Cross-domain mechanical equipment fault diagnosis method and device based on adversarial united domain adaptive graph convolutional network

Abstract

The invention discloses a cross-domain mechanical equipment fault diagnosis method and device based on an adversarial united domain adaptive graph convolutional network, and the method comprises the steps: inputting a vibration signal of to-be-diagnosed mechanical equipment into a pre-trained mechanical equipment fault diagnosis model, and outputting a diagnosis result; wherein the mechanical equipment fault diagnosis model is obtained by training a resistance-based joint domain adaptive graph convolutional network by adopting training data; the adversarial-based joint domain adaptive graph convolutional network comprises a multi-scale feature extraction module with a mixed attention mechanism, a graph structure feature learning module and an improved joint domain adaptive module which are connected in sequence. The multi-scale feature extraction module with the mixed attention mechanism comprises a multi-scale convolution layer with three different convolution kernel sizes, a mixed attention mechanism module, three convolution modules connected in sequence and a full connection layer which are connected in sequence; the graph structure feature learning module comprises a graph construction module and a feature learning module of a double-layer GCN structure, and the graph construction module comprises a self-adaptive similar graph construction module based on an inner product kernel and a graph structure feature extraction module; a loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

Description

Technical Field

[0001] The present invention belongs to the field of cross - domain fault diagnosis, and particularly relates to a cross - domain mechanical equipment fault diagnosis method and device based on an adversarial joint domain adaptive graph convolutional network. Background Art

[0002] As a key component of modern manufacturing systems, mechanical equipment plays an indispensable role in industrial production and economic development. Due to the harsh operating environment and long - term operation, mechanical equipment will inevitably suffer damage and failures, which may lead to downtime losses and even serious accidents. As a part of the Prognostics and Health Management (PHM) technology, intelligent fault diagnosis is the key to ensuring the safe and stable operation of mechanical equipment.

[0003] The advent of the industrial big data era has injected new vitality into intelligent fault diagnosis. In particular, fault diagnosis methods based on deep learning (DL) have become a research hotspot due to their powerful feature representation capabilities. However, due to the high labor costs and limitations of real industries, it is difficult or even impossible to obtain a sufficient amount of labeled fault data. In addition, in real industrial production, due to the change of working conditions, variable - working - condition scenarios are quite common, which inevitably leads to the problem of covariance drift. Therefore, fault diagnosis methods based on transfer learning have emerged.

[0004] Unsupervised domain adaptation (UDA), as a typical technology in transfer learning, has been widely used to solve the above problems. Generally speaking, UDA can be divided into four categories: model - based methods, instance - based methods, mapping - based methods, and adversarial - based methods. Among them, the adversarial - based method has received extensive attention because it can extract more effective and highly robust features. In the adversarial - based UDA method, the same - class labels from different domains are the basis for the application of UDA technology, which ensures that data from different domains can be mapped to the same feature space. In addition, in order to help the feature extractor learn domain - invariant features, a domain classifier is trained based on domain labels. Moreover, data structured information, such as probability distribution, data statistical information, geometric data structure, etc., can be used to reflect the inherent attributes of datasets from different domains. On this basis, by means of three key category information, namely class labels, domain labels, and data structured information, the adversarial - based UDA method builds a bridge between the source domain and the target domain, reducing the distribution difference between different domains.

[0005] However, effectively modeling data structured information and integrating it into a deep - learning model is a challenging task, and most of the existing studies do not consider these factors, which may lead to poor generalization performance of the model. Therefore, how to effectively model data structured information has become an urgent and challenging task in industrial practice and the research community.

[0006] As an extension of DL technology in non-Euclidean space, the Graph Convolutional Network (GCN) has gradually become a research hotspot, providing a new method for effectively extracting the structured information of data. In related research, graph construction and model design are two key issues. In terms of graph construction, the performance of GCN is easily affected by the quality of graph construction. When the constructed graph contains noise and some edges are missing, the bias may be spread and amplified during the process of aggregating adjacent node information, resulting in poor quality of the extracted features and negative transfer effects. In addition, high complexity and low scalability are also two potential problems of graph neural network models, which pose challenges to model design. Summary of the Invention

[0007] Aiming at the problems existing in the prior art, the present invention provides a cross-domain mechanical equipment fault diagnosis method and device based on an adversarial joint domain adaptive graph convolutional network.

[0008] To solve the above technical problems, the present invention is realized through the following technical solutions:

[0009] According to the first aspect of the present invention, there is provided a cross-domain mechanical equipment fault diagnosis method based on an adversarial joint domain adaptive graph convolutional network, including:

[0010] Obtain the vibration signal of the mechanical equipment to be diagnosed;

[0011] Input the vibration signal of the mechanical equipment to be diagnosed into a pre-trained mechanical equipment fault diagnosis model to output a diagnosis result; wherein, the mechanical equipment fault diagnosis model is obtained by training an adversarial joint domain adaptive graph convolutional network using training data, the training data includes source domain data and target domain data, the source domain data includes non-fault vibration signals and different fault vibration signals with different fault labels, and the target domain data includes non-fault vibration signals and different fault vibration signals without fault labels; the adversarial joint domain adaptive graph convolutional network includes a multi-scale feature extraction module with a hybrid attention mechanism, a graph structure feature learning module, and an improved joint domain adaptive module connected in sequence. The multi-scale feature extraction module with a hybrid attention mechanism includes a multi-scale convolutional layer with three different convolutional kernel sizes, a hybrid attention mechanism module, three sequentially connected convolutional modules, and a fully connected layer; the graph structure feature learning module includes a graph construction module and a feature learning module with a double-layer GCN structure. The graph construction module includes an adaptive similarity graph construction module based on an inner product kernel and a graph structure feature extraction module; the loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

[0012] In a possible implementation of the first aspect, regarding the multi-scale feature extraction module with a hybrid attention mechanism, specifically:

[0013] The multi-scale convolutional layer with three different convolutional kernel sizes is used to extract different-scale features in the vibration signal;

[0014] The hybrid attention mechanism module is used to assign weights to different-scale features;

[0015] The three sequentially connected convolutional modules are used to deeply extract different-scale features after weight assignment to obtain different-scale features after deep extraction; the fully connected layer is used to transform the different-scale features after deep extraction to obtain different-scale features after transformation.

[0016] In a possible implementation of the first aspect, regarding the graph structure feature learning module, specifically:

[0017] The adaptive similarity graph construction module based on the inner product kernel is used to construct a graph for the different-scale features after transformation by using the adaptive similarity graph construction method based on the inner product kernel to obtain the optimal graph;

[0018] The graph structure feature extraction module is used to adaptively extract graph structure features from the optimal graph by using the GraphSAGE method;

[0019] The feature learning module with a double-layer GCN structure is used to input the extracted graph structure features into the double-layer GCN to obtain structured information.

[0020] In a possible implementation of the first aspect, regarding the improved joint domain adaptation module, specifically:

[0021] The improved joint domain adaptation module is used to transform the structured information into domain-invariant features and perform fault diagnosis based on the domain-invariant features.

[0022] In a possible implementation of the first aspect, the improved joint domain adaptation loss function, specifically:

[0023]

[0024] Where, is the improved joint domain adaptation loss; is the cross-entropy loss; is the deep coral loss; is the entropy conditional adversarial domain adaptation loss; is the loss based on the multi-kernel maximum mean discrepancy; a, b, c are trade-off parameters; m is the size of the mini-batch; represents the k-th true label; is the output vector of the k-th softmax layer; d is the dimension of the feature; represents the Frobenius norm of the square matrix; CE(·,·) represents the cross-entropy loss; and respectively represent the classification prediction probability of the i-th sample input from the source domain dataset to the model and the true label of the i-th sample input from the source domain dataset to the model; represents the classification prediction probability of the j-th sample input from the target domain dataset to the model; C s is the covariance matrix of the source domain features; C t is the covariance matrix of the target domain features; F(·) is the feature extractor; Φ(·) represents the non-linear mapping function, and H represents the distance calculated after mapping the features to the reproducing Hilbert space; is a sample collected from the source distribution D s ; is a sample collected from the target distribution D t ; D s and D t respectively represent the probability distributions of the source domain and the target domain; represents the expected value, that is, the average value of the random variable; θ ∈ [0,1] is a factor that changes with the training process; w(·) represents the weight generated by the entropy; represents the deep feature of the j-th sample input from the target domain dataset to the model; represents the deep feature of the i-th sample input from the source domain dataset to the model; λ represents the weight parameter for balancing the two loss terms; D(.) represents the domain discrimination operation of the domain discriminator; G(.) is the classification operation of the source classifier; H(·) represents the entropy of the sample; T(·) represents the multi-linear mapping result of the sample; M(.) is the result after mapping the sample.

[0025] In a possible implementation manner of the first aspect, the improved joint domain adaptation module is sequentially stacked by a global average pooling layer, three fully connected layers, and a SoftMax layer.

[0026] According to the second aspect of the present invention, there is provided a cross-domain mechanical equipment fault diagnosis device based on an adversarial joint domain adaptation graph convolutional network, including:

[0027] An acquisition module for acquiring the vibration signal of the mechanical equipment to be diagnosed;

[0028] A diagnostic module, configured to input the vibration signal of the mechanical equipment to be diagnosed into a pre-trained mechanical equipment fault diagnosis model and output a diagnosis result; wherein, the mechanical equipment fault diagnosis model is obtained by training a joint domain adaptive graph convolutional network based on adversarial training using training data, the training data includes source domain data and target domain data, the source domain data includes non-fault vibration signals and different fault vibration signals with different fault labels, and the target domain data includes non-fault vibration signals and different fault vibration signals without fault labels; the joint domain adaptive graph convolutional network based on adversarial training includes a multi-scale feature extraction module with a hybrid attention mechanism, a graph structure feature learning module, and an improved joint domain adaptive module connected in sequence. The multi-scale feature extraction module with a hybrid attention mechanism includes a multi-scale convolutional layer with three different convolutional kernel sizes, a hybrid attention mechanism module, three sequentially connected convolutional modules, and a fully connected layer; the graph structure feature learning module includes a graph construction module and a feature learning module with a two-layer GCN structure, and the graph construction module includes an adaptive similarity graph construction module based on an inner product kernel and a graph structure feature extraction module; the loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

[0029] According to a third aspect of the present invention, there is provided a device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the cross-domain mechanical equipment fault diagnosis method based on the joint domain adaptive graph convolutional network based on adversarial training as described above is implemented.

[0030] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium storing a computer program, and when the computer program is executed by a processor, the cross-domain mechanical equipment fault diagnosis method based on the joint domain adaptive graph convolutional network based on adversarial training as described above is implemented.

[0031] According to a fifth aspect of the present invention, there is provided a computer program product, and when the computer program product is executed by a processor, the cross-domain mechanical equipment fault diagnosis method based on the joint domain adaptive graph convolutional network based on adversarial training as described above is implemented.

[0032] Compared with the prior art, the present invention has at least the following beneficial effects:

[0033] The present invention relates to a cross - domain mechanical equipment fault diagnosis method based on an adversarial joint - domain adaptive graph convolutional network. By designing a multi - scale feature extraction module with a hybrid attention mechanism, it can extract multi - scale features from the vibration signals of mechanical equipment, reduce the loss of key information in the original signals, and thus improve the accuracy of fault diagnosis. This module can adaptively focus on the important parts of the signals and accurately diagnose different types of faults. The adaptive similarity graph construction module based on the inner - product kernel uses an adaptive similarity graph construction method to effectively eliminate noise interference and edge omission problems, avoid bias diffusion and amplification during the feature extraction process, enabling the model to accurately diagnose faults under complex and variable working conditions. In the graph structure feature learning module, the graph construction module and the graph structure feature extraction module adopt the GraphSAGE method to adaptively extract graph structure features from the optimal graph, enhancing the scalability of the model and enabling the model to flexibly handle mechanical equipment fault diagnosis tasks of different scales and complexities. The improved joint - domain adaptive module uses an improved joint - domain adaptive loss function. By proposing a loss strategy based on multi - kernel maximum mean discrepancy and combining it with other loss strategies, it effectively reduces the possible negative transfer effect in the model.

[0034] To make the above - mentioned objects, features, and advantages of the present invention more obvious and understandable, the following specifically enumerates preferred embodiments and, in conjunction with the accompanying drawings, provides detailed descriptions as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the following will briefly introduce the drawings required for use in the description of the specific embodiments. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0036] Figure 1 is a flowchart of a cross - domain mechanical equipment fault diagnosis method based on an adversarial joint - domain adaptive graph convolutional network according to an embodiment of the present invention;

[0037] Figure 2 is the specific structure of the adversarial joint - domain adaptive graph convolutional network proposed by the present invention;

[0038] Figure 3 is the technical flowchart of an embodiment of the present invention;

[0039] Figure 4 is the specific structure of the channel attention network of HAM according to an embodiment of the present invention;

[0040] Figure 5 is the specific structure of the spatial attention network of HAM according to an embodiment of the present invention;

[0041] Figure 6 The operation process of GraphSAGE in the embodiments of the present invention;

[0042] Figure 7 The composition structure of the wind turbine power transmission system diagnostic simulator;

[0043] Figure 8 The diagnostic results of the XJTU dataset;

[0044] Figure 9 The feature visualization of the XJTU dataset, (a) ResNet18. (b) D-CORAL. (c) DANN. (d) DDC. (e) Proposed;

[0045] Figure 10 The confusion matrix analysis of the XJTU dataset. (a) ResNet18. (b) D-CORAL. (c) DANN. (d) DDC. (e) Proposed. Specific embodiments

[0046] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0047] As Figure 1 shown, the embodiments of the present invention provide a cross-domain mechanical equipment fault diagnosis method based on an adversarial joint domain adaptation graph convolutional network, which specifically includes the following steps:

[0048] Step 1, obtain the vibration signal of the mechanical equipment to be diagnosed.

[0049] Specifically, obtain the vibration signal from the mechanical equipment to be diagnosed. Exemplarily, these vibration signals are collected by sensors and undergo preprocessing steps such as denoising, filtering, and normalization to ensure data quality. The preprocessed vibration signals will be used as the input for the subsequent fault diagnosis model.

[0050] Step 2, input the vibration signal of the mechanical equipment to be diagnosed into a pre-trained mechanical equipment fault diagnosis model, and output the diagnostic result.

[0051] Among them, the mechanical equipment fault diagnosis model is obtained by training an adversarial joint domain adaptive graph convolutional network (AJDAGCN) with training data. The training data includes source domain data and target domain data. The source domain data includes non-fault vibration signals with non-fault labels and different fault vibration signals with different fault labels. The target domain data includes non-fault vibration signals and different fault vibration signals without fault labels. As Figure 2 and Figure 3 shown, the adversarial joint domain adaptive graph convolutional network includes a multi-scale feature extraction module with a hybrid attention mechanism, a graph structure feature learning module, and an improved joint domain adaptive module connected in sequence. The multi-scale feature extraction module with a hybrid attention mechanism includes a multi-scale convolutional layer with three different convolutional kernel sizes, a hybrid attention mechanism module, three sequentially connected convolutional modules, and a fully connected layer connected in sequence; the graph structure feature learning module includes a graph construction module and a feature learning module with a double-layer GCN structure. The graph construction module includes an adaptive similarity graph construction module based on an inner product kernel and a graph structure feature extraction module; the loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

[0052] Specifically, the training data set includes source domain data and target domain data. The source domain data contains vibration signals with known fault labels (including non-fault vibration signals and fault vibration signals of different fault types), while the target domain data contains non-fault vibration signals and fault vibration signals with unknown fault labels. The training data set is used to train the model so that it can accurately identify faults under different working conditions. The multi-scale feature extraction module first performs preliminary feature extraction on the input signal through a multi-scale convolutional layer with three different convolutional kernel sizes. Then, the hybrid attention mechanism module is used to weight the extracted features to emphasize important features and suppress noise. Next, the features are further refined through three sequentially connected convolutional modules, and finally, a feature vector is output through the fully connected layer. In the graph construction module, the adaptive similarity graph construction module based on an inner product kernel constructs a graph structure according to the similarity between feature vectors, where nodes represent feature vectors and edges represent the similarity between nodes. Then, the feature learning module with a double-layer GCN structure is used to further learn and extract the features in the graph structure to capture the data structured information.

[0053] More specifically, regarding the multi-scale feature extraction module with the hybrid attention mechanism, the multi-scale convolution layer with three different convolution kernel sizes is used to extract different-scale features in the vibration signal; the hybrid attention mechanism module is used to assign weights to different-scale features; the three sequentially connected convolution modules are used to deeply extract the different-scale features after the weights are assigned, and obtain the different-scale features after deep extraction; the fully connected layer is used to transform the different-scale features after deep extraction, and obtain the transformed different-scale features.

[0054] In the processing of vibration signals, features of different scales often contain different fault information. In order to fully capture this information, the present invention designs a multi-scale convolution layer with three different convolution kernel sizes (64, 32, 16). Through this layer, the vibration signal is converted into feature maps of multiple scales. After obtaining the feature maps of multiple scales, the present invention introduces a hybrid attention mechanism module to weight these features. The hybrid attention mechanism module can quickly locate the important parts in a large amount of information. In the present invention, the hybrid attention mechanism module calculates the correlation between features of different scales and assigns a weight to each feature, and the weight reflects the importance of the feature in the subsequent diagnosis process. In this way, the model can focus more on key features, thereby improving the accuracy of diagnosis. After weighting by the hybrid attention mechanism module, the features are sent to three sequentially connected convolution modules for deep extraction, and the convolution modules further refine the features to extract more refined and discriminative information. It should be understood that each convolution module includes a convolution layer, an activation function and a pooling layer to ensure effective feature extraction and dimensionality reduction. Through three sequentially connected convolution modules, the model can gradually dig out deep fault information hidden in the vibration signal. Finally, the features of different scales after deep extraction are sent to the fully connected layer for transformation. The function of the fully connected layer is to map these features into a new feature space. In the fully connected layer, each feature is converted into a vector of fixed length. These vectors contain the compressed representation of all important information in the vibration signal. These vectors will be used as the input of the graph structure feature learning module for further feature extraction and domain adaptation processing. In summary, the multi-scale feature extraction module with hybrid attention mechanism realizes the comprehensive extraction and weighted processing of features of different scales in the vibration signal through the collaborative work of multi-scale convolutional layers, hybrid attention mechanism modules, sequentially connected convolutional modules and fully connected layers, thereby improving the efficiency and accuracy of feature extraction.

[0055] For example, the features of different scales are cascaded to achieve preliminary fusion, which can be defined as: F multiscale =Concat(F 1 ,F 2 ,F 3 ), where Fmultiscale is the feature after preliminary fusion, F 1 , F 2 and F 3 represent features obtained at different scales.

[0056] Exemplarily, as shown in combination with Figure 4 and Figure 5 , the Hybrid Attention Mechanism (HAM) module includes a channel attention module and a spatial attention module, specifically as follows:

[0057] 1) Channel attention module: First, perform average pooling and max pooling operations on the input feature F to obtain and Then set the learnable parameters α and β, and perform adaptive fusion on and Finally, perform a fast one-dimensional convolution operation C1D on the fusion result 1×k . After further non-linear transformation of the result with an activation function, the final attention weight σ is obtained. The above process can be described as:

[0058]

[0059] where k is the kernel size.

[0060] 2) Spatial attention module: According to the channel attention weights given to each channel, divide the important group F 1 ′ and the less important group F′ 2 . Through multiple experimental tests, the ratio of F 1 ′ is finally determined to be 0.6. Then, similar to the operation of the channel attention module, the spatial attention weights A s,1 and A s,2 are respectively obtained for F 1 ′ and F′ 2 . The above process can be described as:

[0061] A s,1 (F′) = Φ(C1D 1×7 ([Avgpool(F 1 ′); MaxPool(F 1 ′)]))

[0062] A s,2 (F′) = Φ(C1D 1×7 ([Avgpool(F′ 2 ); MaxPool(F′ 2 )]))

[0063] Among them, Φ represents a non-linear operation, including batch normalization and activation functions.

[0064] Given a batch of samples X batch , the whole process of extracting its node features F can be expressed as:

[0065] F = MSCNN-HAM(X batch )

[0066] More specifically, regarding the graph structure feature learning module, the adaptive similarity graph construction module based on the inner product kernel is used to construct a graph for the transformed features of different scales by using the adaptive similarity graph construction method based on the inner product kernel to obtain the optimal graph; the graph structure feature extraction module is used to adaptively extract graph structure features from the optimal graph by using the GraphSAGE method; the feature learning module with a double-layer GCN structure is used to input the extracted graph structure features into the double-layer GCN to obtain structured information.

[0067] In the graph structure feature learning module, first, the adaptive similarity graph construction module based on the inner product kernel is used to construct a graph for the transformed features of different scales. The core is to measure the similarity between them by calculating the inner product between feature vectors and construct a graph structure accordingly. Specifically, for each feature vector, calculate the inner product between it and all other feature vectors, and construct the weight of the edge according to the magnitude of the inner product. The larger the weight, the higher the similarity between the two feature vectors. In this way, a weighted graph can be obtained, where the nodes represent feature vectors and the edges represent the similarity relationship between feature vectors. Since the features of vibration signals are often very complex and variable, the directly constructed graph may contain noise and redundant information. Therefore, the present invention proposes an adaptive similarity graph construction method, which can automatically adjust the graph construction parameters according to the characteristics of the data, such as the sparsity of the graph, to obtain the optimal graph and ensure that the constructed graph can accurately reflect the similarity relationship between feature vectors. After obtaining the optimal graph, then use the graph structure feature extraction module to adaptively extract graph structure features from the graph. Combined with Figure 6As shown in the figure, the GraphSAGE method is adopted. The GraphSAGE method is a graph neural network algorithm based on neighbor aggregation, which can effectively extract the feature representation of nodes from a graph. GraphSAGE learns the embedding representation of nodes by aggregating the neighbor information of nodes. This representation method not only contains the information of the nodes themselves, but also contains the information of their neighbor nodes, thus being able to capture the structural information of the graph. In the present invention, the GraphSAGE method is used to extract the embedding representation of each node from the optimal graph, and these embedding representations are the graph structure features. Through this step, the original high-dimensional feature vectors can be transformed into low-dimensional graph structure features with stronger representation ability. Finally, the extracted graph structure features are input into the feature learning module of the double-layer GCN structure for further learning. GCN (Graph Convolutional Network) can extract local features in the graph through convolutional operations and capture global features by stacking multiple convolutional layers. In the present invention, a double-layer GCN structure is adopted, that is, a neural network model containing two graph convolutional layers. The first layer of GCN performs preliminary convolutional operations on the input graph structure features to extract local features in the graph. The second layer of GCN further performs convolutional operations on the features output by the first layer to capture more global feature information. Through the learning of these two layers of GCN, richer and more strongly represented structured information can be obtained, and this structured information will be used in the subsequent joint domain adaptation module to achieve cross-domain fault diagnosis. In summary, the graph structure feature learning module realizes the comprehensive extraction and learning of the graph structure features in the vibration signal through the collaborative work of the adaptive similarity graph construction module based on the inner product kernel, the graph structure feature extraction module, and the feature learning module of the double-layer GCN structure.

[0068] Exemplarily, through the adaptive similarity graph construction method based on the inner product kernel, the optimal graph is generated during the iteration process, and its process can be expressed by the following equation:

[0069] S i,j = k(x i , x j )

[0070] where S i,j is the edge between node i and node j after sparsification. k(·) represents the inner product kernel, and x i and x j are the corresponding node features. Sparsity is usually a common property of natural graph structures in reality, but the obtained similarity graph is often a dense matrix containing a large number of edges with low confidence. To eliminate the interference of these noises, the present invention adopts the technique of matrix sparsification. Specifically, applying K-Nearest Neighbor (KNN) and only retaining the top K scored edges can be defined as:

[0071]

[0072] Among them, is the edge between node i and node j after sparsification.

[0073] Exemplarily, GraphSAGE forms an inductive framework through two operations: sampling and aggregation. Through the sampling operation, the full-batch training method is transformed into a node-centered mini-batch training method, thereby improving the scalability of the model. In addition, GraphSAGE also extends the method of aggregating neighborhood information, including mean, sum, etc. Applying GraphSAGE to adaptively extract graph structure features from the similarity graph can be expressed as:

[0074]

[0075] Among them, respectively represent the embedding representations of the k-th layer and the (k - 1)-th layer; σ represents the activation function; W k represents the trainable parameter of the k-th layer; is the set of node representations in the neighborhood of the central node; Aggregate(·) represents the aggregation function, and the mean aggregator is selected in this embodiment.

[0076] The improved joint domain adaptation module is used to transform structured information into domain-invariant features and perform fault diagnosis based on the domain-invariant features. Domain-invariant features refer to features that are consistent and stable in different domains. They can cross domain differences and provide reliable information for fault diagnosis. To achieve this goal, the improved joint domain adaptation module adopts an adversarial training strategy. Specifically, the module contains a feature extractor and a domain classifier. The task of the feature extractor is to extract domain-invariant features from the structured information, while the task of the domain classifier is to determine which domain these features come from. During the training process, the feature extractor and the domain classifier perform adversarial training. The feature extractor extracts features that can confuse the domain classifier, that is, making the domain classifier unable to accurately judge the domain to which the features belong. At the same time, the domain classifier continuously improves its classification ability to better identify the domain label of the features. Through this adversarial training, the feature extractor is gradually optimized to be able to extract highly domain-invariant features. After extracting the domain-invariant features, the improved joint domain adaptation module uses these features for fault diagnosis. Since the domain-invariant features are consistent and stable in different domains, they can provide reliable information for fault diagnosis and reduce the impact of domain differences on the diagnosis results.

[0077] Specifically, the improved joint domain adaptation loss function is specifically:

[0078]

[0079] In the formula, is the improved joint domain adaptation loss; is the cross-entropy loss; is the deep coral loss; is the entropy conditional adversarial domain adaptation loss; is the loss based on multi-kernel maximum mean discrepancy; a, b, c are trade-off parameters; m is the size of the mini-batch; represents the k-th true label; is the output vector of the k-th softmax layer; d is the dimension of the feature; represents the square matrix Frobenius norm; CE(·,·) represents the cross-entropy loss; and respectively represent the classification prediction probability of the i-th sample input from the source domain dataset to the model and the true label of the i-th sample input from the source domain dataset to the model; represents the classification prediction probability of the j-th sample input from the target domain dataset to the model; C s is the covariance matrix of the source domain features; C t is the covariance matrix of the target domain features; F(·) is the feature extractor; Φ(·) represents the non-linear mapping function, and H represents the distance calculated after mapping the features to the reproducing Hilbert space; is the sample collected from the source distribution D s ; is the sample collected from the target distribution D t ; D s and D t respectively represent the probability distributions of the source domain and the target domain; represents the expected value, that is, the average value of the random variable; w(·) represents the weight generated by the entropy; represents the deep feature of the j-th sample input from the target domain dataset to the model; represents the deep feature of the i-th sample input from the source domain dataset to the model; λ represents the weight parameter used to balance the two loss terms; D(.) represents the domain discrimination operation of the domain discriminator; G(.) is the classification operation of the source classifier; H(·) represents the entropy of the sample; T(·) represents the multi-linear mapping result of the sample; M(.) is the result after mapping the sample.

[0080] Specifically, in order to obtain effective domain-invariant features, based on the idea of domain adversarial, a joint domain adaptation module is proposed.

[0081] a. Introduce the cross-entropy loss to optimize the model. According to experience, the softmax layer is generally used as the last layer of the classification task, and outputs the probability vector of the input sample belonging to each type.

[0082] After obtaining the probabilities processed by the softmax layer, the cross-entropy loss is calculated to optimize the model. Suppose there are m classes of inputs, which are defined by the following expressions:

[0083]

[0084] where, represents the cross-entropy loss, m represents the size of the mini-batch, represents the k-th true label, is the output vector of the k-th softmax layer.

[0085] b. The Deep coral loss is introduced. A novel regularization constraint term, the Deep coral loss, which has a similar implementation idea to the Maximum Mean Discrepancy (MMD), is used. Suppose C s and C t represent the covariance matrices of the source domain features and the target domain features respectively, then the coral loss can be expressed as:

[0086]

[0087] where, is the Deep coral loss, d represents the dimension of the features, represents the square matrix Frobenius norm.

[0088] c. The Entropy Condition Adversarial Domain Adaptation method is introduced, aiming to use both network feature information and label information during the classification process.

[0089] To fully capture the multimodal structure, a multilinear condition is designed to calculate the cross-covariance dependence between classes and features, and the multilinear mapping can be expressed as:

[0090]

[0091] where, represents the mapping function, represents the outer product, f is the feature representation, and is the classifier prediction. However, the dimension of the mapping result is often very high, leading to parameter explosion. Suppose the dimensions of f and are represented by d f and respectively, then the dimension of the mapping result will be To solve this problem, a random algorithm is defined, and its expression is as follows:

[0092]

[0093] where, represents the randomized multilinear mapping. When the dimension Rf and denote random matrices sampled from f and When ⊙ represents the element-wise product, the set and the final conditional policy can be defined as:

[0094]

[0095] where, if the original value is used, otherwise the simplified value is adopted.

[0096] The domain adaptation method based on the adversarial policy assigns the same importance to each sample. However, the process of domain adaptation may be negatively affected by potentially difficult-to-transfer samples. Therefore, an entropy regulation strategy is applied. Specifically, by calculating the uncertainty predicted by the classifier, each sample will be assigned different weights, and the process can be defined as:

[0097]

[0098] w(S(g)) = 1 + e -S(g)

[0099] where S(·) represents the entropy criterion, N is the number of sample categories, g c represents the probability that the sample belongs to c categories, and w(·) represents the weight generated by the entropy.

[0100] According to the optimization idea of the classical adversarial network, the objective optimization function of the source domain classifier G is defined as follows:

[0101]

[0102] where CE(·,·) represents the cross-entropy loss; respectively represent the classification prediction probability of the i-th sample input from the source domain dataset to the model and the true label of the i-th sample input from the source domain dataset to the model. The domain D and the source classifier

[0103] The objective function of G can be defined as:

[0104]

[0105] where, is the deep feature of the i-th sample input from the source domain dataset to the model, is the deep feature of the j-th sample input from the target domain dataset to the model, The classification prediction probability of the j-th sample input from the target domain dataset to the model. Based on the above objective function, the min-max game can be described as:

[0106]

[0107] Among them, λ is a trade-off parameter.

[0108] Combining the multi-linear adjustment and entropy adjustment strategies, the final optimization objective of CDAN can be defined as:

[0109]

[0110] d. A metric based on multi-kernel maximum mean discrepancy (MK-MMD) is introduced to measure the probability distribution difference, which can be defined as:

[0111]

[0112] Among them, F(·) is the feature extractor, Φ(·) represents the non-linear mapping function, and H represents the distance calculated after mapping the features to the reproducing kernel Hilbert space (RKHS). is the sample collected from the source distribution D s in, is the sample collected from the target distribution D t in. D s and D t respectively represent the probability distributions of the source domain and the target domain, and E represents the expected value, that is, the average value of the random variable.

[0113] e. The overall improved joint domain adaptation loss function is as follows:

[0114]

[0115] Among them, a, b, and c are three trade-off parameters set to obtain better joint effects. After a large number of test experiments, β and γ are selected as 1, and α is set as the kinetic parameter, and its expression can be written as:

[0116]

[0117] Among them, θ ∈ [0,1] is a factor that changes with the training process.

[0118] In one embodiment, the improved joint domain adaptation module is sequentially stacked by a global average pooling layer, three fully connected layers, and a SoftMax layer.

[0119] Exemplarily, such as Figure 2 and Figure 3As shown in the figure, the specific structure and method flow of the adversarial-based joint domain adaptive graph convolutional network provided by the present invention are as follows. First, the original data of the fault simulation platform is collected by vibration sensors, and the labeled samples in the original signal are divided into the source domain for training, and the unlabeled samples are used as the target domain for testing. Second, the constructed AJDAGCN model is trained using the training data. Finally, the test data is input into the trained AJDAGCN model to obtain the diagnostic result. During the diagnosis process, four methods are selected for comparison, including a DL-based method, two traditional transfer methods, and a deep transfer method. The proposed method can achieve state-of-the-art performance.

[0120] Next, the XJTU bearing dataset is used as an example to illustrate the present invention. The XJTU bearing dataset is collected from a fan drive system diagnostic simulator, and its composition is as Figure 7 shown. Due to the modular design, the test bench can be adjusted according to the need to simulate various fault conditions. The type of the experimental bearing is ER16K. The vibration data is collected from the vertical direction at a sampling frequency of 20480Hz. Four states are processed in the experiment, namely internal fault (IF), external fault (OF), ball fault (BF), and normal state (NS). Three speed conditions can be set by controlling the motor, including 1000r / min (R 0 ), 1500r / min (R 1 ), and 2000r / min (R 2 ). Each sample contains 4096 sampling points, and the number of samples in each state is 750. The entire dataset is randomly divided into a training set and a test set at a ratio of 7:3. The source domain and the target domain are divided for the obtained dataset, and a total of 6 transfer learning tasks are set, namely: R 0 →R 1 , R 0 →R 2 , R 1 →R 0 , R 1 →R 2 , R 2 →R 0 and R 2 →R 1 . The signals of each group of source domain and target domain samples are processed, and the original data is normalized by the mean-standard deviation normalization method. The constructed AJDAGCN model is trained using the training data. All methods use the validation set to search for the optimal weights of the model, and the number of iterations for all datasets is set to 150. In addition, each method is repeated ten times to eliminate the influence of accidental factors.

[0121] When training the AJDAGCN model with the training data, by optimizing the training objective Ljoint Through continuous optimization, a well-trained high-performance model for fault diagnosis under PWC is finally generated, achieving the effect of cross-domain diagnosis.

[0122] To prove the superiority of the present invention, four methods were selected for comparison, including one DL-based method, two traditional transfer methods, and one deep transfer method. These 4 methods are as follows: ① Residual Neural Network (ResNet18). As one of the most successful backbones in the field of DL, ResNet has strong learning ability and good generalization performance. In this embodiment, ResNet18, which is widely popular in the field of fault diagnosis, is used as a comparison method. ② Deep CORrelation Alignment (D-CORAL). Deep CORAL extends the CORAL algorithm and applies more powerful non-linear transformations to deep networks. ③ Deep Domain Confusion (DCC). DDC applies an adaptation layer to a deep neural network and uses MMD as the domain confusion loss, belonging to traditional transfer methods. ④ Deep Adversarial Neural Network (DANN). DANN is an advanced domain adaptation method with the idea of adversarial transfer and belongs to deep transfer methods. The comparison results of the five models are as follows:

[0123] 1) The classification accuracy and experimental results are as Figure 8 shown in Table 1. Obviously, the present invention has achieved the best results in all transfer tasks. In addition, the gap between the comparison methods and the method proposed by the present invention is also more obvious. It can be seen from these results that the transfer task R 2 →R 0 is the most difficult. None of the comparison methods have achieved good classification results in this task. The highest accuracy of DDC is 85.92%. However, in this task, the present invention still achieved a relatively high classification accuracy, which is about 13.55% higher than the best result of the comparison methods. This proves the superiority of the present invention.

[0124] Table 1 Classification results of XJTU dataset (%)

[0125]

[0126] 2) Randomly select the transfer task R 0→ R 2 , Figure 9 and Fig. (a)-(d) show the features of all methods learned by the t-SNE method. It can be seen from Figure 9 (a)-(d) that there are varying degrees of confusion among the four types of bearing data. Therefore, the subsequent classifier based on these features cannot achieve good classification accuracy. These analysis results of the XJTU dataset further prove that the present invention has superior ability in learning effective and highly robust features.

[0127] 3) Based on the same task, the detailed classification results of different methods are shown through a confusion matrix, as Figure 10 shown. As can be seen from Figure 10 (a)-(d), ResNet 18, D-CORAL, DANN, and DDC did not achieve good classification results for some types of bearing data. Especially for OF data, except for the present invention, the effects of other comparison methods are not very good. Obviously, the method proposed by the present invention has the highest classification accuracy, which further proves the superiority of the present invention.

[0128] An embodiment of the present invention provides a cross-domain mechanical equipment fault diagnosis device based on an adversarial joint domain adaptive graph convolutional network, including:

[0129] An acquisition module for acquiring the vibration signal of the mechanical equipment to be diagnosed.

[0130] A diagnosis module for inputting the vibration signal of the mechanical equipment to be diagnosed into a pre-trained mechanical equipment fault diagnosis model and outputting a diagnosis result; wherein, the mechanical equipment fault diagnosis model is obtained by training an adversarial joint domain adaptive graph convolutional network using training data, the training data includes source domain data and target domain data, the source domain data includes non-fault vibration signals and different fault vibration signals with different fault labels, the target domain data includes non-fault vibration signals and different fault vibration signals without fault labels; the adversarial joint domain adaptive graph convolutional network includes a multi-scale feature extraction module with a hybrid attention mechanism, a graph structure feature learning module, and an improved joint domain adaptive module connected in sequence, the multi-scale feature extraction module with a hybrid attention mechanism includes a multi-scale convolutional layer with three different convolutional kernel sizes, a hybrid attention mechanism module, three sequentially connected convolutional modules, and a fully connected layer; the graph structure feature learning module includes a graph construction module and a feature learning module with a double-layer GCN structure, the graph construction module includes an adaptive similarity graph construction module based on an inner product kernel and a graph structure feature extraction module; the loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

[0131] All relevant content of each step involved in the embodiment of the foregoing cross - domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network can be cited in the function description of the corresponding functional modules of a cross - domain mechanical equipment fault diagnosis device based on adversarial joint domain adaptive graph convolutional network in the embodiments of the present invention, and will not be elaborated here. The division of modules in the embodiments of the present invention is illustrative, merely a logical function division. In actual implementation, there may be other division methods. In addition, in each embodiment of the present invention, the functional modules can be integrated in one processor, or exist separately physically, or two or more modules can be integrated in one module. The above - mentioned integrated modules can be implemented in the form of hardware or in the form of software functional modules.

[0132] In another embodiment of the present invention, a computer device is provided. The computer device includes a processor and a memory. The memory is used to store a computer program. The computer program includes program instructions. The processor is used to execute the program instructions stored in the computer storage medium. The processor may be a central processing unit (CPU), or may also be other general - purpose processors, digital signal processors (DSPs), application - specific integrated circuits (ASICs), field - programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing core and control core of the terminal, and is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the computer storage medium to implement the corresponding method process or corresponding function. The processor described in the embodiments of the present invention can be used for the operation of a cross - domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network.

[0133] In another embodiment of the present invention, the present invention further provides a storage medium, specifically a computer-readable storage medium (Memory). The computer-readable storage medium is a memory device in a computer device and is used to store programs and data. It can be understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and, of course, the extended storage medium supported by the computer device. The computer-readable storage medium provides a storage space, and this storage space stores the operating system of the terminal. And, in this storage space, there are also stored one or more instructions suitable for being loaded and executed by the processor. These instructions can be one or more computer programs (including program codes). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk memory. One or more instructions stored in the computer-readable storage medium can be loaded and executed by the processor to implement the corresponding steps of the cross-domain mechanical equipment fault diagnosis method related to an adversarial-based joint domain adaptation graph convolutional network in the above embodiment.

[0134] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0135] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram, as well as the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0136] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer-readable memory generate a manufactured article including an instruction device, and this instruction device implements the functions in the flow Figure 1one or more processes and / or blocks Figure 1 the functions specified in one or more blocks.

[0137] These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process. Thus, the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one or more processes and / or blocks Figure 1 the functions specified in one or more blocks.

[0138] The present invention also provides a computer program product, which is used to execute any one of the above-mentioned cross-domain mechanical equipment fault diagnosis methods based on adversarial joint domain adaptive graph convolutional network. Since the computer program product provided by the present invention and the above-mentioned cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network belong to the same inventive concept, the computer program product provided by the present invention has all the advantages of the above-mentioned cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network. Therefore, the beneficial effects of the computer program product provided by the present invention will not be elaborated one by one here.

[0139] In the present invention, terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. mean that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0140] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting it. The protection scope of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the technical field of the present invention can still modify the technical solutions recorded in the foregoing embodiments, or can easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network, characterized in that: include: Obtain vibration signals of mechanical equipment to be diagnosed; The vibration signal of the mechanical equipment to be diagnosed is input into a pre-trained mechanical equipment fault diagnosis model, and a diagnosis result is output; wherein the mechanical equipment fault diagnosis model is obtained by training an adversarial joint domain adaptive graph convolutional network using training data, the training data includes source domain data and target domain data, the source domain data includes non-fault vibration signals and different fault vibration signals with different fault labels, and the target domain data includes non-fault vibration signals and different fault vibration signals without fault labels; the adversarial joint domain adaptive graph convolutional network includes a multi-scale feature extraction module with a hybrid attention mechanism, a graph structure feature learning module and an improved joint domain adaptive module connected in sequence, the multi-scale feature extraction module with a hybrid attention mechanism includes a multi-scale convolution layer with three different convolution kernel sizes connected in sequence, a hybrid attention mechanism module, three sequentially connected convolution modules and a fully connected layer; the graph structure feature learning module includes a graph construction module and a feature learning module with a double-layer GCN structure, the graph construction module includes an adaptive similarity graph construction module based on an inner product kernel and a graph structure feature extraction module; the loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

2. According to claim 1, a cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network is characterized in that: Regarding the multi-scale feature extraction module with hybrid attention mechanism, specifically: The multi-scale convolution layer with three different convolution kernel sizes is used to extract different scale features in the vibration signal; The hybrid attention mechanism module is used to assign weights to features of different scales; The three sequentially connected convolution modules are used to deeply extract the different scale features after the weights are assigned, so as to obtain the different scale features after the deep extraction; the fully connected layer is used to transform the different scale features after the deep extraction, so as to obtain the transformed different scale features.

3. According to claim 2, a cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network is characterized in that: The graph structure feature learning module is as follows: The inner product kernel-based adaptive similarity graph construction module is used to construct a graph for the converted different scale features using an inner product kernel-based adaptive similarity graph construction method to obtain an optimal graph; The graph structure feature extraction module is used to adaptively extract graph structure features from the optimal graph using the GraphSAGE method; The feature learning module of the double-layer GCN structure is used to input the extracted graph structure features into the double-layer GCN to obtain structured information.

4. According to claim 3, a cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network is characterized in that: Regarding the improved joint domain adaptation module, specifically: The improved joint domain adaptation module is used to convert structured information into domain-invariant features, and perform fault diagnosis based on the domain-invariant features.

5. According to claim 1, a cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network is characterized in that: The improved joint domain adaptation loss function is specifically: In the formula, for the improved joint domain adaptation loss; is the cross entropy loss; for deep coral loss; Entropy-conditioned adversarial domain adaptation loss; is the multi-core maximum mean difference loss; a, b, c are trade-off parameters; m is the size of the mini-batch; represents the kth true label; is the output vector of the kth softmax layer; d is the dimension of the feature; represents the Frobenius norm of the square matrix; CE(·,·) represents the cross entropy loss; and They represent the classification prediction probability of the i-th sample input to the model from the source domain dataset and the true label of the i-th sample input to the model from the source domain dataset; represents the classification prediction probability of the jth sample input to the model from the target domain dataset; C s is the covariance matrix of the source domain features; C t is the covariance matrix of the target domain features; F(·) is the feature extractor; Φ(·) represents the nonlinear mapping function, and H represents the distance calculated after mapping the features to the regenerated Hilbert space; is distributed from source D s Samples collected in is from the target distribution D t Samples collected in D s and D t Represent the probability distribution of the source domain and the target domain respectively; represents the expected value, that is, the average value of the random variable; θ∈[0,1] is a factor that changes with the training process; w(·) represents the weight generated by entropy; Represents the deep features of the jth sample input to the model from the target domain dataset; Represents the deep features of the i-th sample input to the model from the source domain dataset; λ represents the weight parameter used to balance the two loss terms; D(.) represents the domain discrimination operation of the domain discriminator; G(.) is the classification operation of the source classifier; H(·) represents the entropy of the sample; T(·) represents the multilinear mapping result of the sample; M(.) is the result after the mapping operation on the sample.

6. According to claim 1, a cross-domain mechanical equipment fault diagnosis method based on adversarial joint domain adaptive graph convolutional network is characterized in that: The improved joint domain adaptation module is composed of a global average pooling layer, three fully connected layers and a SoftMax layer stacked in sequence.

7. A cross-domain mechanical equipment fault diagnosis device based on adversarial joint domain adaptive graph convolutional network, characterized in that: include: An acquisition module, used for acquiring a vibration signal of the mechanical equipment to be diagnosed; A diagnosis module is used to input the vibration signal of the mechanical equipment to be diagnosed into a pre-trained mechanical equipment fault diagnosis model, and output a diagnosis result; wherein the mechanical equipment fault diagnosis model is obtained by training an adversarial joint domain adaptive graph convolutional network using training data, the training data includes source domain data and target domain data, the source domain data includes non-fault vibration signals and different fault vibration signals with different fault labels, and the target domain data includes non-fault vibration signals and different fault vibration signals without fault labels; the adversarial joint domain adaptive graph convolutional network includes a multi-scale feature extraction module with a hybrid attention mechanism, a graph structure feature learning module and an improved joint domain adaptive module connected in sequence, the multi-scale feature extraction module with a hybrid attention mechanism includes a multi-scale convolution layer with three different convolution kernel sizes connected in sequence, a hybrid attention mechanism module, three sequentially connected convolution modules and a fully connected layer; the graph structure feature learning module includes a graph construction module and a feature learning module of a double-layer GCN structure, the graph construction module includes an adaptive similarity graph construction module based on an inner product kernel and a graph structure feature extraction module; the loss function adopted by the improved joint domain adaptive module is an improved joint domain adaptive loss function.

8. A device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements a cross-domain mechanical equipment fault diagnosis method based on an adversarial joint domain adaptive graph convolutional network as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements a cross-domain mechanical equipment fault diagnosis method based on an adversarial joint domain adaptive graph convolutional network as described in any one of claims 1 to 6.

10. A computer program product, characterized in that When the computer program product is executed by a processor, it implements a cross-domain mechanical equipment fault diagnosis method based on an adversarial joint domain adaptive graph convolutional network as described in any one of claims 1 to 6.

Images