted that the test bench fault data only exists in the test dataset. The specific description of the dataset is shown in Table 1.

[0079] Table 1 Dataset description under various states

[0080]

[0081] Figure 3 Shows the evaluation indicators of the images generated by the diffusion model. Due to different evaluation methods of the indicators, the calculation results are normalized. The closer the value is to 0, the worse the effect, and the closer the value is to 1, the better the effect. It can be seen that the images generated by the diffusion model are superior to other models under each indicator evaluation system. From Figure 4 it can be seen that the precision of FADDAN is 96.17%, the recall rate is 95.71%, the F1 score is 95.70%, and the accuracy rate is 95.71%. All indicators are higher than other models. Figure 5 Shows the confusion matrices of several models. It can be seen from it that the correct classification rate of FADDAN for samples is higher than other models. The proposed FADDAN only uses the normal state data of the hydraulic system test bench for training, and can achieve intelligent fault diagnosis of the hydraulic system under the condition of extremely scarce fault samples, and the fault diagnosis accuracy reaches 95.71%.

[0082] In summary, aiming at the problem of extremely scarce hydraulic system fault samples, the present invention proposes an intelligent fault diagnosis method for hydraulic systems based on a diffusion model data augmentation framework. This method constructs a diffusion model to generate high-quality simulation fault data and constructs an enhanced hydraulic system fault dataset; at the same time, it adopts the FADDAN model and only uses the enhanced hydraulic system fault dataset and the experimental normal state data for training to achieve intelligent fault diagnosis of the hydraulic system under zero-sample conditions.

[0083] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered by the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for hydraulic systems based on a data augmentation framework of diffusion models, characterized in that Here are the steps: Step 1: According to the fault type of the hydraulic system, a fault simulation is performed, and data processing is performed on the multi-sensor related signals of the normal state and the fault state obtained by the simulation to obtain a single-channel simulated grayscale image of 128*128 pixels; Step 2: Construct a diffusion model and use the diffusion model to perform data enhancement on the simulated grayscale image to obtain an enhanced hydraulic system fault data set; Step 3: Construct FADDAN, which includes a feature extractor, a label predictor, a gradient reversal layer and two domain discriminators; Step 4: According to the fault type of the hydraulic system, a fault simulation experiment is carried out, and the multi-sensor related signals of the normal state and the fault state collected in the experiment are processed to obtain a single-channel experimental grayscale image of 128*128 pixels. The experimental grayscale image corresponding to the normal state and all images in the enhanced hydraulic system fault data set are used as training sets; Step 5: Use the training set to train FADDAN to obtain a trained fault diagnosis model; Step 6: Use the experimental grayscale image corresponding to the fault state as the test set, and input the test set into the trained fault diagnosis model to perform intelligent fault diagnosis of the hydraulic system.

2. The intelligent fault diagnosis method for hydraulic systems based on the diffusion model data augmentation framework according to claim 1, wherein: In step 1, the fault type of the hydraulic system includes at least one of a hydraulic cylinder leakage fault and a pipeline blockage fault; the multi-sensor related signals include a hydraulic cylinder high-pressure chamber flow, a hydraulic cylinder leakage flow, a hydraulic cylinder high-pressure chamber pressure, a hydraulic cylinder low-pressure chamber pressure, and a control signal size; Data processing is performed on the multi-sensor related signals in the normal state and fault state obtained by fault simulation, that is, the 5-way sensor signals are stacked in the time domain, redundant data are eliminated, and they are normalized to the range of 0 to 1, and then mapped to a single-channel simulated grayscale image of 128*128 pixels.

3. The intelligent fault diagnosis method for hydraulic systems based on the diffusion model data augmentation framework according to claim 1, wherein: In step 2, a diffusion model is constructed as follows: The diffusion model includes the forward process and the reverse process. The mathematical form of the forward process is described as: Among them, x t represents the image generated after t-step diffusion, x0 represents the input image, and the hyperparameter constant α i = 1 - β i , β i is a constant term, a constant less than 1, and ∈ is Gaussian distribution noise; The mathematical expression of the reverse process of the diffusion model is: where x t-1 represents the denoised image obtained at the (t - 1)-th step, α t is a constant, ∈ θ (x t , t) represents the noise estimated by U-Net, σ t represents the variance of the denoising operation at the t-th step, and z represents Gaussian noise.

4. The intelligent fault diagnosis method for hydraulic systems based on the diffusion model data augmentation framework according to claim 1, wherein: In step 2, the diffusion model is used to perform data enhancement on the simulated grayscale image to obtain an enhanced hydraulic system fault data set, as follows: The data enhancement process includes diffusion model training and data enhancement; When training the diffusion model, all simulated grayscale images are sampled twice to obtain a uniformly distributed diffusion step size t and the corresponding Gaussian distribution noise ∈. Next, U-Net is used to estimate the noise added during the forward diffusion process, and the loss function is calculated based on the difference between the noise estimated by U-Net and the actual noise. Finally, the network parameters of the diffusion model are updated through back propagation according to the loss function value to obtain a trained diffusion model. When performing data augmentation, sample a noisy image x from a probability distribution that follows the standard normal distribution T , and use the inverse process of the trained diffusion model to denoise the image for up to T steps until an enhanced simulated grayscale image x0 is generated; The data enhancement process is repeated multiple times to obtain multiple different enhanced simulated grayscale images and an enhanced hydraulic system fault data set.

5. The intelligent fault diagnosis method for a hydraulic system based on a diffusion model data augmentation framework according to claim 1, wherein: In step 3, FADDAN is constructed, which includes a feature extractor, a label predictor, a gradient reversal layer, and two domain discriminators, as follows: The feature extractor includes a 3-layer convolutional neural network, a CBAM attention mechanism, and a 1-layer fully connected layer; the label predictor includes 2-layer fully connected layers; the two domain discriminators are a global domain discriminator and a local domain discriminator respectively, and the above two domain discriminators are both composed of 2-layer fully connected layers; during the forward propagation process, the gradient reversal layer acts as an identity mapping, and during the backward propagation process, this gradient reversal layer multiplies the gradient by the negative identity matrix -I to achieve the reversal of the gradient; The mathematical description of the forward process of the gradient reversal layer is: GRL(x) = x where x is the input image of the gradient reversal layer, and GRL represents the gradient reversal layer; The mathematical description of the backward process of the gradient reversal layer is: The gradient reversal layer is located between the feature extractor and the two domain discriminators, and the two domain discriminators share one gradient reversal layer.

6. The intelligent fault diagnosis method for a hydraulic system based on a diffusion model data augmentation framework according to claim 1, wherein: In step 5, the FADDAN is trained using the training set to obtain a trained fault diagnosis model, specifically as follows: For the training set x i , for the grayscale image sequence numbers \(i = 1,\cdots,a,\cdots,n,\cdots,N\), assume that the first \(n\) are all simulated grayscale images with classification labels, and the last \(n'\) are experimental grayscale images corresponding to the normal state without classification labels. The total number of grayscale images in the training set \(N=n + n'\). Among the first \(n\) simulated grayscale images with classification labels, the first \(a\) are simulated grayscale images corresponding to the normal state with classification labels; Training set x i Entering FADDAN, first, the first two convolutional neural networks in the feature extractor are passed through to extract low-level features, generating a preliminary feature map F. Then, the CBAM attention mechanism in the feature extractor is used to adaptively weight the feature map. Subsequently, the third convolutional neural network in the feature extractor is passed through to extract high-level features, and the core discriminant information is retained through the fully connected layer to obtain the final feature F out , and its mathematical description is: Among them, G f (·; θ f ) is a feature extractor, and θ f are the network parameters of the feature extractor; Final feature F out After the label predictor outputs the class prediction probability, the final feature F out After passing through the gradient reversal layer and the global domain discriminator, the global domain prediction probability is output, and the final feature F out The feature F extracted from the simulation grayscale image corresponding to the normal state in the training set and the experimental grayscale image corresponding to the normal state by the feature extractor o ′ ut After passing through the gradient reversal layer and the local domain discriminator, the local domain prediction probability is output; The loss function L of the label classifier during training cls is as follows: Among them, is the cross-entropy loss function, G y (·; θ y ) is the label classifier, θ y is the network parameter of the label predictor, y is the classification label, and y i is the classification label corresponding to x i ; The loss function L of the global domain discriminator during training dom1 is as follows: Among them, is the binary cross-entropy loss function, G d1 (·; θ d1 ) is the global domain discriminator, θ d1 are the network parameters of the global domain discriminator, is the gradient reversal layer, d1 is the domain label of the global domain discriminator, d 1i is for x i corresponding domain label of the global domain discriminator; The loss function \(L\) of the local domain discriminator during training dom2 is as follows: Among them, G d2 (·; θ d2 ) is the local domain discriminator, θ d2 are the network parameters of the local domain discriminator, d2 is the domain label of the local domain discriminator, d 2i is the local domain discriminator domain label corresponding to x i ; Total training loss function L total is as follows: L total = L cls + λ1L dom1 + λ2L dom2 where λ1 is the loss weight of the global domain discriminator, and λ2 is the loss weight of the local domain discriminator; after performing the above operations on all the images in the training set, a trained fault diagnosis model is obtained.

Images