A method for intermittent fault diagnosis in low-dimensional interconnect networks based on graph attention mechanism

Abstract

This application belongs to the field of interconnection network reliability and fault diagnosis technology, and discloses a method for intermittent fault diagnosis of low-bandwidth long interconnection networks based on graph attention mechanism. Targeting the hierarchical recursive structure and high connectivity of the network, under the PMC fault diagnosis model, a multi-round testing strategy is used to obtain test symptoms within the node's neighborhood. A feature vector is constructed for each node using local statistical feature extraction methods, and preprocessed using zero-padding and noise enhancement techniques. Finally, a graph attention network model is constructed, and the importance weights of neighboring node test results are dynamically learned using the attention mechanism to achieve accurate diagnosis of node fault states in the network. This application leverages the powerful local information aggregation capability of graph attention networks to overcome the diagnostic limitations of traditional algorithms, maintaining high diagnostic accuracy and robustness even with high fault rates, incomplete test symptoms, and large network sizes.

Description

Technical Field

[0001] This application belongs to the field of interconnection network reliability and fault diagnosis technology, specifically relating to a method for diagnosing intermittent faults in low-dimensional long interconnection networks based on graph attention mechanism. Background Technology

[0002] With the development of high-performance computing and data centers, the scale of interconnected networks is growing rapidly, increasing the probability of system component failures. Among various types of failures, intermittent failures are more difficult to detect and diagnose than permanent failures due to their randomness, transient nature, and repetitiveness. Intermittent failures cause processors to perform normally at some times and abnormally at others, and this uncertainty poses a significant challenge to system maintenance.

[0003] Augmentation -Yuan -cube( As an important type of low-profile, long-range interconnected network topology, the network has... There are nodes, each node is represented as... tuple ,in . The network is organized in a hierarchical structure. isomorphic Subnetworks, this hierarchical recursive structure gives the network high connectivity, low diameter and good scalability.

[0004] Traditional system-level fault diagnosis algorithms are typically based on specific diagnostic models such as the PMC model, and their effectiveness is limited by the system's theoretical diagnostic capability. When the number of faulty nodes in the network exceeds the theoretical limit, traditional algorithms often cannot guarantee the accuracy of the diagnosis. Furthermore, traditional algorithms usually assume the ability to acquire complete test symptoms, while in practical applications, due to communication congestion or link failures, only partial symptom data is often obtained. For interconnected networks with complex hierarchical structures, existing technologies have the following shortcomings:

[0005] 1. Insufficient utilization of network topology information: Traditional methods often treat symptom data as isolated test results, failing to effectively utilize the connectivity between nodes and the locality of fault propagation; 2. Lack of adaptive testing strategies: Fixed-round testing cannot adapt to networks of different sizes and complexities, potentially leading to wasted resources or insufficient diagnostic information; 3. Limited feature extraction methods: Failure to design feature extraction methods that can effectively reflect the time-varying characteristics of intermittent faults; 4. Weak generalization ability of diagnostic models: Traditional algorithms are highly dependent on complete test symptoms, and their performance degrades significantly when data is missing or noise is present.

[0006] In recent years, fault diagnosis methods based on neural networks have been proposed. However, most existing methods train the entire network as a whole, ignoring the local characteristics of fault propagation, which leads to the introduction of irrelevant noise. Furthermore, they are highly dependent on large-scale training data and are difficult to adapt to dynamically changing network environments. Summary of the Invention

[0007] To achieve the above objectives, this application provides a method for diagnosing intermittent faults in low-bandwidth long interconnected networks based on graph attention mechanisms. This method utilizes the powerful local information aggregation capability of graph attention networks to overcome the diagnostic limitations of traditional algorithms. It can maintain high diagnostic accuracy and robustness even when the fault rate is high, the test symptoms are incomplete, and the network scale is large. It is suitable for real-time fault detection in large-scale data centers and high-performance computing clusters.

[0008] To achieve the above objectives, this application employs the following technical solution:

[0009] This application discloses a method for diagnosing intermittent faults in low-dimensional long interconnect networks based on graph attention mechanisms, characterized by the following steps:

[0010] Step 1, Build The network topology model, the Network includes There are nodes, each node is represented as... tuple ,in The network is organized according to a hierarchical structure. ,in Indicates the first Isomorphism The sub-network, based on the PMC diagnostic model, for Nodes in the network perform Rounds of mutual testing were conducted, and symptom data were collected in each round. , The dimension parameter represents the length of the coordinate tuple for each node. The base parameter represents the boundary of the value range for each coordinate component. For nodes, For neighboring nodes;

[0011] Step 2: Process the collected data Seasonal symptom data Perform local statistical analysis, that is, for the current node and adjacent nodes Furthermore, statistical analysis is not performed on non-adjacent nodes; each node is calculated. For all its neighboring nodes The test anomaly ratio, extract each node Local statistical eigenvectors ;

[0012] Step 3: Extract local statistical feature vectors Preprocessing is performed, which includes dimensionality unification and adaptive feature enhancement. Dimensionality unification yields a unified feature vector, and adaptive feature enhancement yields an enhanced feature vector.

[0013] Step 4: Construct a fault diagnosis model based on a graph attention network. This model includes a feature transformation layer, a multi-head attention mechanism layer, and a classification layer. Using the enhanced feature vector obtained in Step 3 as input, the enhanced feature vector is mapped to a high-dimensional latent feature space representing the fault state of the nodes. The multi-head attention mechanism layer then classifies the model based on the symptom data. The credibility is adaptively assigned to the aggregation weight, and the classification layer outputs a probability vector of each node as a normal node or an intermittently faulty node.

[0014] Step 5: Using the enhanced feature vector obtained from the adaptive feature enhancement process in Step 3, train the fault diagnosis model based on the graph attention network. Then, use the trained fault diagnosis model to analyze new targets. All in the network Each node is classified into different fault states to determine whether it is a normal node or an intermittent fault node.

[0015] A further improvement of this application is that, in step 1, the... The node connectivity of a network is defined as follows: for a node and neighboring nodes A node is valid if and only if one of the following two conditions is met. with neighboring nodes There are connected edges:

[0016] (1) There exists a unique dimension , , making And for all ,have ;

[0017] (2) For all dimensions , They all The network is layered according to the first dimension coordinate, forming isomorphic Subnetwork.

[0018] A further improvement of this application is that, in step 1, the number of test rounds... , It is a constant.

[0019] A further improvement in this application is that step 2 specifically involves: for Every node in the network compute nodes For each neighbor node The percentage of test anomalies:

[0020]

[0021] in, For the number of test rounds, Indicates the number of collected Round test symptom data;

[0022] Node For all its neighboring nodes Test anomaly ratio Combine them to form nodes Local statistical eigenvectors , It is a node The set of all neighboring nodes.

[0023] A further improvement of this application is that, in step 3, the dimensional unification process takes into account... The maximum degree of a node in the network is specifically obtained by: Maximum node degree in the network For nodes with a degree less than the maximum node degree Local statistical eigenvectors Zero-padding is applied to the end of the local statistical eigenvector to make the local statistical eigenvectors more stable. The length is uniform This yields a unified feature vector.

[0024] A further improvement of this application is that, in step 3, the adaptive feature enhancement processing is based on the network size. The noise parameters are adjusted as follows: Gaussian noise is added to the unified feature vector obtained through dimensionality unification, with the noise parameters being... ,in The adjustment coefficient is used to obtain the enhanced feature vector, which is then truncated within the range [0,1].

[0025] A further improvement in this application is that, in step 4,

[0026] The feature transformation layer uses each node in The enhanced feature vectors received in the round of testing are used as input. The enhanced feature vectors are linearly projected through the learnable weight matrix and nonlinear activation is applied. The enhanced feature vectors of each node are embedded into the high-dimensional latent feature space to obtain the initial node embedding vector of each node.

[0027] The multi-head attention mechanism layer is for each node. neighboring nodes Assign weights, aggregate neighbor node information, and when aggregating neighbor node information, for the first... Each attention head can be used to learn attention vectors. For nodes with neighboring nodes After concatenating the embedding vectors, a linear transformation is performed, and the attention coefficients are calculated:

[0028]

[0029] The aggregate weights are obtained after softmax normalization:

[0030]

[0031] in, Let represent the transpose of the learnable attention vector of the i-th attention head. For learnable weight matrix, and Representing nodes respectively and neighboring nodes The feature representation vector obtained after the previous transformation This represents a vector concatenation operation; the fault diagnosis model based on graph attention networks aggregates weights. Adaptively assign aggregation weights to trusted neighbor nodes to infer the fault state of the target node under the condition that the faulty node generates random test interference;

[0032] The classification layer is a binary classification output layer, and the output nodes are two-dimensional normalized probability vectors of normal nodes or intermittent fault nodes.

[0033] A further improvement in this application is that, in step 5, the fault state classification method is as follows: using symptom data with labeled node fault state tags collected through multiple rounds of PMC diagnostic model testing in a simulation environment, the fault diagnosis model based on graph attention network is trained, and the target... The enhanced feature vectors of all nodes in the network are constructed into graph data according to the topology. This data is then input into a well-trained fault diagnosis model based on graph attention network. After feature transformation and aggregation of neighborhood attention weights, a two-dimensional normalized probability vector is output for each node. The fault state is classified by taking the maximum probability, thereby determining whether the node is a normal node or an intermittent fault node.

[0034] The beneficial effects of this application are:

[0035] This application combines multi-round test statistical features and graph attention mechanism to accurately capture intermittent fault patterns with high accuracy;

[0036] This application maintains a high diagnostic accuracy even when the number of faulty nodes exceeds the traditional theoretical diagnostic accuracy by as much as 50%, thus breaking through the theoretical limit.

[0037] This application has a high tolerance for partial syndromes and can effectively diagnose even when some test results are lost, demonstrating strong robustness.

[0038] This application utilizes the local aggregation characteristics of graph attention networks to perform diagnosis based solely on the local neighborhood information of nodes, which aligns with the actual diagnostic scenarios of distributed systems. Attached Figure Description

[0039] Figure 1 This is a flowchart of the fault diagnosis method proposed in this application.

[0040] Figure 2 This is a schematic diagram of the multi-round mutual testing process of this application.

[0041] Figure 3 This is a schematic diagram of the fault diagnosis model based on graph attention network in this application.

[0042] Figure 4 This application Network topology diagram, showing augmentation -Yuan - Hierarchical recursive properties of cube networks.

[0043] Figure 5 This is a comparison of the accuracy of this application (FaultGAT) with the traditional neural network diagnostic method RNNIFDCOM based on full-map signs in a simulation environment. Detailed Implementation

[0044] The embodiments of the present invention will be disclosed below with reference to the drawings. For clarity, many practical details will be described in the following description. However, it should be understood that these practical details are not intended to limit the present invention. That is, in some embodiments of the present invention, these practical details are not essential. In addition, for the sake of simplicity, some conventional structures and components will be shown in the drawings in a simple schematic manner.

[0045] like Figures 1-2 As shown, a method for intermittent fault diagnosis of low-gap long interconnection networks based on graph attention mechanism is proposed, targeting a network with... Enhanced per processor -Yuan -cube( The diagnostic method for this network includes the following steps:

[0046] Step 1, Build The network topology model, the Network includes There are nodes, each node is represented as... tuple ,in The network is organized according to a hierarchical structure. ,in Indicates the first Isomorphism The sub-network, based on the PMC diagnostic model, for Nodes in the network perform Rounds of mutual testing were conducted, and symptom data were collected in each round. , The dimension parameter represents the length of the coordinate tuple for each node. The base parameter represents the boundary of the value range for each coordinate component. For nodes, For neighboring nodes.

[0047] In this step, the The node connectivity of a network is defined as follows: for a node and neighboring nodes A node is valid if and only if one of the following two conditions is met. with neighboring nodes There are connected edges:

[0048] (1) There exists a unique dimension , , making And for all ,have ;

[0049] (2) For all dimensions , They all The network is layered according to the first dimension coordinate, forming isomorphic Subnetwork.

[0050] like Figure 2 shown A diagram illustrating the mutual testing and extraction of wheels. Figure 4 For The following is a diagram illustrating a hierarchical topology, exemplified by the example network topology. .according to Adaptive network size setting for test rounds .in The number of rounds is the base test round. In each round... In the graph, for each directed edge... ,node For neighboring nodes Tests were conducted. Test results and symptom data were collected. Follow these rules:

[0051] If node Normal and neighboring nodes If the current performance is normal, then ;

[0052] If node Normal and neighboring nodes If it is an intermittently faulty node and is currently in an active fault state, then ;

[0053] like Normal, neighboring nodes If a node is experiencing intermittent failure but is currently in a temporarily normal state, then ;

[0054] If node If the node is an intermittent fault and the test results are unreliable regardless of its current state, then... A random value that is either 0 or 1, where, .

[0055] Step 2: Process the collected data Seasonal symptom data Perform local statistical analysis, that is, for the current node and adjacent nodes Furthermore, statistical analysis is not performed on non-adjacent nodes; each node is calculated. For all its neighboring nodes The test anomaly ratio, extract each node Local statistical eigenvectors Specifically: for Every node in the network compute nodes For each neighbor node The percentage of test anomalies:

[0056]

[0057] in, For the number of test rounds, Indicates the number of collected Round test symptom data;

[0058] Node For all its neighboring nodes Test anomaly ratio Combine them to form nodes Local statistical eigenvectors , It is a node The set of all neighboring nodes.

[0059] Step 3: Extract local statistical feature vectors Preprocessing is performed, including dimensionality unification and adaptive feature enhancement.

[0060] In step 3, a unified feature vector is obtained through dimensionality unification, and an enhanced feature vector is obtained through adaptive feature enhancement. The dimensionality unification process considers… The maximum degree of a node in the network is specifically obtained by: Maximum node degree in the network For nodes with a degree less than the maximum node degree Local statistical eigenvectors Zero-padding is applied to the end of the local statistical eigenvector to make the local statistical eigenvectors more stable. The length is uniform This yields a unified feature vector.

[0061] In step 3, the adaptive feature enhancement process is based on the network size. The noise parameters are adjusted as follows: Gaussian noise related to the network size is added to the unified feature vector obtained by dimensionality unification. The noise parameters are... ,in The adjustment coefficient is used to obtain the enhanced feature vector, which is then truncated within the range [0,1].

[0062] Step 4: Construct a fault diagnosis model based on a graph attention network (GAT). Using the enhanced feature vector obtained in step 3 as input, map the enhanced feature vector to a high-dimensional latent feature space representing the fault state of the node. The multi-head attention mechanism layer then applies the symptom data... The credibility is adaptively assigned to the aggregation weight, and the classification layer outputs a probability vector of each node as a normal node or an intermittently faulty node.

[0063] like Figure 3 As shown, the fault diagnosis model includes a feature transformation layer, a multi-head attention mechanism layer, and a classification layer. The feature transformation layer uses the values ​​of each node in the classification layer to determine the fault diagnosis model. The enhanced feature vectors received in the round of testing are used as input. The enhanced feature vectors are linearly projected through the learnable weight matrix and nonlinear activation is applied. The enhanced feature vectors of each node are embedded into the high-dimensional latent feature space to obtain the initial node embedding vector of each node.

[0064] The multi-head attention mechanism layer is for each node. neighboring nodes Assign weights, aggregate neighbor node information, and when aggregating neighbor node information, for the first... Each attention head can be used to learn attention vectors. For nodes with neighboring nodes After concatenating the embedding vectors, a linear transformation is performed, and the attention coefficients are calculated:

[0065]

[0066] The aggregate weights are obtained after softmax normalization:

[0067]

[0068] in, Let represent the transpose of the learnable attention vector of the i-th attention head. For learnable weight matrix, and Representing nodes respectively and neighboring nodes The feature representation vector obtained after the previous transformation This represents a vector concatenation operation; the fault diagnosis model based on graph attention networks aggregates weights. Adaptively assign aggregation weights to trusted neighbor nodes to infer the fault state of the target node under the condition that the faulty node generates random test interference;

[0069] The classification layer is a binary classification output layer, and the output nodes are two-dimensional normalized probability vectors of normal nodes or intermittent fault nodes.

[0070] Step 5: Using the enhanced feature vector obtained from the adaptive feature enhancement process in Step 3, train the fault diagnosis model based on the graph attention network. Then, use the trained fault diagnosis model to analyze new targets. All in the network Each node is classified into a fault state to determine whether it is a normal node or an intermittent fault node. In this step, the fault state classification method is as follows: using symptom data with labeled node fault state tags collected through multiple rounds of PMC diagnostic model testing in a simulation environment, a fault diagnosis model based on a graph attention network (GAT) is trained, and the target... The enhanced feature vectors of all nodes in the network are constructed into graph data according to the topology. This data is then input into a well-trained fault diagnosis model based on graph attention network (GAT). After feature transformation and aggregation of neighborhood attention weights, a two-dimensional normalized probability vector is output for each node. The fault state is classified by taking the maximum probability, thereby determining whether the node is a normal node or an intermittent fault node.

[0071] like Figure 5 As shown, tests were conducted in a simulation environment for scenarios with different failure rates, such as 10% to 50%, and different symptom missing rates, such as 10% to 50%. The results show that the accuracy of this application (FaultGAT) is close to 100% within the theoretical diagnostic limit. Even in high failure rate scenarios exceeding the theoretical limit, the accuracy remains above 90%, significantly outperforming the traditional neural network diagnostic method RNNIFDCOM based on full-map symptoms.

[0072] This application leverages the powerful local information aggregation capability of graph attention networks to overcome the diagnostic limitations of traditional algorithms. It maintains high diagnostic accuracy and robustness even with high failure rates, incomplete test symptoms, and large network scales, making it suitable for real-time fault detection in large-scale data centers and high-performance computing clusters.

[0073] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for diagnosing intermittent faults in low-dimensional long interconnect networks based on graph attention mechanism, characterized in that: The diagnostic method includes the following steps: Step 1, Build The network topology model, the Network includes There are nodes, each node is represented as... tuple ,in The network is organized according to a hierarchical structure. ,in Indicates the first Isomorphism The subnetwork, based on the PMC diagnostic model, for Nodes in the network perform Rounds of mutual testing were conducted, and symptom data were collected in each round. ,in, For dimension parameters, For cardinality parameter, For nodes, For neighboring nodes; Step 2: Process the collected data Symptom data from round testing Perform local statistical analysis, that is, for the current node and adjacent nodes Furthermore, statistical analysis is not performed on non-adjacent nodes; each node is calculated. For nodes All neighboring nodes The test anomaly ratio, extract each node Local statistical eigenvectors ; Step 3: Extract local statistical feature vectors Preprocessing is performed, which includes dimensionality unification and adaptive feature enhancement. Dimensionality unification yields a unified feature vector, and adaptive feature enhancement yields an enhanced feature vector. Step 4: Construct a fault diagnosis model based on a graph attention network. This model includes a feature transformation layer, a multi-head attention mechanism layer, and a classification layer. The obtained enhanced feature vectors are used as input, and these vectors are mapped to a high-dimensional latent feature space representing the fault state of nodes. The multi-head attention mechanism layer then applies the symptom data... The credibility is adaptively assigned to the aggregation weight, and the classification layer outputs a probability vector of each node as a normal node or an intermittently faulty node. Step 5: Using the enhanced feature vector obtained from the adaptive feature enhancement process in Step 3, train the fault diagnosis model based on the graph attention network. Then, use the trained fault diagnosis model to analyze new targets. All in the network Each node is classified into different fault states to determine whether it is a normal node or an intermittent fault node.

2. The intermittent fault diagnosis method for low-circumference long interconnection networks based on graph attention mechanism according to claim 1, characterized in that: In step 1, the The node connectivity of a network is defined as follows: for a node and neighboring nodes A node is valid if and only if one of the following two conditions is met. with neighboring nodes There are connected edges: (1) There exists a unique dimension , , making And for all ,have ,in, Indicates the first 1 node Indicates the first One node; (2) For all dimensions , They all The network is layered according to the first dimension coordinate, forming isomorphic Subnetwork.

3. The intermittent fault diagnosis method for low-circumference long interconnection networks based on graph attention mechanism according to claim 1, characterized in that: In step 1, the number of test rounds , It is a constant. The number of rounds for the basic test.

4. The intermittent fault diagnosis method for low-circumference long interconnection networks based on graph attention mechanism according to claim 1, characterized in that: Step 2 specifically involves: For Every node in the network compute nodes For each neighbor node The percentage of test anomalies: in, For the number of test rounds, Indicates the number of collected Round test symptom data; Node For nodes All neighboring nodes Test anomaly ratio Combine them to form nodes Local statistical eigenvectors , It is a node The set of all neighboring nodes.

5. The method for diagnosing intermittent faults in low-circumference long interconnect networks based on graph attention mechanism according to claim 1, characterized in that: In step 3, the dimensional unification process considers The maximum degree of a node in the network is specifically obtained by: Maximum node degree in the network For nodes with a degree less than the maximum node degree Local statistical eigenvectors Zero-padding is applied to the end of the local statistical eigenvector to make the local statistical eigenvectors more stable. The length is uniform This yields a unified feature vector.

6. The intermittent fault diagnosis method for low-circumference long interconnection networks based on graph attention mechanism according to claim 5, characterized in that: In step 3, the adaptive feature enhancement process is based on the network size. The noise parameters are adjusted as follows: Gaussian noise is added to the unified feature vector obtained through dimension unification to obtain an enhanced feature vector. This enhanced feature vector is then truncated to the range [0,1]. The noise parameters are as follows: , This is the adjustment coefficient.

7. The intermittent fault diagnosis method for low-circumference long interconnection networks based on graph attention mechanism according to claim 1, characterized in that: In step 4, The feature transformation layer uses each node in The enhanced feature vectors received in the round of testing are used as input. The enhanced feature vectors are linearly projected through the learnable weight matrix and nonlinear activation is applied. The enhanced feature vectors of each node are embedded into the high-dimensional latent feature space to obtain the initial node embedding vector of each node. The multi-head attention mechanism layer is for each node. neighboring nodes Assign weights, aggregate neighbor node information, and when aggregating neighbor node information, for the first... Each attention head can be used to learn attention vectors. For nodes with neighboring nodes After concatenating the embedding vectors, a linear transformation is performed, and the attention coefficients are calculated: The aggregate weights are obtained after softmax normalization: in, Let represent the transpose of the learnable attention vector of the i-th attention head. For learnable weight matrix, and Representing nodes respectively and neighboring nodes The feature representation vector obtained after the previous transformation This represents a vector concatenation operation; the fault diagnosis model based on graph attention networks aggregates weights. Adaptively assign aggregation weights to trusted neighbor nodes to infer the fault state of the target node under the condition that the faulty node generates random test interference; The classification layer is a binary classification output layer, and the output nodes are two-dimensional normalized probability vectors of normal nodes or intermittent fault nodes.

8. The intermittent fault diagnosis method for low-circumference long interconnection networks based on graph attention mechanism according to claim 1, characterized in that: In step 5, the fault state classification method is as follows: using symptom data with labeled node fault state tags collected through multiple rounds of PMC diagnostic model testing in a simulation environment, the fault diagnosis model based on graph attention network is trained, and the target... The enhanced feature vectors of all nodes in the network are constructed into graph data according to the topology. This data is then input into a well-trained fault diagnosis model based on graph attention network. After feature transformation and aggregation of neighborhood attention weights, a two-dimensional normalized probability vector is output for each node. The fault state is classified by taking the maximum value of the probability vector, thereby determining whether the node is a normal node or an intermittent fault node.

Images