Bearing cross-domain fault diagnosis system and method based on meta-learning domain adversarial graph convolutional network

Abstract

The invention discloses a bearing cross-domain fault diagnosis system and method based on a meta-learning domain adversarial graph convolutional network, and particularly relates to the technical field of mechanical fault diagnosis. Multi-source bearing vibration signals are integrated, and a cross-domain graph structure data set including node features and an adjacent matrix is constructed; performing adversarial training through a feature extractor and a domain classifier of the domain adversarial graph convolutional network, and combining a gradient inversion layer to extract domain invariant features; carrying out internal circulation task adaptation and external circulation element parameter updating by utilizing a element learning framework, and optimizing network parameters; and finally carrying out fault diagnosis on the target domain signal. And the total loss function of the system fuses task classification loss, domain adversarial loss and a graph structure regularization item, so that the cross-domain diagnosis precision is improved. The method effectively solves the problem of model generalization caused by domain difference, is suitable for bearing fault diagnosis scenes with few samples and multiple working conditions, and has the advantages of high robustness and high diagnosis precision.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and more specifically, to a bearing cross-domain fault diagnosis system and method based on a meta-learning domain adversarial graph convolutional network. Background Art

[0002] In the field of intelligent operation and maintenance of industrial equipment, bearings are key rotating components, and their fault diagnosis accuracy directly affects system reliability. Traditional machine learning-based fault diagnosis methods rely on a large amount of labeled data under a single working condition. When applied to cross-domain scenarios with different working conditions (such as speed and load changes) or different equipment, the model generalization ability is greatly reduced due to the significant difference in data distribution between the source domain and the target domain. In the existing technology, although deep learning-based methods can automatically extract features, they lack effective modeling of multi-sensor spatial correlations, and cross-domain migration requires reliance on a large amount of target domain labeled data, which makes it difficult to meet the actual needs of data scarcity in industrial scenarios.

[0003] In recent years, domain adversarial learning has improved cross-domain transfer capabilities by aligning feature distributions between the source and target domains. However, existing methods are mostly based on Euclidean spatial data, making them difficult to directly apply to unstructured, multi-sensor correlation scenarios. While graph convolutional networks (GCNs) can effectively model spatial relationships in graph-structured data, they lack the ability to align features across domains. Furthermore, traditional deep learning models are prone to overfitting in low-sample scenarios, resulting in low training efficiency and difficulty adapting to the demands of rapid deployment in industrial settings.

[0004] Therefore, the integration of graph-structured data modeling, domain adversarial learning, and meta-learning techniques to achieve structured utilization of multi-source data, extraction of domain-invariant features, and rapid adaptation to small-sample scenarios in cross-domain bearing fault diagnosis remains a pressing technical challenge. This paper proposes an efficient cross-domain fault diagnosis solution by constructing a meta-learning domain adversarial graph convolutional network, significantly improving the model's robustness and diagnostic accuracy in complex industrial environments. Summary of the Invention

[0005] In order to overcome the above-mentioned defects of the prior art, the present invention provides a bearing cross-domain fault diagnosis system and method based on a meta-learning domain adversarial graph convolutional network to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a bearing cross-domain fault diagnosis method based on a meta-learning domain adversarial graph convolutional network, comprising the following steps:

[0007] Acquire multi-source bearing vibration signals and construct a cross-domain graph structure dataset;

[0008] Inputting the graph structure data into a domain adversarial graph convolutional network, and extracting domain invariant features through adversarial training of a domain classifier and a feature extractor;

[0009] A meta-learning framework is used to optimize the domain adversarial graph convolutional network, including inner-loop task adaptation and outer-loop meta-parameter update.

[0010] Fault diagnosis of bearing vibration signals in the target domain is performed based on the optimized network;

[0011] Among them, the total loss function L of the domain adversarial graph convolutional network total for:

[0012]

[0013] Where, L task is the task classification loss, L domain is the domain adversarial loss, L graph is a regularization term based on graph structure, and γ are hyperparameters.

[0014] Preferably, constructing a cross-domain graph structure dataset includes:

[0015] Convert the vibration signal of each sensor into node features, which include time domain statistics, frequency domain energy and wavelet packet coefficients;

[0016] Construct an adjacency matrix based on the physical layout of sensors or signal correlation, and define the edge weights as:

[0017]

[0018] Among them, f i and f j is the eigenvector of nodes i and j, and σ is the scale parameter.

[0019] Preferably, the domain adversarial loss is implemented by a gradient reversal layer, and its expression is:

[0020]

[0021] Among them, d k is the domain label (source domain is 0, target domain is 1), D(·) is the domain classifier, h k It is the feature representation output by the feature extractor.

[0022] Preferably, the graph structure-based regularization term is defined as:

[0023]

[0024] Used to constrain the smoothness of adjacent node features.

[0025] Preferably, the inner loop task adaptation process of the meta-learning framework is:

[0026] For each meta-task, the initial parameters are updated by gradient descent as:

[0027]

[0028] The outer loop meta parameters are updated as follows:

[0029]

[0030] Among them, α and β are the inner and outer loop learning rates respectively.

[0031] A system for executing the above-mentioned bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network comprises:

[0032] A graph structure building module for converting multi-source vibration signals into graph structure data;

[0033] The domain adversarial graph convolutional network module includes a feature extractor, a task classifier, and a domain classifier, and extracts domain-invariant features through adversarial training;

[0034] A meta-learning optimization module, configured to optimize network parameters through inner-loop task adaptation and outer-loop meta-parameter update;

[0035] The fault diagnosis module is used to output the fault category for the target domain data.

[0036] Preferably, the feature extractor adopts a multi-layer graph convolutional network, and each layer is calculated as:

[0037]

[0038] in, is the normalized adjacency matrix, W ( l ) is a trainable parameter, is the activation function.

[0039] A computer-readable storage medium stores a program, which implements the above-mentioned method when executed by a processor.

[0040] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the program.

[0041] Technical effects and advantages of the present invention:

[0042] 1. By combining domain adversarial learning with meta-learning, the distribution difference between the source domain and the target domain is significantly reduced. This solves the problem of decreased diagnostic accuracy caused by different data distribution in traditional methods across working conditions and equipment scenarios, achieves domain-invariant extraction of bearing fault features, and improves model generalization capabilities.

[0043] 2. Convert multi-sensor vibration signals into graph-structured data, characterize the physical layout of sensors or signal correlation through the adjacency matrix, effectively capture spatial correlation features, and combine the hierarchical feature extraction capabilities of the graph convolutional network (GCN) to improve the characterization accuracy of complex vibration signals.

[0044] 3. The meta-learning framework enables the model to converge efficiently in a small number of sample scenarios through rapid adaptation of inner-loop tasks and update of outer-loop meta-parameters, reducing dependence on large-scale labeled data. At the same time, the graph structure regularization term is used to constrain the smoothness of adjacent node features, preventing overfitting and enhancing model stability.

[0045] 4. The system adopts a modular design that combines graph structure construction, domain adversarial network, meta-learning optimization, and fault diagnosis. Each module has a clear division of labor and strong collaboration, which facilitates engineering implementation and maintenance and is suitable for real-time fault diagnosis needs in industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the diagnostic method of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] As attached Figure 1 The bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network shown in the figure includes the following steps:

[0049] Acquire multi-source bearing vibration signals and construct a cross-domain graph structure dataset;

[0050] Inputting the graph structure data into a domain adversarial graph convolutional network, and extracting domain invariant features through adversarial training of a domain classifier and a feature extractor;

[0051] A meta-learning framework is used to optimize the domain adversarial graph convolutional network, including inner-loop task adaptation and outer-loop meta-parameter update.

[0052] Fault diagnosis of bearing vibration signals in the target domain is performed based on the optimized network;

[0053] Among them, the total loss function L of the domain adversarial graph convolutional network total for:

[0054]

[0055] Where, L task is the task classification loss, L domain is the domain adversarial loss, L graph is a regularization term based on graph structure, and γ are hyperparameters.

[0056] In the specific implementation, the multi-source bearing vibration signals are constructed into a cross-domain graph structure dataset, so that the spatial correlation between sensors can be captured with the help of the graph structure. Then, the domain adversarial graph convolutional network is used to extract features that do not change with the domain through adversarial training of feature extractors and domain classifiers, thereby reducing the distribution difference between the source domain and the target domain. Finally, the network is optimized using the meta-learning framework. The inner loop completes task adaptation and the outer loop realizes meta-parameter update, thereby improving the generalization ability of the model in cross-domain scenarios, effectively solving the problem of insufficient model generalization ability caused by domain differences in cross-domain fault diagnosis, and significantly improving the fault diagnosis accuracy of bearings in different working conditions or equipment.

[0057] The constructing of a cross-domain graph structure dataset includes:

[0058] Convert the vibration signal of each sensor into node features, which include time domain statistics, frequency domain energy and wavelet packet coefficients;

[0059] Construct an adjacency matrix based on the physical layout of sensors or signal correlation, and define the edge weights as:

[0060]

[0061] Among them, f i and f j is the eigenvector of nodes i and j, and σ is the scale parameter.

[0062] In specific implementation, the vibration signal of each sensor is converted into node features containing time domain statistics, frequency domain energy, and wavelet packet coefficients to fully describe the signal characteristics. Based on the physical layout of the sensors or signal correlation, the edge weights are calculated using the Gaussian kernel function to construct the adjacency matrix. The formula is: In this way, the similarity between nodes is preserved, and the spatial correlation and signal feature similarity between multiple sensors can be effectively captured through the graph structure, providing a more structured input for the subsequent graph convolutional network, thereby improving the efficiency of feature extraction.

[0063] The domain adversarial loss is implemented through the gradient reversal layer, and its expression is:

[0064]

[0065] Among them, d kis the domain label (source domain is 0, target domain is 1), D(·) is the domain classifier, h k It is the feature representation output by the feature extractor.

[0066] In specific implementation, the gradient reversal layer (GRL) is used to implement the domain adversarial loss, which is expressed as During training, the gradient reversal layer reverses the gradient, forcing the feature extractor to generate features that are difficult to distinguish by the domain classifier, thereby learning domain-invariant features, which effectively reduces the distribution difference between the source domain and the target domain, allowing the model to achieve effective knowledge transfer in the target domain (even if labeled data is scarce).

[0067] The graph-based regularization term is defined as:

[0068]

[0069] Used to constrain the smoothness of adjacent node features.

[0070] In specific implementation, the regularization term is used to constrain the smoothness of adjacent node features, so that similar nodes are closer in the feature space, which conforms to the local smoothness assumption of the graph. This can maintain the geometric properties of the graph structure, improve the coherence of the features, effectively prevent model overfitting, and enhance the model's adaptability to graph structure data.

[0071] The inner loop task adaptation process of the meta-learning framework is:

[0072] For each meta-task, the initial parameters are updated by gradient descent as:

[0073]

[0074] The outer loop meta parameters are updated as follows:

[0075]

[0076] Among them, α and β are the inner and outer loop learning rates respectively.

[0077] In specific implementation, the inner loop is implemented by gradient descent Fast adaptation to a single meta-task, the outer loop is based on the meta-loss of multiple tasks:

[0078]

[0079] Updating meta-parameters enables the model to have the ability to "learn to learn", thereby significantly improving the model's adaptability in small-sample scenarios, accelerating model convergence, and improving the robustness of cross-domain diagnosis.

[0080] A system for cross-domain bearing fault diagnosis based on meta-learning domain adversarial graph convolutional networks, comprising:

[0081] A graph structure building module for converting multi-source vibration signals into graph structure data;

[0082] The domain adversarial graph convolutional network module includes a feature extractor, a task classifier, and a domain classifier, and extracts domain-invariant features through adversarial training;

[0083] A meta-learning optimization module, configured to optimize network parameters through inner-loop task adaptation and outer-loop meta-parameter update;

[0084] The fault diagnosis module is used to output the fault category for the target domain data.

[0085] The feature extractor uses a multi-layer graph convolutional network, and each layer is calculated as:

[0086]

[0087] in, is the normalized adjacency matrix, W ( l ) is a trainable parameter, is the activation function.

[0088] In the specific implementation, a multi-layer graph convolutional network is used, and the calculation method of each layer is By normalizing the adjacency matrix and trainable parameters W ( l ) Perform convolution operation and then pass through activation function By introducing nonlinearity, hierarchical graph features can be extracted, so that more abstract features can be extracted layer by layer, effectively capturing the complex relationships in the graph structure, and thus improving the accuracy of fault diagnosis.

[0089] A computer-readable storage medium stores a program, which implements the above-mentioned method when executed by a processor.

[0090] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor implements the above-described method when executing the program.

[0091] Specifically, the present invention provides a bearing cross-domain fault diagnosis method based on a meta-learning domain adversarial graph convolutional network, comprising the following steps:

[0092] Step 1: Build a cross-domain graph structure dataset

[0093] Data acquisition: Multiple sensors are used to collect vibration signals of bearings under different operating conditions (such as different speeds and loads) to form a multi-source dataset, which contains source domain (labeled) and target domain (no / few label) data.

[0094] Feature Engineering:

[0095] For the vibration signal of each sensor, time domain statistics such as mean, variance, kurtosis, etc. are calculated to describe the time domain distribution characteristics of the signal.

[0096] Perform Fourier transform on the signal and calculate the frequency domain energy to capture the fault characteristic frequency.

[0097] The wavelet packet decomposition method is used to extract the wavelet packet coefficients to characterize the energy distribution of the signal in different frequency bands.

[0098] Graph structure construction:

[0099] Node setting: The feature vector of each sensor is used as the node of the graph, that is, the feature of node i is f i .

[0100] Adjacency matrix construction:

[0101] If the adjacency matrix is ​​constructed based on the physical layout of the sensors, the edge weight is set to a non-zero value when two sensors are physically adjacent. If the adjacency matrix is ​​constructed based on signal correlation, the Gaussian kernel function is used to calculate the edge weight A. ij , where the scale parameter σ can be optimized through cross-validation.

[0102] Normalize the adjacency matrix and get in (I is the identity matrix), is a diagonal matrix whose elements

[0103] Step 2: Domain-Adversarial Graph Convolutional Network Training

[0104] Network architecture:

[0105] Feature extractor: consists of multiple layers of graph convolutional layers. The calculation of each layer is as described in claim 7. For example, 2-3 layers are set, the output dimension of the first layer is set to 64, and the output dimension of the second layer is set to 32.

[0106] Task classifier: It is a fully connected layer used to classify faults based on the extracted features. The output dimension is the number of fault categories.

[0107] Domain classifier: It is also a fully connected layer used to determine whether the feature comes from the source domain or the target domain, with an output dimension of 2 (activated by softmax).

[0108] Adversarial Training:

[0109] The goal of the feature extractor is to minimize the task classification loss L task(such as cross entropy loss), while deceiving the domain classifier through the gradient reversal layer, that is, minimizing the domain adversarial loss L domain .

[0110] The goal of the domain classifier is to maximize L domain , thereby accurately distinguishing the source domain and target domain features.

[0111] The total loss function is The hyperparameters and γ are determined by grid search.

[0112] Step 3: Meta-learning framework optimization

[0113] Inner loop task adaptation:

[0114] For each meta-task (a few-shot task sampled from the source domain), we use the initial meta-parameters (θ) to perform gradient descent and update the parameters to The learning rate α is set to 0.01-0.1.

[0115] Outer loop meta parameter update:

[0116] Calculate the average loss of multiple meta-tasks and then update the meta-parameters

[0117]

[0118] Among them, the learning rate β is set to 0.001-0.01.

[0119] Repeat the inner and outer loop processes until the model converges.

[0120] Step 4: Target domain fault diagnosis

[0121] The vibration signal of the target domain is constructed into graph structure data according to step 1.

[0122] The constructed data is input into the optimized domain adversarial graph convolutional network.

[0123] The fault category probability is output by the task classifier, and the category with the highest probability is selected as the diagnosis result.

[0124] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network, characterized by: The following steps are involved: Acquire multi-source bearing vibration signals and construct a cross-domain graph structure dataset; Inputting the graph structure data into a domain adversarial graph convolutional network, and extracting domain invariant features through adversarial training of a domain classifier and a feature extractor; A meta-learning framework is used to optimize the domain adversarial graph convolutional network, including inner-loop task adaptation and outer-loop meta-parameter update. Fault diagnosis of bearing vibration signals in the target domain is performed based on the optimized network; Among them, the total loss function L of the domain adversarial graph convolutional network total for: Where, L task is the task classification loss, L domain is the domain adversarial loss, L graph is a regularization term based on graph structure, and γ are hyperparameters.

2. The bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network according to claim 1 is characterized in that: The constructing of a cross-domain graph structure dataset includes: The vibration signal of each sensor is converted into node features, which include time domain statistics, frequency domain energy and wavelet packet coefficients; Construct an adjacency matrix based on the physical layout of sensors or signal correlation, and define the edge weights as: Among them, f i and f j is the eigenvector of nodes i and j, and σ is the scale parameter.

3. The bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network according to claim 1 is characterized in that: The domain adversarial loss is implemented through the gradient reversal layer, and its expression is: Among them, d k is the domain label (source domain is 0, target domain is 1), D(·) is the domain classifier, h k It is the feature representation output by the feature extractor.

4. The bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network according to claim 1 is characterized in that: The graph-based regularization term is defined as: Used to constrain the smoothness of adjacent node features.

5. The bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network according to claim 1 is characterized in that: The inner loop task adaptation process of the meta-learning framework is: For each meta-task, the initial parameters are updated by gradient descent as: The outer loop meta parameters are updated as follows: Among them, α and β are the inner and outer loop learning rates respectively.

6. A system for executing the bearing cross-domain fault diagnosis method based on meta-learning domain adversarial graph convolutional network according to any one of claims 1 to 5, characterized in that: include: A graph structure building module for converting multi-source vibration signals into graph structure data; The domain adversarial graph convolutional network module includes a feature extractor, a task classifier, and a domain classifier, and extracts domain-invariant features through adversarial training; A meta-learning optimization module, configured to optimize network parameters through inner-loop task adaptation and outer-loop meta-parameter update; The fault diagnosis module is used to output the fault category for the target domain data.

7. The bearing cross-domain fault diagnosis system based on meta-learning domain adversarial graph convolutional network according to claim 6 is characterized in that: The feature extractor uses a multi-layer graph convolutional network, and each layer is calculated as: in, is the normalized adjacency matrix, W (l) is a trainable parameter, is the activation function.

8. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

9. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 5 is implemented.

Images