Variable working condition fault diagnosis method and device based on adaptive anchor point graph attention

Abstract

This invention discloses a method and apparatus for fault diagnosis under varying operating conditions based on adaptive anchor point graph attention, belonging to the fields of mechanical fault diagnosis and graph neural network technology. This invention performs adaptive asymmetric truncated smoothing on vibration signals under different operating conditions to remove extreme noise from varying operating conditions, and then adaptively generates cross-domain anchor points based on feature space distribution. A highly reliable topology graph is reconstructed based on the denoised features, and deep fusion and implicit alignment of features are achieved through global cross-domain path (GCP) and local topology path (LTA), improving the model's ability to learn and process sparse edge samples. Finally, efficient extraction and alignment of deep domain invariant features are achieved through unsupervised adaptive joint optimization combined with multi-kernel maximum mean difference (MMD). This invention effectively overcomes the oversmoothing defect of traditional graph neural networks, achieves deep fusion of global and local features, and significantly improves the model's ability to represent sparse and isolated samples.

Description

Technical Field

[0001] This invention relates to the fields of mechanical fault diagnosis and graph neural network technology, specifically to an unsupervised adaptive fault diagnosis method for rotating machinery under varying operating conditions based on an adaptive anchor point graph attention network, which is applicable to fault identification and diagnosis of mechanical equipment under scenarios of changing operating conditions and data distribution shifts. Background Technology

[0002] In industrial production, the stable operation of core mechanical equipment such as bearings, fans, and pumps is crucial for ensuring production continuity, and timely and accurate fault diagnosis of these components has significant engineering value. With the development of deep learning technology, intelligent fault diagnosis methods based on neural networks have become mainstream. However, in real-world industrial scenarios, factors such as varying equipment operating loads, wear and aging, and environmental noise can lead to significant distribution differences between the source and target domain data for fault diagnosis—a problem known as covariate shift under varying operating conditions. Furthermore, fault labeling data in the target domain is often scarce in real-world industrial settings. Therefore, Unsupervised Domain Adaptation (UDA) has become a core technological direction for addressing this problem.

[0003] Currently, fault diagnosis methods based on unsupervised domain adaptation mainly reduce the distribution difference between the source and target domains through feature alignment and adversarial training. However, existing methods have the following key drawbacks: First, most methods only focus on modeling category labels and domain labels, ignoring the correlation and distribution characteristics between samples. This makes it difficult to capture the global structural information of equipment fault data, and it is prone to the problem of missing contextual information in sparse data scenarios.

[0004] Second, when processing fault data, traditional graph attention networks (GAT) have a receptive field limited to the first-order neighbors of nodes. Stacking multiple layers of GAT can lead to oversmoothing problems, resulting in a decrease in the discriminability of fault features and an inability to effectively extract domain-invariant and domain-separable features.

[0005] Third, existing anchor-based graph neural network methods use a fixed number of anchors, which cannot adaptively adjust according to the inherent characteristics of fault data such as sparsity and scale. In sparse fault data under varying operating conditions, they are prone to insufficient information coverage or waste of computational resources, making it difficult to balance the effectiveness of feature extraction and computational efficiency.

[0006] Fourth, fault data under varying operating conditions suffers from the dual problems of missing context in sparse regions and overly smoothed features in dense regions. Existing methods cannot solve both problems simultaneously, resulting in limited generalization performance and accuracy of fault diagnosis.

[0007] Furthermore, traditional methods have a relatively simple way of aggregating sample features and do not consider the centrality and spatial location information of nodes in the graph, which further reduces the ability of graph attention networks to represent fault features and makes it difficult to meet the engineering requirements of high-precision fault diagnosis under varying working conditions. Summary of the Invention

[0008] This invention addresses the problems of existing variable operating condition fault diagnosis methods in unsupervised adaptive scenarios, such as insufficient capture of global structural information of samples, overly smooth graph attention networks, inability of anchor points to adaptively match data features, and imbalance in fault feature representation in sparse or dense regions. It proposes a variable operating condition fault diagnosis method and device based on adaptive anchor point graph attention.

[0009] In a first aspect, this invention proposes a variable-condition fault diagnosis method based on adaptive anchor graph attention, comprising: Step S1: Collect vibration signals of rotating machinery under source domain operating conditions and target domain variable operating conditions respectively, and generate a preliminary node feature matrix after preprocessing; Step S2: Perform adaptive asymmetric truncation smoothing on the preliminary node feature matrix to generate denoised smooth features; Step S3: Based on the variance distribution, feature dimension, and number of samples in the batch data of the smoothed feature, adaptively calculate the number of cross-domain anchor points and randomly initialize and generate the cross-domain anchor point feature matrix; Step S4: Calculate the cosine similarity between nodes in the sample using the smoothed features after noise reduction, construct the initial sparse adjacency matrix using the Top-k function; calculate the shortest path matrix between any nodes using the breadth-first search algorithm, and perform binarization reconstruction mapping on the graph structure to generate the reconstructed topology graph; Step S5: Simultaneously input the reconstructed topology graph and the cross-domain anchor feature matrix into the dual-path unified graph attention network; extract global cross-domain features and local topology features through the global cross-domain path GCP and the local topology path LTA respectively, and perform deep fusion through residual connections and layer normalization to output high-order node representations; Step S6: The high-order node representations are jointly optimized using a three-layer unsupervised domain adaptive network architecture to complete fault classification under varying operating conditions.

[0010] Furthermore, in step S2, the step of performing adaptive asymmetric truncated smoothing on the preliminary node feature matrix to generate denoised smooth features specifically includes: Based on the median, maximum, and minimum values ​​of each feature dimension of the preliminary node feature matrix in the current batch of data, calculate the positive and negative offset distances that measure the offset direction and severity of the impact noise under varying operating conditions. Based on the relative ratio of positive offset distance to negative offset distance, the total truncation ratio threshold is adaptively decomposed into an upper truncation rate and a lower truncation rate. For each feature dimension, based on the upper and lower truncation rates, extreme feature values ​​at the top and bottom of the ranking are removed, the robust truncation mean of the remaining features is calculated, and the features are normalized using the L2 norm to generate denoised smooth features.

[0011] Furthermore, in step S3, the step of adaptively calculating the number of cross-domain anchor points based on the variance distribution of the smoothed features, the feature dimension, and the number of samples in the batch data, and randomly initializing and generating the cross-domain anchor point feature matrix specifically includes: The number of cross-domain anchor points is calculated using the following formula. : ; in, This indicates the number of samples in the batch data. This represents the initial feature dimension extracted from the node. The overall variance representing the feature. Indicates hyperparameters; The generated cross-domain anchor feature matrix A is represented as follows: ; in, Indicates the anchor point. This represents the initial feature dimension extracted from the node. Represents the set of real numbers. This indicates the smoothing feature after noise reduction.

[0012] Furthermore, in step S4, the constructed initial sparse adjacency matrix B is represented as: ; in, This represents the smoothed features after noise reduction. The Top-k function is used to extract the top k most relevant edges for each node. The breadth-first search (BFS) algorithm is used to calculate the shortest path matrix D between any nodes, which is represented as: D = BFS(B); The elements in D Represents i nodes To the j-th node The shortest path length; Perform a binarized reconstruction mapping on the graph structure to generate a reconstructed topological graph mapping matrix. ; in, ; in, Represents the mapping matrix for reconstructing the topology graph. The elements in.

[0013] Furthermore, in step S5, the step of extracting global cross-domain features through the global cross-domain path GCP specifically includes: All nodes in the source and target domains All with anchor points Perform fully connected attentional interactions and achieve a global interaction attention score. Represented as: ; The global interaction attention score is converted into a normalized global attention weight using the softmax function. , represented as: ; Through global attention weights The global cross-domain features are obtained by calculating the transformation features of all anchor points. , represented as: ; in, This indicates a concatenation operation; m takes the value of LeakyReLU is the activation function; For learnable weight matrix, For learnable attention vectors, This represents the dimension of the latent space features after mapping by the weight matrix. This represents the initial feature dimension extracted from the node. This represents a nonlinear activation function used to introduce nonlinear expressive power into the weighted summation of global cross-domain features.

[0014] Furthermore, in step S5, the step of extracting local topological features through the local topological path LTA specifically includes: The nodes in the reconstructed topology graph The in-degree and out-degree are specified as learnable centrality embedding vectors. and , obtain node Enhanced features : ; Learnable scalar biases assigned based on spatial location information in a predefined reconstructed topology graph. Enhanced features and inverse decay function Obtain the local attention score that includes centrality and spatial awareness. , represented as: ; The local attention scores are normalized using the softmax function to calculate the local attention weights. And expressed as: ; Based on the local attention weights and enhancement features Calculate local topological features , represented as: ; in, This indicates a splicing operation. This indicates the relationship between nodes in the reconstructed topology graph. The set of neighboring nodes that are connected; Learnable weight matrix For learnable attention vectors, This represents the dimension of the latent space features after mapping by the weight matrix. This represents the initial feature dimension extracted from the node. This represents a nonlinear activation function used to introduce nonlinear expressive power into the weighted summation of global cross-domain features.

[0015] Furthermore, in step S5, the step of deep fusion through residual connections and layer normalization to output high-order node representations involves high-order node representations. Represented as: ; in, Indicates to The output results of each attention head are concatenated. It is a layer normalization operation. This represents the global cross-domain feature output of the h-th attention head. This represents the local topological feature output of the h-th attention head.

[0016] Furthermore, the three-layer unsupervised domain adaptive network architecture includes a feature extractor, a domain discriminator, and a fault classifier. Through weighted joint optimization of classification loss, domain alignment adversarial loss, and multi-kernel maximum mean difference loss, it achieves efficient extraction of domain-invariant fault features and accurate fault diagnosis of unlabeled target domains, ultimately completing the fault classification task under varying operating conditions.

[0017] Secondly, the present invention also proposes a variable operating condition fault diagnosis device based on adaptive anchor graph attention, applied to the variable operating condition fault diagnosis method as described in the first aspect, specifically including: The signal acquisition and feature extraction module is used to acquire vibration signals of rotating machinery under source domain operating conditions and target domain variable operating conditions, and generate a preliminary node feature matrix after preprocessing. An adaptive asymmetric truncation smoothing module is used to perform adaptive asymmetric truncation smoothing on the preliminary node feature matrix to generate smoothed features after noise reduction. An adaptive anchor point generation module is used to adaptively calculate the number of cross-domain anchor points based on the variance distribution of the smoothed features, the feature dimension, and the number of samples in the batch data, and to randomly initialize and generate a cross-domain anchor point feature matrix. The topology graph construction and reconstruction module is used to calculate the cosine similarity between nodes in the sample using the smoothed features after noise reduction, construct the initial sparse adjacency matrix through the Top-k function, calculate the shortest path matrix between any nodes using the breadth-first search algorithm, and perform binarization reconstruction mapping on the graph structure to generate the reconstructed topology graph. The dual-path unified graph attention module is used to synchronously input the reconstructed topology graph and the cross-domain anchor feature matrix into the dual-path unified graph attention network; global cross-domain features and local topology features are extracted by global cross-domain path GCP and local topology path LTA respectively, and deep fusion is performed by residual connection and layer normalization to output high-order node representations. The unsupervised domain adaptive joint optimization module is used to jointly optimize the high-order node representation through a three-layer unsupervised domain adaptive network architecture to complete fault classification under varying operating conditions.

[0018] Thirdly, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0019] The adaptive fault diagnosis method for unsupervised domain under varying operating conditions proposed in this invention, based on an adaptive anchor graph attention network, has several significant advantages over existing technologies: An adaptive asymmetric truncation smoothing mechanism based on the extreme value offset ratio was designed.

[0020] This invention utilizes the relative distance between the median and extreme values ​​in the feature space to adaptively sense the bias direction of impact energy. When a sudden change in operating conditions introduces a strong unidirectional impact, the model dynamically tilts the truncation index towards the high-energy noise side, achieving targeted removal of extreme outliers. This maximizes the preservation of true background fault characteristics and improves the model's robustness under harsh operating conditions.

[0021] Second, we designed an adaptive generation strategy for cross-domain anchor points based on feature space, which breaks through the limitations of the traditional "build the graph first and then generate anchor points". It does not rely on graph topology information, but can dynamically generate cross-domain anchor points before graph construction based only on the variance distribution, dimension and sample number of batch features. This avoids the waste of computing resources and uses anchor points as cross-domain bridges to allow source domain and target domain nodes to interact with them, realizing implicit feature distribution alignment in advance and completely breaking the traditional physical isolation between domains.

[0022] Third, the constructed dual-path unified graph attention network, through the parallel architecture of "global cross-domain path + local topology path", allows nodes to not only mine the fine intra-class structure under the same working conditions, but also absorb the macro-context of cross-domain hubs. It can obtain the global receptive field without stacking network layers, effectively overcoming the oversmoothing defect of traditional graph neural networks, and achieving deep fusion of global and local features. While preserving the intra-class compactness and inter-class separability of fault features, it greatly improves the model's ability to represent sparse and isolated samples. Attached Figure Description

[0023] Figure 1 This is a flowchart of the variable operating condition fault diagnosis method proposed in this invention; Figure 2 This is a diagram of the dual-path unified graph attention network architecture in the variable operating condition fault diagnosis method proposed in this invention. Figure 3 This is a schematic diagram of the HP-to-0HP confusion matrix visualization results in one embodiment of the variable operating condition fault diagnosis method proposed in this invention; Figure 4 This is a schematic diagram of the visualization results of the 2HP-to-3HP confusion matrix in one embodiment of the variable operating condition fault diagnosis method proposed in this invention; Figure 5 This is a schematic diagram of the HP-to-0HP_t-SNE feature distribution visualization results in one embodiment of the variable operating condition fault diagnosis method proposed in this invention; Figure 6 This is a schematic diagram of the visualization results of the 2HP-to-3HP_t-SNE feature distribution in one embodiment of the variable operating condition fault diagnosis method proposed in this invention; Figure 7 This is a framework diagram of the variable operating condition fault diagnosis device based on adaptive anchor graph attention proposed in this invention. Detailed Implementation

[0024] The present application will now be described in further detail with reference to the accompanying drawings, but the scope of protection of the present application is not limited to the specific embodiments.

[0025] To address the challenges of fault diagnosis in real-world industrial scenarios, such as inconsistent data distribution due to varying operating conditions and a lack of labeled data in the target domain, an unsupervised adaptive method based on an adaptive anchor point graph attention network is employed. Graph neural networks, as a powerful deep learning method for data processing, are based on the aggregation principle of node features and topological structure. This makes them more efficient in capturing complex relationships and distribution characteristics between samples, and the results more readily reflect the inherent manifold structure of the data. This invention innovatively introduces a distribution-based adaptive anchor point generation mechanism, which not only effectively solves the oversmoothing problem that easily occurs in deep layers of traditional graph networks but also overcomes local graph construction errors caused by noise interference. It is an effective method for solving cross-domain fault diagnosis problems under varying operating conditions.

[0026] like Figure 1 As shown, this embodiment proposes a fault diagnosis method for variable operating conditions based on adaptive anchor graph attention, which specifically includes the following steps: Step S1: Collect vibration signals of rotating machinery under source domain operating conditions and target domain variable operating conditions respectively, and generate a preliminary node feature matrix after preprocessing.

[0027] In practical applications with varying operating conditions, it is necessary to acquire the vibration signal set from the source domain. The target domain's variable-condition vibration signal set Xt is used to perform layered sampling of all signals from both vibration signal sets through a sliding window, forming a sequence containing... Batch data of individual samples Batch data is input into a wide-kernel convolutional neural network. Preliminary one-dimensional spatial feature extraction is performed to generate a preliminary node feature matrix X, which is represented as: ; in, This is the initial node feature matrix. This indicates the initial feature dimension extracted from the node.

[0028] Step S2: Perform adaptive asymmetric truncation smoothing on the preliminary node feature matrix to generate denoised smooth features.

[0029] In this embodiment, to prevent severe covariate shifts caused by varying operating conditions from resulting in cross-domain breaks or noisy edges in subsequent graph construction, adaptive asymmetric truncation smoothing based on the extreme value shift ratio is applied to the initial features. For each feature dimension of the initial node feature matrix, the median of the current batch of features is calculated. Maximum value and minimum value Calculate the positive offset distance. Negative offset distance To measure the direction and severity of impact noise under varying operating conditions: ; ; The total truncation ratio threshold is set based on the relative proportion of the offset distance. Adaptive decomposition with upper limit truncation rate With lower limit truncation rate : ; ; Finally, based on the calculated asymmetric truncation rate across each feature dimension, items ranked at the top are removed. Proportion and bottom Calculate the robust truncated mean of the remaining features from the extreme eigenvalues ​​of the proportion. And perform feature normalization to obtain smooth features after noise reduction. : ; in, It is the L2 norm. To prevent extremely small constants from being divided by zero, the interference of extreme outliers on the characteristic distribution can be mitigated.

[0030] Step S3: Based on the variance distribution, feature dimension, and number of samples in the batch data of the smoothed feature, adaptively calculate the number of cross-domain anchor points and randomly initialize and generate the cross-domain anchor point feature matrix.

[0031] In this embodiment, the number of cross-domain anchor points is adaptively calculated based on the number of samples, feature dimensions, and the overall variance of the features. : ; in, The number of samples in the batch data is the maximum number of samples. The larger the number, the more anchor points are needed to cover the global information of all samples and avoid insufficient information coverage. The overall variance of the features is represented. The larger the variance, the more discrete the data distribution and the stronger the sparsity. More anchor points are needed to supplement the contextual information of sparse regions and solve the problem of missing context in sparse regions of fault data under varying operating conditions. This is a hyperparameter.

[0032] After determining the number, the adaptive anchor point feature matrix is ​​randomly initialized. , Representing the set of real numbers, the generated adaptive anchor feature matrix It will serve as a reference input for the Global Cross-Domain Path (GCP) in the subsequent dual-path unified graph attention network. This represents the initial feature dimension extracted from the node. The higher the feature dimension, the richer the detailed features of the fault, requiring more anchor points to capture the global correlation in the high-dimensional features and avoid insufficient feature representation. By introducing a set of learnable global anchor points, the physical isolation between the sample nodes of the source domain and the target domain is broken, providing a unified reference benchmark for subsequent global cross-domain feature alignment.

[0033] Step S4: Calculate the cosine similarity between nodes in the sample using the smoothed features after noise reduction, construct the initial sparse adjacency matrix using the Top-k function; calculate the shortest path matrix between any nodes using the breadth-first search algorithm, and perform binarization reconstruction mapping on the graph structure to generate the reconstructed topology graph.

[0034] In this embodiment, it should be noted that in the representation learning of complex data, the quality of the graph structure directly determines the effect of subsequent information aggregation. The feature-based topology graph generation and reconstruction module aims to break the limitations of the original isolated data and adaptively establish and optimize the topological connection network between samples through feature similarity measurement in the latent space.

[0035] In this embodiment, the smoothing feature after noise reduction is utilized. Calculate the cosine similarity matrix between nodes And use the Top-k function to extract the most relevant nodes for each node. Edges, construct the initial sparse adjacency matrix : ; By traversing the initial sparse graph using the breadth-first search (BFS) algorithm, the shortest path matrix between any two nodes in the network can be systematically calculated. Matrix elements Represents the i-th node To the j-th node The shortest path length records the minimum number of edges a signal needs to traverse to propagate from nodes. For node pairs that are completely disconnected in the current local topology and have no path reachable, the system uses a special marker value of -1 for placeholder isolation. This pathfinding process elevates the simple local connectivity state into a topological distance metric with a global perspective.

[0036]

[0037] After obtaining the global connectivity information, the final step is to perform a binarized reconstruction mapping of the graph structure, generating the final reconstructed single-layer graph structure mapping matrix. Among them, the mapping matrix for reconstructing a single-layer graph structure. The elements used Specifically, it means: ; This mapping mechanism considers nodes as logically related as long as a reachable path exists between them, regardless of the path length. This reconstruction method significantly enhances the robustness of the graph, fills in edge gaps that may be caused by feature noise, and ultimately forms a structurally complete and logically rigorous topological network.

[0038] Step S5: Simultaneously input the reconstructed topology graph and cross-domain anchor feature matrix into the dual-path unified graph attention network; extract global cross-domain features and local topology features through the global cross-domain path GCP and the local topology path LTA respectively, and perform deep fusion through residual connections and layer normalization to output high-order node representations.

[0039] like Figure 2 As shown, the dual-path unified graph attention network in this embodiment adopts a dual-branch parallel feature extraction architecture, including a global cross-domain path (GCP) and a local topology path (LTA). In the GCP channel, adaptive anchor features are introduced to interact fully with node features to model long-distance global dependencies across nodes. In the LTA channel, the reconstructed topology graph is input, and a spatial relationship mapping and inverse decay attention mechanism are constructed based on the shortest path matrix to capture multi-order local topological information. The dual-branch output features are deeply fused through residual connections and layer normalization to achieve a collaborative representation of global distributed semantics and local manifold structure, enhancing the model's feature discriminativeness under varying conditions.

[0040] In this embodiment, the steps for extracting global cross-domain features through the Global Cross-Domain Path (GCP) specifically include: Force all nodes in the source and target domains All with anchor points Perform fully connected attention-based interactions to achieve implicit cross-domain alignment in advance. The global interaction attention score is calculated as follows: ; in, This represents the concatenation operation; LeakyReLU is the activation function. For learnable weight matrix, For learnable attention vectors; in sparse graphs, the expected value of attention score follows a heavy-tailed distribution and contains a large number of negative values, which can maintain gradient flow even under negative attention scores, ensuring the smoothness and traversability of the optimization space.

[0041] The attention scores are converted into normalized weights using the softmax function, ensuring that the sum of the weights of each node to all anchors is 1. This normalization process yields the global attention weights. : ; The global context features of a node are obtained by weighting and summing the transformation features of all anchor points using global attention weights. These global context features are then fused with the original node features to calculate the final global backflow features. The formula is as follows: ; In the above formula, This indicates a concatenation operation; m takes the value of LeakyReLU is the activation function. For learnable weight matrix, For learnable attention vectors, This represents the dimension of the latent space features after mapping by the weight matrix. This represents the initial feature dimension extracted from the node. This represents a nonlinear activation function used to introduce nonlinear expressive power into the weighted summation of global cross-domain features.

[0042] The nodes are individually selected from global anchors and global information is integrated. This not only utilizes the cross-domain global characteristics of anchors to supplement the context of sparse regions, but also avoids interference from irrelevant global patterns through attention weights. In the end, high-order features that take into account both local details and global context are obtained, providing a more robust representation for fault diagnosis under varying operating conditions.

[0043] In this embodiment, the step of extracting local topological features through local topological path (LTA) specifically includes the following:

[0044] Traditional single-layer graph attention networks (GATs) have limited receptive fields, while multi-layer stacking can easily lead to oversmoothing problems. To address this, this approach introduces node centrality encoding, spatial location encoding, and attention decay factors to flexibly capture fine local and global structural information.

[0045] To measure the critical role of nodes in information propagation and domain confrontation, the graph will be reconstructed. Nodes in The in-degree and out-degree are specified as learnable centrality embedding vectors. and Concatenate it to the original feature to obtain the enhanced feature of node i. :

[0046] Define the spatial relationship function between nodes using the shortest path matrix D. Value is a node With nodes The shortest path length between, i.e. Based on this spatial relationship function, learnable scalar biases are assigned to the spatial location information in the graph. The specific allocation logic is as follows: In this embodiment, a set of one-dimensional learnable space bias parameter tables is predefined. , To determine the maximum path length, different shortest path distance values ​​are mapped to corresponding bias entries in this parameter table. That is, nodes with the same distance share the same scalar bias, such as... All nodes share the bias , All nodes of 2 share For node pairs with no connected paths, i.e. Then a specific independent far-end bias parameter is assigned. This embodiment can directly perceive and model spatial dependencies of different orders using only a single-layer network, effectively alleviating the oversmoothing problem that easily occurs in deep graph neural networks. Simultaneously, by sharing bias parameters according to path distance, the number of parameters is significantly reduced, while fully preserving the structural information of disconnected nodes. The aforementioned spatial bias, as a learnable parameter, is trained end-to-end with the entire model and automatically optimized and updated through backpropagation. This dynamically corrects the attention weights, enhances the model's ability to model global structural information, and thus improves the feature representation and adaptive performance of cross-domain fault diagnosis.

[0047] In this embodiment, an inverse decay function is introduced to strengthen the dominant role of nearest neighbor nodes in feature interactions. in The shortest path length between nodes and satisfying > The attention weights of neighbors of different orders can be dynamically scaled, preserving the structural information of higher-order neighbors while appropriately attenuating noise interference from distant neighbors. This attenuation factor works synergistically with the learnable spatial bias to construct a local attention score that incorporates centrality and spatial awareness. This allows the model to focus on key local interactions based on prior constraints during end-to-end training, while also adaptively adjusting the contribution intensity of nodes at different distances through learnable parameters.

[0048] ; in, Indicates a splicing operation; A learnable weight matrix is ​​used to map the features of nodes and anchors to a unified implicit space. For learnable attention vectors, This represents the dimension of the latent space features after mapping by the weight matrix. This indicates the initial feature dimension extracted from the node.

[0049] To ensure the numerical stability of the feature aggregation process and to reasonably allocate the information contribution of each node within the local neighborhood, a softmax function is used to evaluate the nodes. The attention scores within its local neighborhood are normalized. The local attention weights are then calculated from this normalization. The specific formula is as follows: ; in, This indicates the relationship between nodes in the reconstructed topology graph. The set of neighboring nodes with interconnected relationships. This local attention weight. The value ranges from 0 to 1, dynamically reflecting the relationship between neighboring nodes after combining shortest path decay and node centrality. Features of the target node The actual structural importance.

[0050] After obtaining this weight, the enhanced features within the local neighborhood are weighted and summed to generate local topological features with a wide range of information perception. : ; In this embodiment, the process of deep fusion through residual connections and layer normalization to output high-order node representations specifically includes: By leveraging residual connections and LayerNorm, global and local features are deeply fused to obtain high-order node representations. : ; in, Indicates to The output results of each attention head are concatenated. It is a layer normalization operation. This represents the global cross-domain feature output of the h-th attention head. This represents the local topological feature output of the h-th attention head.

[0051] Step S6: The high-order node representations are jointly optimized using a three-layer unsupervised domain adaptive network architecture to complete fault classification under varying operating conditions.

[0052] In this embodiment, to achieve cross-domain feature alignment between source domain labeled data and target domain unlabeled data under varying operating conditions, while ensuring the accuracy of fault classification, this embodiment constructs a three-layer adversarial adaptive network architecture that integrates a feature extractor, a domain discriminator, and a fault classifier. Through weighted joint optimization of classification loss, domain alignment adversarial loss, and multi-core maximum mean difference loss, it achieves efficient extraction of domain-invariant fault features and accurate fault diagnosis of the unlabeled target domain, ultimately completing the fault classification task under varying operating conditions.

[0053] Classification loss ( To ensure accurate label predictions, the cross-entropy loss function is used to measure the difference between the true and predicted labels. The cross-entropy loss is defined as: ; in, Indicates the first sampled from the source domain. Data for each fault sample; Represents source domain samples The corresponding actual fault category label; This represents the source domain dataset and its corresponding data distribution; This represents the feature extraction result of the i-th sample in the source domain; This represents the prediction result of the label classifier; ) represents the cross-entropy loss; It represents the mathematical expectation.

[0054] Domain alignment adversarial loss ( To address the issue of intra-domain covariate shift, label classifiers trained on source domain data often perform poorly when processing target domain data. This paper introduces a domain discriminator to determine the source of extracted features and trains a feature extractor to mislead the discriminator's judgment. In this adversarial process, the model learns the domain invariance of the features, thus achieving equilibrium. This paper selects binary cross-entropy loss as the domain alignment loss, expressed as follows: ; in, Indicates the first sampled from the source domain. Data for each fault sample; Indicates the first sampled from the target domain Data for each fault sample; This represents the source domain dataset and its corresponding data distribution; This represents the target domain dataset and its corresponding data distribution; and D represents the feature extraction results of the i-th sample in the source domain and the j-th sample in the target domain, respectively; Returns a binary value (0 or 1) to distinguish the sample source, where 0 indicates the sample comes from the source domain. This indicates that the sample comes from the target domain; It represents the mathematical expectation.

[0055] Feature distribution alignment loss ( When aligning the feature representations of the source and target domains, the MMD metric is used to achieve the alignment loss based on the feature distribution differences, as expressed below: ; in, Indicates the first sampled from the source domain. Data for each fault sample; Indicates the first sampled from the target domain Data for each fault sample; This represents the source domain dataset and its corresponding data distribution; This represents the target domain dataset and its corresponding data distribution; and Let represent the feature extraction results of the i-th sample in the source domain and the j-th sample in the target domain, respectively; In the regenerating nucleus Hilbert space The square of the second norm; This represents the kernel mapping function, used to implicitly map the features extracted by the network to a high-dimensional regenerating kernel Hilbert space. middle; It represents the mathematical expectation.

[0056] Overall optimization goal: ; Where θ and μ are weight decay parameters.

[0057] To verify the effectiveness of the proposed graph attention network-based cross-condition bearing fault diagnosis method in cross-condition migration scenarios, experimental verification was conducted using the CWRU bearing fault public dataset. The CWRU dataset contains four typical operating conditions, denoted as 0HP, 1HP, 2HP, and 3HP, corresponding to speeds of 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm, respectively. Each condition includes a normal state and three typical fault types: inner ring fault, outer ring fault, and rolling element fault. The fault diameters cover three severity levels: 0.007 inches, 0.014 inches, and 0.021 inches, constituting a total of 10 bearing health states. Each sample is a one-dimensional vibration signal segment of length 1024, which can be used to characterize the dynamic response features of the bearing system under different operating load conditions.

[0058] In terms of specific category definitions, the 10 bearing health states in the dataset can be represented as follows: Category 0 represents a healthy state; Categories 1, 2, and 3 represent inner ring failures with failure diameters of 0.007, 0.014, and 0.021 inches, respectively; Categories 4, 5, and 6 represent outer ring failures with failure diameters of 0.007, 0.014, and 0.021 inches, respectively; and Categories 7, 8, and 9 represent rolling element failures with failure diameters of 0.007, 0.014, and 0.021 inches, respectively. Here, different failure diameters represent the severity of the failure; the larger the failure diameter, the more severe the failure.

[0059] To verify the migration diagnostic capability of this invention under different operating conditions, this experiment constructed two cross-condition migration tasks. One is a light-load migration task, i.e., a migration task from 1 HP to 0 HP, and the other is a heavy-load migration task, i.e., a migration task from 2 HP to 3 HP. The two sets of tasks respectively cover typical operating condition changes of decreasing and increasing load, which can effectively verify the generalization and migration capabilities of the model under different data distributions.

[0060] The experiment employed an unsupervised adaptive domain setting, using one type of work condition data as the source domain and another as the target domain. The source domain data, with true labels, was randomly divided into training and validation sets in an 8:2 ratio for supervised model training and parameter optimization. The target domain data, without labels, was randomly divided into adversarial training, validation, and test sets in a 4:3:3 ratio. The adversarial training and validation sets were used to assist in domain distribution alignment learning, while the test set was used solely for final model performance evaluation. The classification accuracy on the target domain test set was used as the core evaluation metric for the model. The experimental results are shown in Table 1. Table 1 Experimental results for the two transfer tasks

[0061] The proposed method demonstrates excellent performance on both sets of cross-load migration tasks: the classification accuracy is 96.93% on the 1HP→0HP task and 98.25% on the 2HP→3HP task, with an average accuracy of 97.59% for both sets of tasks. The overall performance is stable, indicating that the method can effectively overcome the data distribution offset under different load conditions and realize knowledge transfer from the source domain to the target domain. It maintains high-precision fault identification capability under both typical scenarios of load reduction and load increase, verifying its effectiveness and generalization in cross-load condition bearing fault diagnosis.

[0062] To further verify the classification performance and feature separability of the proposed method in bearing fault diagnosis across operating conditions, sections 3-6 sequentially present the confusion matrix and t-SNE feature distribution visualization results under two sets of transfer tasks. Figure 3 , 5 Visualization results of the confusion matrix and feature distribution for the 1HP→0HP task. Figure 4 , 6 The confusion matrix and feature distribution visualization results for the 2HP→3HP task are shown. The confusion matrix quantifies the model's accuracy in identifying various bearing health states. The identification accuracy of each fault category in both tasks is at a high level, with only a small number of cross-class misclassifications. The t-SNE feature distribution map further shows that the model can map the features of different fault categories to a feature space with strong intra-class clustering and clear inter-class boundaries, achieving effective distribution alignment between the source and target domains.

[0063] In summary, the graph attention network fault diagnosis method based on adaptive anchor points proposed in this embodiment has demonstrated its effectiveness in varying operating conditions through validation on two typical cross-load migration tasks on the CWRU dataset. Experimental results show that this method not only maintains significant diagnostic accuracy in multi-directional operating condition migration but also effectively suppresses class aliasing within the target domain. The cross-domain aggregation performance in the confusion matrix and T-SNE space demonstrates that the proposed method effectively achieves manifold alignment of cross-load features, improving the structural consistency and global stability of the target domain prediction results. This provides a reliable technical solution for cross-domain migration diagnosis of rolling bearings under complex varying load environments.

[0064] According to the purpose of the embodiments of this application, such as Figure 7 As shown, this embodiment also provides a variable operating condition fault diagnosis device 70 based on adaptive anchor graph attention, applied to perform the diagnosis method described above, specifically including: The signal acquisition and feature extraction module 701 is used to acquire vibration signals of rotating machinery under source domain operating conditions and target domain variable operating conditions, and generate a preliminary node feature matrix after preprocessing. The adaptive asymmetric truncation smoothing module 702 is used to perform adaptive asymmetric truncation smoothing on the preliminary node feature matrix to generate smoothed features after noise reduction. The adaptive anchor point generation module 703 is used to adaptively calculate the number of cross-domain anchor points based on the variance distribution of the smooth feature, the feature dimension and the number of samples in the batch data, and randomly initialize and generate a cross-domain anchor point feature matrix. The topology graph construction and reconstruction module 704 is used to calculate the cosine similarity between nodes in the sample using the smoothed features after noise reduction, construct the initial sparse adjacency matrix through the Top-k function, calculate the shortest path matrix between any nodes using the breadth-first search algorithm, and perform binarization reconstruction mapping on the graph structure to generate the reconstructed topology graph. The dual-path unified graph attention module 705 is used to synchronously input the reconstructed topology graph and the cross-domain anchor feature matrix into the dual-path unified graph attention network; global cross-domain features and local topology features are extracted by global cross-domain path GCP and local topology path LTA respectively, and deep fusion is performed by residual connection and layer normalization to output high-order node representations. The unsupervised domain adaptive joint optimization module 706 is used to jointly optimize the high-order node representation through a three-layer unsupervised domain adaptive network architecture to complete fault classification under varying operating conditions.

[0065] This application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the methods described in the above embodiments.

[0066] This application may be a system, apparatus, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this application.

[0067] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technical improvements to the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of this application is defined by the appended claims.

Claims

1. A variable working condition fault diagnosis method based on adaptive anchor point graph attention, characterized in that, include: Step S1: Collect vibration signals of rotating machinery under source domain operating conditions and target domain variable operating conditions respectively, and generate a preliminary node feature matrix after preprocessing; Step S2: Perform adaptive asymmetric truncation smoothing on the preliminary node feature matrix to generate denoised smooth features; Step S3: Based on the variance distribution, feature dimension, and number of samples in the batch data of the smoothed feature, adaptively calculate the number of cross-domain anchor points and randomly initialize and generate the cross-domain anchor point feature matrix; Step S4: Calculate the cosine similarity between nodes in the sample using the smoothed features after noise reduction, construct the initial sparse adjacency matrix using the Top-k function; calculate the shortest path matrix between any nodes using the breadth-first search algorithm, and perform binarization reconstruction mapping on the graph structure to generate the reconstructed topology graph; Step S5: Simultaneously input the reconstructed topology graph and the cross-domain anchor feature matrix into the dual-path unified graph attention network; Global cross-domain features and local topological features are extracted by global cross-domain path (GCP) and local topological path (LTA), respectively, and then deep fusion is performed by residual connection and layer normalization to output high-order node representations. Step S6: The high-order node representations are jointly optimized using a three-layer unsupervised domain adaptive network architecture to complete fault classification under varying operating conditions.

2. The method of claim 1, wherein, Step S2, which involves performing adaptive asymmetric truncation smoothing on the initial node feature matrix to generate denoised smooth features, specifically includes: Based on the median, maximum, and minimum values ​​of each feature dimension of the preliminary node feature matrix in the current batch of data, calculate the positive and negative offset distances that measure the offset direction and severity of the impact noise under varying operating conditions. Based on the relative ratio of positive offset distance to negative offset distance, the total truncation ratio threshold is adaptively decomposed into an upper truncation rate and a lower truncation rate. For each feature dimension, based on the upper and lower truncation rates, extreme feature values ​​at the top and bottom of the ranking are removed, the robust truncation mean of the remaining features is calculated, and the features are normalized using the L2 norm to generate denoised smooth features.

3. The method of claim 2, wherein, Step S3, which involves adaptively calculating the number of cross-domain anchor points based on the variance distribution, feature dimension, and number of samples in the batch data of the smoothed feature, and randomly initializing and generating the cross-domain anchor point feature matrix, specifically includes: The number of cross-domain anchor points is calculated according to the following formula : ； wherein, represents the number of samples of the batch data, represents the overall variance of the feature, represents a hyperparameter; The generated cross-domain anchor feature matrix A is represented as follows: ； wherein, denotes an anchor point, denotes a raw feature dimension preliminarily extracted by a node, denotes a real set, denotes a smoothed feature after noise reduction.

4. The adaptive anchor graph attention based variable operating condition fault diagnosis method according to claim 3, characterized in that, In step S4, the initial sparse adjacency matrix B is represented as: ； wherein, denotes the smoothed features after noise reduction, and the Top-k function is used to extract the top k edges most relevant to each node. The breadth-first search (BFS) algorithm is used to calculate the shortest path matrix D between any nodes, which is represented as: D = BFS(B) wherein the elements in D denotes the shortest path length from the i-th node to the j-th node ; Binarization reconstruction mapping is performed on the graph structure to generate a reconstruction topology graph mapping matrix ; in, ； wherein, denotes an element in the reconstructed topological map matrix denotes an element in the reconstructed topological map matrix 5. The method of claim 4, wherein, Step S5, specifically the step of extracting global cross-domain features through the Global Cross-Domain Path (GCP), includes: All nodes of the source domain and the target domain are fully connected with the anchor points for full connection attention interaction Global interaction attention scores are represented as: ； The global interaction attention score is converted into a normalized global attention weight by a softmax function is represented as: ； By global attention weights The global cross-domain features are calculated by the transformation features of all anchor points , denoted as: ； in, This indicates a concatenation operation; m takes the value of LeakyReLU is the activation function; For learnable weight matrix, For learnable attention vectors, This represents the dimension of the latent space features after mapping by the weight matrix. This represents the initial feature dimension extracted from the node. This represents a nonlinear activation function used to introduce nonlinear expressive power into the weighted summation of global cross-domain features.

6. The adaptive anchor graph attention based variable operating condition fault diagnosis method according to claim 5, characterized in that, Step S5, the step of extracting local topological features through the local topological path LTA, specifically includes: specifying the in-degree and out-degree of the nodes in the reconstructed topological graph as learnable centrality embedding vectors , resulting in enhanced features of the nodes :​​​ ； Learnable scalar biases assigned based on spatial location information in a predefined reconstructed topology graph. Enhanced features and inverse decay function Obtain the local attention score that includes centrality and spatial awareness. , is represented as: ； The local attention weight is calculated by normalizing the local attention score by a softmax function and is expressed as: ； According to the local attention weight and the enhanced feature The local topological feature is calculated is expressed as: ； in, This indicates a splicing operation. This indicates the relationship between nodes in the reconstructed topology graph. The set of neighboring nodes that are connected; Learnable weight matrix For learnable attention vectors, This represents the dimension of the latent space features after mapping by the weight matrix. This indicates the initial feature dimension extracted from the node.

7. The adaptive anchor graph attention based variable operating condition fault diagnosis method according to claim 6, characterized in that, In step S5, the high-order node representation is output through deep fusion via residual connections and layer normalization. Represented as: ； wherein, denotes a concatenation operation on the output results of the attention heads; is a layer normalization operation, denotes the global cross-domain feature output result of the h-th attention head, denotes the local topological feature output result of the h-th attention head.

8. The adaptive anchor graph attention based variable operating condition fault diagnosis method according to claim 7, characterized in that, The three-layer unsupervised domain adaptive network architecture includes a feature extractor, a domain discriminator, and a fault classifier. Through weighted joint optimization of classification loss, domain alignment adversarial loss, and multi-kernel maximum mean difference loss, it achieves efficient extraction of domain-invariant fault features and accurate fault diagnosis of unlabeled target domains, ultimately completing the fault classification task under varying operating conditions.

9. A variable working condition fault diagnosis apparatus based on adaptive anchor graph attention, configured to perform the variable working condition fault diagnosis method according to any one of claims 1-8, characterized in that, Specifically, it includes: The signal acquisition and feature extraction module is used to acquire vibration signals of rotating machinery under source domain operating conditions and target domain variable operating conditions, and generate a preliminary node feature matrix after preprocessing. An adaptive asymmetric truncation smoothing module is used to perform adaptive asymmetric truncation smoothing on the preliminary node feature matrix to generate smoothed features after noise reduction. An adaptive anchor point generation module is used to adaptively calculate the number of cross-domain anchor points based on the variance distribution of the smoothed features, the feature dimension, and the number of samples in the batch data, and to randomly initialize and generate a cross-domain anchor point feature matrix. The topology graph construction and reconstruction module is used to calculate the cosine similarity between nodes in the sample using the smoothed features after noise reduction, construct the initial sparse adjacency matrix through the Top-k function, calculate the shortest path matrix between any nodes using the breadth-first search algorithm, and perform binarization reconstruction mapping on the graph structure to generate the reconstructed topology graph. The dual-path unified graph attention module is used to synchronously input the reconstructed topology graph and the cross-domain anchor feature matrix into the dual-path unified graph attention network; global cross-domain features and local topology features are extracted by global cross-domain path GCP and local topology path LTA respectively, and deep fusion is performed by residual connection and layer normalization to output high-order node representations. The unsupervised domain adaptive joint optimization module is used to jointly optimize the high-order node representation through a three-layer unsupervised domain adaptive network architecture to complete fault classification under varying operating conditions.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 8.

Images