Power transmission system fault correlation analysis method and system based on feature map matching

Abstract

The application relates to a power transmission system fault correlation analysis method and system based on feature graph matching and belongs to the technical field of fault diagnosis. By performing three-dimensional feature coding on the multi-physical quantity signal data of a power transmission system, unified feature extraction and standardized representation of the multi-physical quantity signal are realized; an improved fast Apriori algorithm is designed according to the three-dimensional feature coding data to extract the correlation rules between the multi-physical quantity signals in the power transmission system, and the stability, physical correlation and statistical correlation of the correlation rules are comprehensively considered to quantize the value of the feature rules and convert the value into a structured fault feature graph; a graph matching algorithm is used to mine the fault feature graph and construct a correlation fault model of the power transmission system. The method can fuse the feature information of multiple physical quantity signals of the power transmission system, reveal the potential signal or mechanism commonality among some faults of the system, and provide a theoretical basis and visual support for carrying out multi-fault identification.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a method and system for fault correlation analysis of power transmission systems based on feature map matching. Background Technology

[0002] With the advancement of global energy structure transformation and the trend towards green and low-carbon development, new energy technologies have developed rapidly. Compared with traditional diesel-powered vehicles, new energy vehicles (such as hybrid and electric vehicles) exhibit significant advantages in power response, energy efficiency, emission control, and noise suppression. Hybrid systems, through the coordinated operation of an electric motor and a diesel engine, can achieve faster torque response and stronger acceleration from a standstill, meeting the requirements for high energy density and quiet operation, and possessing zero emissions and low noise characteristics in low-speed driving and standby states. Compared with traditional transmission systems, the new generation of powertrain systems has simplified mechanical structures, resulting in a more compact overall system structure, while also possessing higher integration and information levels. Therefore, fault correlation analysis of powertrain systems, thereby providing effective assistance for system fault diagnosis, has always been a research hotspot in the field of fault diagnosis.

[0003] The damage morphologies of various faults in power transmission systems often involve multiple complex physical mechanisms. Their formation and evolution are influenced by a combination of mechanical, thermal, electrical, magnetic, and hydraulic factors, exhibiting high correlation and nonlinear characteristics. Traditional diagnostic methods have achieved certain results in fault feature extraction and classification, possessing high identification accuracy. However, because they are mostly based on single fault modes or local signals, they often suffer from limitations in diagnostic perspective and fragmented fault response mechanisms, making it difficult to comprehensively capture the high-dimensional coupling characteristics and fault propagation laws among multiple physical quantity signals in the system.

[0004] In engineering practice, current fault correlation analysis techniques typically rely on a deep understanding of prior information such as system fault mechanisms, evolution paths, and propagation mechanisms. Correlation metrics are selectively introduced during model construction to enhance feature extraction capabilities and improve fault identification accuracy and model generalization. However, for highly complex powertrain systems with tightly coupled operating states, real challenges remain, such as difficulties in acquiring prior knowledge, scarcity of fault samples, and unclear fault mechanisms. Due to a lack of systematic knowledge support, it is difficult to effectively model the causal relationships and propagation logic among multi-source fault features, often forcing existing methods to rely on purely data-driven strategies. Under complex and variable operating conditions, these methods are susceptible to noise interference and data bias, leading to insufficient reliability of diagnostic results and high false diagnosis rates. Therefore, for complex systems like powertrain systems, there is an urgent need to explore a hybrid fault correlation analysis framework that integrates limited prior knowledge with the advantages of data-driven approaches. It is necessary to combine feature graph mining and graph matching networks for fault correlation analysis of powertrain systems. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of low reliability and high misdiagnosis rate in existing transmission system fault diagnosis methods, and to provide a power transmission system fault correlation analysis method and system based on feature map matching.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for fault correlation analysis of a power transmission system based on feature map matching, comprising the following steps: Multiple physical quantity signal data of the power transmission system are collected and multiple feature indicators are extracted. The Relief algorithm is used to optimize the feature indicators and obtain the dominant features. Fuzzy C-means clustering is used to cluster the dominant features. Based on the clustering results, the dominant feature points are discretized. The clustering results, dominant features and discrete intervals are uniformly encoded to construct three-dimensional feature coding data. Based on three-dimensional feature encoding data, an improved fast Apriori algorithm is designed to mine highly correlated bivariate frequent itemsets and extract the association rules between multiple physical quantity signals in the power transmission system. Based on the correlation rules between multiple physical quantity signals extracted from the power transmission system, a multi-index entropy-hierarchical analysis weighted feature rule evaluation system is constructed. By integrating stability, physical correlation and statistical correlation, the value of feature rules is quantified, and the weighted feature rules are transformed into a structured fault feature map. Based on the constructed fault feature map, a graph matching network is used to mine the correlation matrix between different fault modes. Then, a spectral clustering strategy is used to divide the correlation matrix to obtain the correlation hierarchy structure between each fault mode and construct a fault correlation model for the power transmission system. Real-time acquisition of various physical quantity signal data of the power transmission system, and correlation analysis is performed through the constructed fault correlation model.

[0007] In the step of acquiring various physical quantity signal data of the power transmission system and extracting multiple feature indicators, the Relief algorithm is used to optimize the feature indicators to obtain dominant features. Fuzzy C-means clustering is then used to cluster the dominant features. Based on the clustering results, the dominant feature points are discretized. The clustering results, dominant features, and discrete intervals are uniformly encoded to construct three-dimensional feature-encoded data. The formula for optimizing the feature indicators using the Relief algorithm to obtain the dominant features is as follows:

[0008] in, For the first Features at the next iteration Weight value, For the sample In category The Middle The nearest neighbor samples, For category Prior probability, For the sample With its kind Neighbor In features The degree of difference.

[0009] Fuzzy C-means clustering is used, through membership functions The clustering results are obtained by measuring the degree to which dominant feature points belong to different cluster centers, using the following formula:

[0010] in, As the dominant feature point, For the first Cluster centers, This is the fuzzy coefficient (usually between 1.5 and 2.5). The distance between the dominant feature point and the cluster center is the Euclidean distance.

[0011] The specific method for the step of designing an improved fast Apriori algorithm based on three-dimensional feature encoding data to mine highly correlated bivariate frequent itemsets and extract the association rules between multiple physical quantity signals in the power transmission system is as follows: Based on the constructed three-dimensional feature encoding data, a minimum support threshold is set, a candidate option set is generated starting from a single item, and a hash tree is constructed to store the candidate option set. A preset hash function is used to map the combination of items in the transaction to the corresponding leaf nodes in the hash tree, and the candidate option set is matched on these nodes. When a match is successful, the support count of the candidate option set is updated. Based on the obtained support count, the candidate itemsets that are less than the minimum support threshold are pruned to obtain the corresponding frequent itemsets. Using the obtained frequent itemsets, higher-order candidate itemsets and frequent itemsets are iteratively generated until no new candidate itemsets can be generated. For each frequent itemset, its non-empty subsets are used as the preceding terms of the rules, utilizing support data. Calculate the corresponding confidence level and improvement The association rules between multiple physical quantity signals in the power transmission system are selected based on the calculated confidence and lift.

[0012] The formulas for calculating the confidence level and lift are as follows: ;

[0013] in, For itemsets Support For itemsets Support For itemsets Support level.

[0014] The specific method for constructing a multi-index entropy-hierarchical analysis weighted feature rule evaluation system based on the correlation rules between multiple physical quantity signals extracted from the power transmission system, integrating stability, physical correlation, and statistical correlation to quantify the value of feature rules, and transforming the weighted feature rules into a structured fault feature map is as follows: A feature rule evaluation system is constructed by introducing three evaluation indicators: stability, physical correlation, and statistical correlation. The stability indicator is support. Confidence level and improvement Physical correlation refers to whether there is a physical connection; statistical correlation refers to the Spearman correlation coefficient. and maximum mutual information coefficient ; By using the entropy weight method for objective evaluation and the analytic hierarchy process (AHP) for subjective evaluation, the sub-weights of each evaluation index in the feature rule evaluation system are weighted to obtain the objective weights of the entropy weight method and the subjective weights of the AHP method. The objective weights of the entropy weight method and the subjective weights of the analytic hierarchy process are added together to obtain the final value of the feature rule. Then, the preceding and following terms of the feature rule are regarded as nodes of the graph model, and the value of the feature rule is used as the weight of the edge to construct the fault feature graph.

[0015] In the step of weighting the various evaluation indicators of the feature rule evaluation system using the objective entropy weight method and the subjective analytic hierarchy process (AHP) to obtain the objective weights of the entropy weight method and the subjective weights of the AHP, ... Entropy weight method calculates the information entropy of each indicator. To determine its weight in the overall evaluation: ; ; in, Information entropy for feature indicators, To ensure the constant for entropy normalization, For the first The sample at the th The proportion of each indicator The total number of samples, For the feature index weights, For the number of indicators; The system constructs a judgment matrix for each layer, comparing each pair of elements within the same layer. It calculates the weight vector of the judgment matrix and verifies its consistency. Finally, it multiplies and sums the weights of the alternative layer relative to the criterion layer with the weights of the criterion layer relative to the target layer to obtain the total weight of the alternative layer relative to the target layer. ; in, For the weight vector, Let I be the eigenvector component corresponding to the largest eigenvalue of the judgment matrix, and let I be the order of the judgment matrix.

[0016] The specific method for constructing a fault association model of the power transmission system by using a graph matching network to mine the correlation matrix between different fault modes based on the constructed fault feature map, and then using a spectral clustering strategy to partition the correlation matrix to obtain the hierarchical structure of the correlation between each fault mode is as follows: The nodes and edges of the fault feature graph are encoded using a graph neural network, and multiple rounds of information propagation are performed in the graph structure. Then, the similarity between graphs is measured by comprehensively calculating node similarity and attention similarity, and the total similarity matrix is ​​obtained. The total similarity matrix is ​​divided based on the spectral clustering strategy, and similar graphs are grouped into the same cluster to construct the standard Laplacian matrix. Based on the constructed standard Laplace matrix, the DB index is used to classify fault modes using K-means clustering, determine the optimal number of fault mode categories, and construct a fault association model.

[0017] The DB index is calculated as follows: ; in, The number of clusters; For the first The average distance within each cluster; For the first The average distance within each cluster; Cluster center and The Euclidean distance between them.

[0018] Secondly, the present invention provides a power transmission system fault correlation analysis system based on feature map matching, comprising: The data acquisition and encoding module is used to acquire various physical quantity signal data of the power transmission system and extract multiple feature indicators. The Relief algorithm is used to optimize the feature indicators and obtain the dominant features. Fuzzy C-means clustering is used to cluster the dominant features. Based on the clustering results, the dominant feature points are discretized. The clustering results, dominant features and discrete intervals are uniformly encoded to construct three-dimensional feature encoding data. The association rule extraction module is used to design an improved fast Apriori algorithm based on three-dimensional feature encoding data to mine frequently used itemsets of highly correlated bivariate variables and extract association rules between multiple physical quantity signals in the power transmission system. The fault feature map construction module is used to construct a multi-index entropy-hierarchical analysis weighted feature rule evaluation system based on the correlation rules between multiple physical quantity signals extracted in the power transmission system. It integrates stability, physical correlation and statistical correlation, quantifies the value of feature rules, and transforms the weighted feature rules into a structured fault feature map. The fault association model construction module is used to mine the association matrix between different fault modes based on the constructed fault feature map using a graph matching network, and then use a spectral clustering strategy to divide the association matrix to obtain the association hierarchy structure between each fault mode, thereby constructing a fault association model for the power transmission system. Real-time acquisition of various physical quantity signal data of the power transmission system, and correlation analysis is performed through the constructed fault correlation model.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This method enables three-dimensional feature encoding of multi-physical quantity signal data in powertrain systems. It effectively characterizes the variation patterns of these signals using standardized methods. Then, it extracts association rules between these signals based on the 3D encoded data. The value of these feature rules is quantified from multiple perspectives, including stability, physical correlation, and statistical correlation, transforming them into a structured fault feature map. Finally, a graph matching algorithm is used to mine the fault feature map and construct an associated fault model for the powertrain system. This method provides unified feature extraction and standardized characterization of multi-physical quantity signals in powertrain systems, and uncovers the potential logical relationships between various fault modes, providing prior knowledge for fault diagnosis.

[0020] Furthermore, this invention is based on the multi-physical quantity signals of the power transmission system. These signals not only have different characteristic patterns, but also often exhibit highly coupled characteristics in actual operation, implying complex deep relationships, which helps to uncover the correlation between power transmission system faults.

[0021] Furthermore, the improved fast Apriori algorithm designed in this invention can further accelerate the efficiency of model data processing based on effectively revealing the intrinsic relationship between various feature indicators.

[0022] Furthermore, the multi-index entropy-hierarchical analysis weighted feature rule evaluation method adopted in this invention can comprehensively evaluate feature rules, integrate subjective and objective weights, and make the weight evaluation more scientific and reasonable.

[0023] Furthermore, the comprehensive similarity calculation method proposed in this invention can better achieve feature map matching and improve the accuracy and indicativeness of the association matrix.

[0024] Furthermore, the association rule model structure proposed in this invention has explicitness and saliency, and can provide effective prior knowledge for fault diagnosis.

[0025] Furthermore, the power transmission system fault correlation analysis method of the present invention is simple, reliable, accurate, reasonable, and easy to implement in engineering applications. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the working principle of the fault correlation analysis model construction method involved in this invention; Figure 2 This is a flowchart illustrating the process flow of the fault correlation analysis method involved in this invention. Figure 3 This is a flowchart illustrating the overall process of the three-dimensional feature encoding method involved in this invention. Figure 4 This invention relates to a feature rule value weight evaluation structure diagram; Figure 5 This is a structural diagram of the fault association model involved in the present invention; Figure 6 This is a feature discrete interval partitioning diagram related to Embodiment 2 of the present invention; Figure 7 This is a fault feature diagram related to Embodiment 2 of the present invention; Figure 8 This refers to the fault correlation matrix involved in Embodiment 2 of the present invention; Figure 9 This is the power transmission system fault association model involved in Embodiment 2 of the present invention. Detailed Implementation

[0027] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0028] Example 1 like Figures 1-2 As shown, a fault correlation analysis method for power transmission systems based on feature map matching includes the following steps: S1: Collect various physical quantity signal data of the power transmission system and extract multiple feature indicators. Use the Relief algorithm to optimize the feature indicators and obtain the dominant features. Use fuzzy C-means clustering to cluster the dominant features. Discretize the dominant feature points according to the clustering results. Unify the clustering results, dominant features and discrete intervals to construct three-dimensional feature coding data. S2: Based on three-dimensional feature encoding data, an improved fast Apriori algorithm is designed to mine highly correlated bivariate frequent itemsets and extract the association rules between multiple physical quantity signals in the power transmission system. S3: Based on the correlation rules between multiple physical quantity signals extracted in the power transmission system, a multi-index entropy-hierarchical analysis weighted feature rule evaluation system is constructed. The system integrates stability, physical correlation and statistical correlation to quantify the value of feature rules and transforms the weighted feature rules into a structured fault feature map. S4: Based on the constructed fault feature map, use graph matching network to mine the correlation matrix between different fault modes, and then use spectral clustering strategy to divide the correlation matrix to obtain the correlation hierarchy structure between each fault mode and construct a fault correlation model of the power transmission system. S5: Real-time acquisition of various physical quantity signal data of the power transmission system, and correlation analysis through the constructed fault correlation model.

[0029] Specifically, in S1, such as Figure 3As shown, various physical quantity signal data of the power transmission system are collected and multiple feature indicators are extracted. The Relief algorithm is used to optimize the feature indicators and obtain the dominant features. Fuzzy C-means clustering is used to cluster the dominant features. Based on the clustering results, the dominant feature points are discretized. The signal type, dominant features and discrete intervals are uniformly encoded to construct three-dimensional feature coding data.

[0030] First, acquire signal data of various physical quantities such as electrical, mechanical, and thermodynamic quantities of the power transmission system, extract time-domain feature indicators, frequency-domain feature indicators, and time-frequency domain feature indicators of different types of signals, and use the Relief algorithm to select and sort the different advantageous features of various types of signals in the power transmission system.

[0031] The feature weight update formula for the Relief algorithm is: ; in, For the first Features at the next iteration Weight value, For the sample In category The Middle The nearest neighbor samples, For category Prior probability, For the sample With its kind Neighbor In features The degree of difference.

[0032] Secondly, fuzzy C-means clustering is used, with membership functions... The degree to which dominant feature points belong to different cluster centers is measured to obtain the clustering result, i.e., the signal type. Based on the clustering result, the dominant feature points are discretized to obtain the discrete intervals corresponding to the dominant feature points: ; in, As the dominant feature point, For the first Cluster centers, This is the fuzzy coefficient (usually between 1.5 and 2.5). The distance between the dominant feature point and the cluster center is the Euclidean distance.

[0033] Finally, the signal type (clustering results), dominant features, and discrete intervals are uniformly encoded to construct three-dimensional feature-coded data.

[0034] Specifically, in S2, based on three-dimensional feature encoding data, an improved fast Apriori algorithm is designed to mine highly correlated bivariate frequent itemsets and extract the association rules between multiple physical quantity signals in the power transmission system.

[0035] First, based on the constructed 3D feature encoding data, a minimum support threshold is set. Candidate item sets are generated starting from a single item, and a hash tree is constructed to store these sets. A preset hash function is used to map item combinations in a transaction to corresponding leaf nodes in the hash tree. Candidate item sets are then matched at these nodes. Upon successful matching, the support count of the candidate item set is updated. ; Secondly, based on the obtained support count, candidate itemsets that are less than the minimum support threshold are pruned to obtain the corresponding frequent itemsets. Then, using the obtained frequent itemsets, higher-order candidate itemsets and frequent itemsets are iteratively generated until no new candidate itemsets can be generated.

[0036] Finally, for each frequent itemset, its non-empty subsets are used as the preceding terms of the rules, and the support data is utilized. Calculate the corresponding confidence level and improvement The correlation rules between multiple physical quantity signals in the powertrain system are selected based on the calculated confidence and lift. The calculation formula is as follows: ;

[0037] in, For itemsets Support For itemsets Support For itemsets Support level.

[0038] Specifically, in S3, such as Figure 4 As shown, a multi-index entropy-hierarchical analysis weighted feature rule evaluation method is proposed, which integrates stability, physical correlation and statistical correlation to quantify the value of feature rules, and transforms the weighted feature rules into a structured fault feature map.

[0039] First, we introduce three evaluation indicators: stability, physical correlation, and statistical correlation, to construct a feature rule evaluation system. The stability indicator is support. Confidence level and improvement Physical correlation refers to whether there is a physical connection; statistical correlation refers to the Spearman correlation coefficient. and maximum mutual information coefficient : ; in, and As characteristic variables, For the sample size, For characteristic variables The The rank of each observation; For characteristic variables The The rank of each observation. for The rank mean, for The rank mean.

[0040] ; ; in, and For the number of grid rows and columns, For mutual information, For data points in the grid The probability, For data points in Edge frequencies on the axis, For data points Edge frequencies on the axis.

[0041] Secondly, by employing the entropy weighting method for objective evaluation and the analytic hierarchy process (AHP) for subjective evaluation, the sub-weights of each evaluation indicator in the feature rule value evaluation system are weighted. The entropy weighting method calculates the information entropy of each indicator. To determine its weight in the overall evaluation: ; ; in, Information entropy for feature indicators, To ensure the constant for entropy normalization, For the first The sample at the th The proportion of each indicator For the feature index weights, This refers to the number of indicators.

[0042] Construct a judgment matrix and compare each pair of elements at the same level; calculate the weight vector of the judgment matrix and verify its consistency; multiply the weights of the alternative layer relative to the criterion layer and the weights of the criterion layer relative to the target layer and sum them to obtain the total weight of the alternative layer relative to the target layer. ; in, For the weight vector, Let I be the eigenvector component corresponding to the largest eigenvalue of the judgment matrix, and let I be the order of the judgment matrix.

[0043] Finally, regarding the objective weights of the entropy weight method Subjective weights in the analytic hierarchy process Add them together to get the final value of the feature rules. Then, treating the preceding and following terms of the feature rule as nodes in a graph model, and using the value of the feature rule as the weight of the edge, a fault feature graph is constructed: ; Specifically, in S4, based on the constructed fault feature map, a graph matching network is used to mine the correlation matrix between different fault modes. Then, a spectral clustering strategy is used to partition the correlation matrix to obtain the hierarchical structure of the correlation between each fault mode, constructing a structure such as... Figure 5 The fault association model of the powertrain system is shown below. First, a graph neural network is used to encode the nodes and edges of the fault feature graph, and multiple rounds of information propagation are performed in the graph structure. Then, the similarity between graphs is measured by comprehensively calculating node similarity and attention similarity.

[0044] The node similarity calculation method is based on three-dimensional feature encoding decomposition, using three-dimensional characters to represent nodes and calculating node similarity: ; in, For nodes, For the first Each character weight, For the first Character similarity, such as ,but ,otherwise .

[0045] Attention similarity calculation methods are based on graph convolutional networks, which extract features and then calculate the graph feature matrix. and The dot product is used to obtain the attention similarity matrix, and the total similarity matrix is ​​obtained by weighted combination of the node similarity matrix and the attention similarity matrix. ; ; in, These are the weight parameters for node similarity; These are the weight parameters for attention similarity.

[0046] Secondly, the correlation matrix (total similarity matrix) is divided based on the spectral clustering strategy, and similar graphs (i.e., those with similar failure modes) are grouped into the same cluster; Constructing the standard Laplace matrix based on the correlation matrix : ; in, It is an incidence matrix. It is a diagonal matrix.

[0047] Finally, using the DB index, K-means clustering is employed to classify fault modes, determine the optimal number of fault mode categories, and construct a fault association model.

[0048] ; in, The number of clusters; For the first The average distance within each cluster; For the first The average distance within each cluster; Cluster center and The Euclidean distance between them.

[0049] Example 2 This embodiment verifies the effectiveness of the invention by combining a gear transmission system dataset containing multiple physical quantity signals.

[0050] Figure 1 This is a flowchart illustrating the working principle of a power transmission system fault association analysis model construction method based on feature map matching proposed in this invention. The associated faults of the gear transmission system will be analyzed according to this flowchart below.

[0051] Figure 2 This is a flowchart of the process flow of a power transmission system fault association analysis method based on feature map matching proposed in this invention. Its main framework includes five parts: three-dimensional feature encoding of multi-physical quantity signals of the power transmission system, extraction of association rules between signals, quantification of feature rule weights, construction of fault feature maps, establishment of fault association models, and fault analysis.

[0052] S1: First, based on the acquisition of various physical quantity signal data such as electrical, mechanical and thermodynamic signals of the power transmission system by sensors, the time domain feature index, frequency domain feature index and time-frequency domain feature index of different types of signals are extracted respectively, and the Relief algorithm is used to select and sort the different advantageous features of various types of signals of the power transmission system.

[0053] The feature weight update formula for the Relief algorithm is: ; in, For the first Features at the next iteration Weight value, For the sample In category The Middle The nearest neighbor samples, For category Prior probability, For the sample With its kind Neighbor In features The degree of difference.

[0054] Secondly, fuzzy C-means clustering is used, with membership functions... Measuring the degree to which dominant feature points belong to different cluster centers: ; in, As the dominant feature point, For the first Cluster centers, This is the fuzzy coefficient (usually between 1.5 and 2.5). The distance between the dominant feature point and the cluster center is the Euclidean distance.

[0055] Based on the clustering results, the dominant feature points are discretized to obtain the discrete intervals corresponding to the dominant feature points, such as... Figure 6 As shown, fuzzy C-means clustering divides continuous signal features into three discrete intervals.

[0056] S2: First, based on the constructed 3D feature encoding data, a minimum support threshold is set. Candidate item sets are generated starting from a single item, and a hash tree is constructed to store these sets. A preset hash function is used to map item combinations in a transaction to corresponding leaf nodes in the hash tree. Candidate item sets are then matched at these nodes. Upon successful matching, the support count of the candidate item set is updated. ; Secondly, based on the obtained support count, candidate itemsets that are less than the minimum support threshold are pruned to obtain the corresponding frequent itemsets. Then, using the obtained frequent itemsets, higher-order candidate itemsets and frequent itemsets are iteratively generated until no new candidate itemsets can be generated.

[0057] Finally, for each frequent itemset, its non-empty subsets are used as the preceding terms of the rules, and the support data is utilized. Calculate the corresponding confidence level and improvement .

[0058] ; ; S3: First, we introduce three evaluation indicators: stability, physical correlation, and statistical correlation, to construct a feature rule evaluation system. The stability indicator is support. Confidence level and improvement Physical correlation refers to whether there is a physical correlation, such as correlation in failure modes, equipment structure, and model measurement; statistical correlation refers to the Spearman correlation coefficient. and maximum mutual information coefficient : ; in, and As characteristic variables, for The rank mean, for The rank mean.

[0059] ; ; in, and As characteristic variables, and For the number of grid rows and columns, For mutual information, For data points in the grid The probability, For data points in Edge frequencies on the axis, For data points Edge frequencies on the axis.

[0060] Secondly, by employing the entropy weighting method for objective evaluation and the analytic hierarchy process (AHP) for subjective evaluation, the sub-weights of each evaluation index for the value of the feature rules are weighted. The entropy weighting method calculates the information entropy of each index. To determine its weight in the overall evaluation: ; ; in, Information entropy for feature indicators, To ensure constantness in entropy normalization, For the first The sample at the th The proportion of each indicator For the feature index weights, This refers to the number of indicators.

[0061] Construct a judgment matrix and compare each pair of elements at the same level; calculate the weight vector of the judgment matrix and verify its consistency; multiply the weights of the alternative layer relative to the criterion layer and the weights of the criterion layer relative to the target layer and sum them to obtain the total weight of the alternative layer relative to the target layer. ; in, For the weight vector, These are the eigenvector components corresponding to the largest eigenvalue of the judgment matrix. The weighted results of the analytic hierarchy process (AHP) are shown in Table 1: Table 1 Weighted Results of Analytic Hierarchy Process (AHP)

[0062] Finally, regarding the objective weights of the entropy weight method Subjective weights in the analytic hierarchy process The summation yields the final value of the feature rules as shown in Table 2. This reflects its overall contribution: ; Table 2 Comprehensive Weight Values ​​of Feature Rules

[0063] Then, treating the preceding and following terms of the feature rule as nodes in a graph model, and using the value of the feature rule as the weight of the edge, a graph is constructed as follows: Figure 7 The fault feature map shown can systematically reveal the structural differences and internal logic of various faults in the feature space, which helps to visualize and distinguish fault types and conduct fault correlation analysis in the future.

[0064] S4: First, use graph neural networks to encode the features of nodes and edges in the fault feature graph, and perform multiple rounds of information propagation in the graph structure. Then, measure the similarity between graphs by comprehensively calculating node similarity and attention similarity.

[0065] The node similarity calculation method is based on three-dimensional feature encoding decomposition, using three-dimensional characters to represent nodes and calculating node similarity: ; in, For nodes, For the first Each character weight, For the first Character similarity, such as ,but ,otherwise .

[0066] Attention similarity calculation methods are based on graph convolutional networks, which extract features and then calculate the graph feature matrix. and The dot product is used to obtain the attention similarity matrix. The node similarity matrix and the attention similarity matrix are combined to obtain the following result: Figure 8 The overall similarity matrix shown is as follows: ; ; in, These are the weight parameters for node similarity; These are the weight parameters for attention similarity.

[0067] Secondly, based on the spectral clustering strategy, a standard Laplace matrix is ​​constructed. : ; in, It is an incidence matrix. It is a diagonal matrix.

[0068] Finally, the DB index was used to classify the failure modes using K-means clustering:

[0069] in, The number of clusters; For the first The average distance within each cluster; For the first The average distance within each cluster; Cluster center and The Euclidean distance between them.

[0070] S5: Fault types were classified and analyzed using spectral clustering based on the correlation matrix of the fault graph model. The number of cluster centers was set to 3 to 6, and the clustering results are shown in Table 3. This result effectively reveals the potential correlation patterns among fault types.

[0071] Table 3 Spectral Clustering Results

[0072] The minimum DB value is found when the number of cluster centers is 5, indicating that the clustering structure is optimal at this point. Further, a clustering model was constructed as follows: Figure 9The powertrain system fault association model shown vividly illustrates the similarity relationships and hierarchical structure among various faults. Based on clustering results, the fault association model not only clearly reflects the aggregation characteristics of different faults in the feature graph space, but also reveals the potential commonalities in signals or mechanisms among certain faults, providing a theoretical basis and visualization support for multi-fault identification.

[0073] Example 3 A powertrain system fault correlation analysis system based on feature map matching, comprising: The data acquisition and encoding module is used to acquire various physical quantity signal data of the power transmission system and extract multiple feature indicators. The Relief algorithm is used to optimize the feature indicators and obtain the dominant features. Fuzzy C-means clustering is used to cluster the dominant features. Based on the clustering results, the dominant feature points are discretized. The clustering results, dominant features and discrete intervals are uniformly encoded to construct three-dimensional feature encoding data. The association rule extraction module is used to design an improved fast Apriori algorithm based on three-dimensional feature encoding data to mine frequently used itemsets of highly correlated bivariate variables and extract association rules between multiple physical quantity signals in the power transmission system. The fault feature map construction module is used to construct a multi-index entropy-hierarchical analysis weighted feature rule evaluation system based on the correlation rules between multiple physical quantity signals extracted in the power transmission system. It integrates stability, physical correlation and statistical correlation, quantifies the value of feature rules, and transforms the weighted feature rules into a structured fault feature map. The fault association model construction module is used to mine the association matrix between different fault modes based on the constructed fault feature map using a graph matching network, and then use a spectral clustering strategy to divide the association matrix to obtain the association hierarchy structure between each fault mode, thereby constructing a fault association model for the power transmission system. The analysis module collects various physical quantity signal data of the power transmission system in real time and performs correlation analysis through the constructed fault correlation model.

[0074] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for fault association analysis of power transmission systems based on feature map matching, characterized in that, Includes the following steps: By collecting electrical, mechanical, and thermal signals from the power transmission system using sensors, time-domain, frequency-domain, and time-frequency-domain feature indices of different signal types are extracted. The Relief algorithm is used to optimize the feature indices and obtain the dominant features. Fuzzy C-means clustering is used to cluster the dominant features, and the dominant features are discretized based on the clustering results. The clustering results, dominant features, and discrete intervals are uniformly encoded to construct three-dimensional feature coding data. Based on three-dimensional feature encoding data, an improved fast Apriori algorithm is designed to mine highly correlated bivariate frequent itemsets and extract the association rules between multiple physical quantity signals in the power transmission system. Based on the correlation rules between multiple physical quantity signals extracted from the power transmission system, a multi-index entropy-hierarchical analysis weighted feature rule evaluation system is constructed. By integrating stability, physical correlation and statistical correlation, the value of feature rules is quantified, and the weighted feature rules are transformed into a structured fault feature map. Based on the constructed fault feature map, a graph matching network is used to mine the correlation matrix between different fault modes. Then, a spectral clustering strategy is used to divide the correlation matrix to obtain the correlation hierarchy structure between each fault mode and construct a fault correlation model for the power transmission system. Real-time acquisition of multi-physical quantity signal data from the power transmission system, followed by correlation analysis using a constructed fault correlation model; The specific method for designing an improved fast Apriori algorithm based on three-dimensional feature encoding data to mine highly correlated bivariate frequent itemsets and extract the association rules between multiple physical quantity signals in the power transmission system is as follows: Based on the constructed three-dimensional feature encoding data, a minimum support threshold is set, a candidate option set is generated starting from a single item, and a hash tree is constructed to store the candidate option set. A preset hash function is used to map the combination of items in the transaction to the corresponding leaf nodes in the hash tree, and the candidate option set is matched on these nodes. When a match is successful, the support count of the candidate option set is updated. Based on the obtained support count, the candidate itemsets that are less than the minimum support threshold are pruned to obtain the corresponding frequent itemsets. Using the obtained frequent itemsets, higher-order candidate itemsets and frequent itemsets are iteratively generated until no new candidate itemsets can be generated. For each frequent itemset, its non-empty subsets are used as the preceding terms of the rules, utilizing support data. Calculate the corresponding confidence level and improvement Based on the calculated confidence and lift, association rules between multiple physical quantity signals in the power transmission system are selected. The specific method for constructing a fault association model for the power transmission system by using a graph matching network to mine the correlation matrix between different fault modes based on the constructed fault feature map, and then using a spectral clustering strategy to divide the correlation matrix to obtain the hierarchical structure of the correlation between each fault mode is as follows: The nodes and edges of the fault feature graph are encoded using a graph neural network, and multiple rounds of information propagation are performed in the graph structure. Then, the similarity between graphs is measured by comprehensively calculating node similarity and attention similarity, and the total similarity matrix is ​​obtained. The total similarity matrix is ​​divided based on the spectral clustering strategy, and similar graphs are grouped into the same cluster to construct the standard Laplacian matrix. Based on the constructed standard Laplace matrix, the DB index is used to classify fault modes using K-means clustering, determine the optimal number of fault mode categories, and construct a fault association model.

2. The method for fault association analysis of power transmission systems based on feature map matching according to claim 1, characterized in that, The formulas for calculating the confidence level and lift are as follows: ； in, For itemsets Support For itemsets Support For itemsets Support level.

3. The method for fault association analysis of power transmission systems based on feature map matching according to claim 1, characterized in that, The specific method for constructing a multi-index entropy-hierarchical analysis weighted feature rule evaluation system based on the correlation rules between multiple physical quantity signals extracted from the power transmission system, integrating stability, physical correlation, and statistical correlation to quantify the value of feature rules, and transforming the weighted feature rules into a structured fault feature map is as follows: A feature rule evaluation system is constructed by introducing three evaluation indicators: stability, physical correlation, and statistical correlation. The stability indicator is support. Confidence level and improvement Physical correlation refers to whether there is a physical connection; statistical correlation refers to the Spearman correlation coefficient. and maximum mutual information coefficient ; By using the entropy weight method for objective evaluation and the analytic hierarchy process (AHP) for subjective evaluation, the sub-weights of each evaluation index in the feature rule evaluation system are weighted to obtain the objective weights of the entropy weight method and the subjective weights of the AHP method. The objective weights of the entropy weight method and the subjective weights of the analytic hierarchy process are added together to obtain the final value of the feature rule. Then, the preceding and following terms of the feature rule are regarded as nodes of the graph model, and the value of the feature rule is used as the weight of the edge to construct the fault feature graph.

4. The method for fault association analysis of power transmission systems based on feature map matching according to claim 3, characterized in that, In the step of weighting the various evaluation indicators of the feature rule evaluation system using the objective entropy weight method and the subjective analytic hierarchy process (AHP) to obtain the objective weights of the entropy weight method and the subjective weights of the AHP, ... Entropy weight method calculates the information entropy of each indicator. To determine its weight in the overall evaluation: ； ； in, Information entropy for feature indicators, To ensure the constant for entropy normalization, For the first The sample at the th The proportion of each indicator The total number of samples, For the feature index weights, For the number of indicators; The Analytic Hierarchy Process (AHP) is used for evaluating the characteristic rule weights of power transmission systems, decomposing the decision problem into an objective layer, a criterion layer, and a solution layer. Construct a judgment matrix and compare each pair of elements at the same level; calculate the weight vector of the judgment matrix and verify its consistency; multiply the weights of the alternative layer relative to the criterion layer and the weights of the criterion layer relative to the target layer and sum them to obtain the total weight of the alternative layer relative to the target layer. ； in, For the weight vector, Let I be the eigenvector component corresponding to the largest eigenvalue of the judgment matrix, and let I be the order of the judgment matrix.

5. A powertrain system fault correlation analysis system based on feature graph matching, based on the powertrain system fault correlation analysis method based on feature graph matching as described in any one of claims 1 to 4, characterized in that, include: The data acquisition and encoding module is used to collect electrical, mechanical, and thermal multi-physical quantity signal data of the power transmission system through sensors, extract time-domain feature indicators, frequency-domain feature indicators, and time-frequency domain feature indicators of different types of signals, use the Relief algorithm to optimize the feature indicators and obtain the dominant features; use fuzzy C-means clustering to cluster the dominant features, discretize the dominant feature points according to the clustering results, and uniformly encode the clustering results, dominant features, and discrete intervals to construct three-dimensional feature encoding data; The association rule extraction module is used to design an improved fast Apriori algorithm based on three-dimensional feature encoding data to mine frequently used itemsets of highly correlated bivariate variables and extract association rules between multiple physical quantity signals in the power transmission system. The fault feature map construction module is used to construct a multi-index entropy-hierarchical analysis weighted feature rule evaluation system based on the correlation rules between multiple physical quantity signals extracted in the power transmission system. It integrates stability, physical correlation and statistical correlation, quantifies the value of feature rules, and transforms the weighted feature rules into a structured fault feature map. The fault association model construction module is used to mine the association matrix between different fault modes based on the constructed fault feature map using a graph matching network, and then use a spectral clustering strategy to divide the association matrix to obtain the association hierarchy structure between each fault mode, thereby constructing a fault association model for the power transmission system. The analysis module collects multi-physical signal data of the power transmission system in real time and performs correlation analysis through the constructed fault correlation model.

Images