A fault diagnosis method based on a multi-scale chebyshev convolution fusion network

Abstract

This invention provides a fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network, belonging to the field of fault diagnosis technology. First, a system graph data based on relational attributes and physical connections between components is constructed. Second, using the system graph data as input, a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network is constructed. Third, the fault diagnosis model is trained to obtain a trained fault diagnosis model. Finally, the trained fault diagnosis model based on the multi-scale Chebyshev convolutional fusion network is used to diagnose typical system faults under conditions of random data loss. This invention enables the synergistic utilization of physical connection data and monitoring data, effectively achieving the fusion of multi-scale domain information and compensating for the decreased accuracy of traditional fault diagnosis due to data loss.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology and relates to a fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network. In particular, it relates to a fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network for the conditions of missing random node data and missing fault node data in typical systems. Background Technology

[0002] To address the challenges of analysis and diagnosis under conditions of missing data, scholars both domestically and internationally have conducted extensive research and proposed several solutions. For example, Chinese invention patent CN119471181B provides a smart grid fault early warning and diagnosis system that uses graph convolutional networks to model the power grid topology to achieve fault location and diagnosis. Chinese invention patent CN110212528B provides a method for reconstructing missing distribution network measurement data based on generative adversarial networks and dual semantic awareness. This method uses generative adversarial networks (GANs) to learn data features through game theory, reconstructing the missing data before analysis.

[0003] While existing technical solutions can address data loss or perform fault diagnosis to some extent, they all have significant limitations: the diagnostic method based on graph convolutional networks proposed in Chinese invention patent CN119471181B typically uses a single or fixed-scale convolutional kernel, failing to weightedly fuse multi-level neighborhood information. When data is missing at fault nodes, the receptive field is limited and the information compensation capability is insufficient, leading to decreased diagnostic stability. On the other hand, the two-stage method of "reconstruction first, then diagnosis" proposed in Chinese invention patent CN110212528B faces significant reconstruction difficulties and high errors when key fault nodes are missing. Furthermore, the reconstruction errors accumulate in subsequent diagnostic stages, and there is a lack of a mechanism for multi-scale information compensation using the system's physical topology. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network. This method can effectively utilize prior knowledge such as the physical connection relationships, fault mechanisms, and topological structures of typical systems to construct system graph data. Furthermore, by using a multi-scale Chebyshev convolutional fusion network to weighted aggregate information from multiple neighborhood nodes, it achieves efficient extraction of global fault features and accurate pattern recognition even under conditions of missing random node data and missing fault node data, significantly improving the diagnostic accuracy and robustness under data loss conditions.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network includes the following steps: Step 1: Construct system diagram data based on relational attributes and physical connection attributes between components; specifically: The system graph data method based on relational attributes treats each component of a typical system as a vertex and establishes an initial relation matrix through the physical connection attributes between components to represent the relationships between them. Considering that indirect influences exist even between components without direct physical connections, relational attributes are calculated between components to uncover the relationships between indirectly connected components. These attributes are then used to update the relation matrix, representing the indirect relationships between components in the typical system. Since the indirect influence between indirectly connected components diminishes to negligible levels, for all components, the relational information between a component and its neighboring components within a certain order is selected. A new weighted relational matrix is ​​calculated using a first-order Gaussian kernel function to represent the propagation of fault information among components in the typical system. Finally, the system graph data is constructed, denoted as […]. ,in, These consist of a vertex set, an edge set, a weighted relation matrix, and a fault feature matrix, respectively. The construction process is as follows: Step 1.1: Using each component of a typical system as a vertex, denote it as the vertex set. ; Step 1.2: Establish an initial relationship matrix based on the physical connection attributes of each component, denoted as... Where R represents a matrix and n represents the number of vertices; (definition) As vertices i and vertex j The physical connection attribute value, and ; Step 1.3, Determine the physical connection attribute value The correlation between the size and indirect relationship of components is determined by using a first-order Gaussian kernel function to calculate the relationship attributes and update the initial relationship matrix, resulting in an updated weighted relationship matrix, denoted as . . Specifically: Step 1.3.1, Physical connection attribute value under negative correlation conditions The smaller the value, the greater the weight of the indirect association between components. Combined with Dijkstra's algorithm, the relationship attribute is the shortest distance between each vertex and the physical connection attributes of all other vertices. Specifically: Step 1.3.1.1: Calculate the relation attributes using Dijkstra's algorithm. The relation attributes are the shortest distances between each vertex and all other vertices in terms of physical connectivity. Obtain the updated relation matrix using these attributes. ; Step 1.3.1.2: Sort the relational attributes of each vertex in ascending order, and take the first few vertices of vertex x. The set of neighboring nodes is denoted as _____. and each vertex The minimum relational attribute composition of the order neighbor nodes Order submatrix ; Step 1.3.1.3: Calculate the kernel radius using a first-order Gaussian kernel function. : ,in, Indicates the distance from vertex i to... The distance between neighboring nodes, where i represents the vertex index and n represents the total number of vertices; Step 1.3.1.4: Calculate the weighted relation matrix for each vertex. : like , ; like , ; in, This represents the distance between vertex i and vertex j; The attribute value represents the relationship between vertex i and vertex j; Step 1.3.1.5: If the weighted relation matrix obtained in step 1.3.1.3 is an asymmetric matrix, then convert it into a symmetric matrix. Specifically, if... and ,but .

[0006] Step 1.3.2, Physical connection attribute values ​​under positive correlation conditions The larger the value, the greater the weight of the indirect association between components. The relationship attribute is the reciprocal of the shortest distance between each vertex and the physical connection attributes of all vertices. Specifically: Step 1.3.2.1: Calculate the shortest distance of physical connectivity between each vertex and all other vertices using Dijkstra's algorithm. Based on these shortest distances, calculate the relation attributes, which are the reciprocals of the shortest distances of the physical connectivity attributes. Use these relation attributes to obtain the updated relation matrix. ; Step 1.3.2.2: Sort the relational attributes of each vertex from smallest to largest, and select the vertex. x The former The set of neighboring nodes is denoted as _____. and each vertex The minimum relational attribute composition of the order neighbor nodes Order submatrix .

[0007] Step 1.3.2.3, Calculate the kernel radius : ; Step 1.3.2.4: Calculate the weighted relation matrix for each vertex. : like , ; like , ; Step 1.3.2.5: If the weighted relation matrix is ​​asymmetric, then convert it to a symmetric matrix. Specifically, if... and ,but .

[0008] Step 1.4: Construct an edge set E based on the weighted relation matrix obtained in Step 1.3, denoted as E. ,in, This represents the first edge. Indicates the second edge. Indicates the m-th edge; when When, the edge exists, when At that time, the edge does not exist.

[0009] Step 1.5: Obtain monitoring data by monitoring or simulating typical systems, and construct a fault feature matrix based on the monitoring data of each node. ,in, This represents the fault characteristic data of the nth vertex; Step 1.6, the completed system diagram data is denoted as... Where V is the set of vertices, each vertex representing a system component; E is the set of edges, each edge representing the relationship between two components; and A is the weighted relation matrix, i.e., the numerical values ​​of the association relationships between components. This is the fault characteristic matrix, which consists of the monitoring parameters of each component.

[0010] The system diagram data enables the modeling of indirect relationships between typical system components, effectively representing the information propagation relationships between components. The larger the relationship value, the stronger the indirect relationship and information propagation relationship between components.

[0011] Step 2: Using the system graph data constructed in Step 1 as input, build a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network, which includes an input module, a multi-scale Chebyshev convolutional fusion module, a fully connected module, and a graph classification module; specifically: To address the conditions of random node missingness and fault node missingness, a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network is constructed. This model utilizes multi-scale Chebyshev graph convolutional kernels to learn the feature information of multiple neighboring nodes. The multi-scale neighboring node information of missing data nodes is weighted and aggregated to obtain fault information of the missing nodes. Furthermore, global fault features are extracted to achieve fault pattern recognition under conditions of random data missingness. The model construction process is as follows: Step 2.1, Input module: The input module receives the completed system diagram data, where the fault feature matrix of the monitoring data is denoted as... ,in n The number of vertices. This refers to the fault characteristic dimension of the monitoring data.

[0012] Step 2.2, Multi-scale Chebyshev Convolution Fusion Module: The multi-scale Chebyshev convolution fusion module includes multiple Chebyshev graph convolution kernels. Specifically, the degree of the Chebyshev polynomial in each Chebyshev convolution... k The size of the convolution kernel determines the size of the Chebyshev graph convolution, indicating that the convolution of the central node will aggregate the values ​​of the convolution kernel. k The feature information of the neighborhood is limited. Traditional Chebyshev graph convolution kernels typically only include one scale, thus restricting the scope of neighborhood node feature information learned by graph convolution and limiting the learning process. Specifically: The multi-scale Chebyshev convolutional fusion module is constructed by stacking multiple multi-scale Chebyshev convolutional fusion layers, each layer including Chebyshev graph convolution kernels at multiple scales. Multi-scale Chebyshev graph convolution operations are performed in each layer, weighted and aggregated fault feature information from different scale neighborhoods. The final output is a two-dimensional enhanced feature matrix that fuses fault information from multiple scale neighborhoods. ,in n The number of vertices. To enhance the feature dimension of the feature matrix.

[0013] Single-layer multi-scale Chebyshev convolutional fusion layer weighted aggregation i Chebyshev plot convolution kernels of different scales. j The multi-scale Chebyshev convolutional fusion layer includes i Each Chebyshev plot convolution kernel has a corresponding kernel. The Chebyshev plot convolution kernel scales are respectively ,different The Chebyshev plot convolution kernel propagates fault information in different ranges of the neighborhood.

[0014] Regarding the first j The input of the multi-scale Chebyshev convolutional fusion layer is processed by Chebyshev graph convolution kernels. Then, its output is recorded as We utilize weighted aggregation to achieve weighted aggregation of Chebyshev diagram convolution kernels at different scales, thereby obtaining enhanced fault features that fuse neighborhood information at different scales, denoted as... The weighted aggregation process is as follows: (1) in, The weighted aggregation operator controls the fusion weights of Chebyshev graph convolution kernels at different scales.

[0015] Step 2.3, Fully Connected Module: The fully connected module expands the obtained enhanced feature matrix into a one-dimensional fault feature vector, and then extracts global fault features through a fully connected layer. Specifically, the fully connected module converts the two-dimensional enhanced feature matrix output by the multi-scale Chebyshev convolution fusion module into a one-dimensional fault feature vector. Expanded into a one-dimensional fault feature vector, the length of the one-dimensional fault feature vector is After processing through multiple fully connected layers, global fault characteristics of a typical system are extracted.

[0016] Step 2.4, Graph Classification Module: The global fault features are fed into the graph classification module, which uses the Log-Softmax classifier to output the fault modes of typical systems.

[0017] Step 3: Train the fault diagnosis model constructed in Step 2 to obtain the trained fault diagnosis model; specifically: A training set is established based on typical system data without missing data to train a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network. Specifically: monitoring data without missing data from typical systems is collected to establish a training set. The monitoring data is obtained through simulation or data acquisition of typical systems. System graph data without missing data is constructed to train the fault diagnosis model based on the multi-scale Chebyshev convolutional fusion network.

[0018] Step 4: Based on the fault diagnosis model trained in Step 3, perform fault diagnosis on typical systems under conditions of random data loss; specifically: Random data loss includes missing random nodes and missing faulty nodes, as defined below: (1) Random node missing; Random node missing refers to the absence of all sensor measurement data points of random components in a typical system. The randomness of this data missing situation is reflected in the randomness of the system components and their number. Therefore, the impact of random node missing data conditions on the accuracy of fault diagnosis is also random.

[0019] (2) The fault node is missing; Missing fault nodes refer to the absence of all sensor measurement data for a component in a typical system that has failed. This data loss signifies a loss of fault information, significantly impacting the accuracy of fault diagnosis. Furthermore, missing fault nodes also include the absence of sensor measurement data for all random components other than the failed component. This situation, on top of the loss of fault information, further increases the randomness of the impact on fault diagnosis accuracy.

[0020] To address the issues of missing data at random nodes and missing data at faulty nodes, a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network, trained beforehand, is used to diagnose typical system faults. Specifically, during the testing phase, a test set is established to accommodate these conditions. System graph data under missing data conditions is constructed. The trained fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network feeds this missing data into the input module. The multi-scale Chebyshev convolutional fusion module obtains a two-dimensional enhanced feature matrix under these conditions. This matrix is ​​then processed by a fully connected module to extract global fault features under these missing data conditions. Finally, a graph classification module is used to diagnose typical system faults under these missing data conditions.

[0021] The present invention has the following beneficial effects: (1) This invention proposes a fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network. The system graph data is constructed based on the physical connection attributes of typical systems, and a fault diagnosis model based on multi-scale Chebyshev is established. The collaborative utilization of physical connection data and monitoring data is realized, providing a new solution for fault diagnosis under the conditions of missing random node data and missing fault node data.

[0022] (2) This invention proposes a system graph data construction based on relational attributes. Based on the physical connection attributes of typical systems, relational attributes are calculated for the correlation between physical connection attributes and indirect relationships between components. This realizes the modeling of indirect relationships between components of typical systems, which can effectively characterize the information propagation relationship between components and provide support for subsequent diagnosis.

[0023] (3) The present invention constructs a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network, which weights and aggregates multi-scale domain information, effectively realizing the fusion of multi-scale domain information and making up for the shortcomings of traditional fault diagnosis due to the decrease in fault diagnosis accuracy caused by data loss.

[0024] (4) Through simulation data verification, compared with traditional fault diagnosis methods, the method proposed in this invention can effectively improve the fault diagnosis accuracy under data missing conditions. Attached Figure Description

[0025] Figure 1 This is a flowchart of the present invention; Figure 2 Build visualization results for power grid system diagram data. Detailed Implementation

[0026] The present invention will be further described below with reference to specific implementation examples.

[0027] A fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network includes the following steps: Step 1: Construct system diagram data based on relational attributes and physical connection attributes between components; specifically: In this specific embodiment, fault simulation and data generation are conducted using an IEEE 123-node distribution network as the object. IEEE 123 is a 123-node distribution network simulation system provided by IEEE. This distribution network model has a voltage level of 4.16kV and its structure consists of overhead lines, distribution transformers, disconnecting switches, parallel capacitor banks, and other auxiliary facilities, and is equipped with three-phase balanced or unbalanced loads as well as single-phase loads. The distribution network contains 123 buses, increasing the complexity of the distribution network model and the scope of fault impact. The IEEE 123 system power grid model is built and simulated using Simulink software.

[0028] For three common fault modes in power distribution networks, including single-phase grounding, two-phase grounding, and two-phase short circuit, the Three-PhaseFault module in Simulink is used to inject faults according to the principles of each fault mode and collect simulation data.

[0029] Based on the three typical fault modes mentioned above, including single-phase grounding, two-phase grounding, and two-phase short circuit fault injection methods, model simulations were completed with a sampling frequency of 10 kHz and a sampling time of 0.1 s. Simulations were performed under normal conditions and under a total of 260 modes, including 123 nodes with a-phase grounding, 3 two-phase branch line nodes with a / b-phase grounding and a / b-phase short circuits, and 65 three-phase main line nodes with b / c-phase grounding and a / b-phase short circuits. Details are shown in Table 1.

[0030] Table 1: Details of 260 Fault Simulations

[0031] To make the ideal voltage source in the simulation model closer to the actual voltage in engineering practice, this embodiment takes 200 data points from the stable period of each simulation data set, adds random Gaussian white noise with an SNR of 45, and uses a moving average with a window length of 10 and a step size of 10 to extract 10 data features as 20 sample points. Taking node 1 as an example, under four states—normal, phase a ground, phase b / c ground, and phase a / b short circuit—the six parameter data of this node after random noise addition are ( , , , , , )like Figure 2 As shown.

[0032] The system graph data method based on relational attributes treats each component of a typical system as a vertex and establishes an initial relation matrix through the physical connection attributes between components to represent the relationships between them. Considering that indirect influences exist even between components without direct physical connections, relational attributes are calculated between components to uncover the relationships between indirectly connected components. These attributes are then used to update the relation matrix, representing the indirect relationships between components in the typical system. Since the indirect influence between indirectly connected components diminishes to negligible levels, the relational information between a component and its neighboring components within a certain order is selected. A new weighted relational matrix is ​​calculated using a first-order Gaussian kernel function to represent the propagation of fault information among components in the typical system. Finally, the system graph data is constructed, denoted as […]. ,in, These consist of a vertex set, an edge set, a weighted relation matrix, and a fault feature matrix, respectively. The construction process is as follows: Step 1.1: Using each component of a typical system as a vertex, denote it as the vertex set. In this embodiment, the 123 bus nodes in the power grid system are treated as 123 vertices, denoted as the vertex set. .

[0033] Step 1.2: Establish an initial relationship matrix based on the physical connection attributes of each component, denoted as... Where R represents matrix representation and n represents the number of vertices; definition As vertices i and vertex j The physical connection attribute value, and ; Step 1.3, Determine the physical connection attribute value The correlation between the size and indirect relationship of components is determined by using a first-order Gaussian kernel function to calculate the relationship attributes and update the initial relationship matrix, resulting in an updated weighted relationship matrix, denoted as . . Specifically: Step 1.3.1, Physical connection attribute value under negative correlation conditions The smaller the value, the greater the weight of the indirect association between components. Combined with Dijkstra's algorithm, the relationship attribute is the shortest distance between each vertex and the physical connection attributes of all other vertices. Specifically: Step 1.3.1.1: Calculate the relation attributes using Dijkstra's algorithm. The relation attributes are the shortest distances between each vertex and all other vertices in terms of physical connectivity. Obtain the updated relation matrix using these attributes. ; Step 1.3.1.2: Sort the relational attributes of each vertex from smallest to largest, and select the vertex. The former The set of neighboring nodes is denoted as _____. and each vertex The minimum relational attribute composition of the order neighbor nodes Order submatrix In a specific embodiment, take =20.

[0034] Step 1.3.1.3: Calculate the kernel radius using a first-order Gaussian kernel function. : ,in, Indicates the distance from vertex i to... The distance between neighboring nodes, where i represents the vertex index and n represents the total number of vertices; Step 1.3.1.4: Calculate the weighted relation matrix for each vertex. : like , ; like , ; in, This represents the distance between vertex i and vertex j; The attribute value represents the relationship between vertex i and vertex j; Step 1.3.1.5: If the weighted relation matrix obtained in step 1.3.1.3 is an asymmetric matrix, then convert it into a symmetric matrix. Specifically, if... and ,but .

[0035] Step 1.3.2, Physical connection attribute values ​​under positive correlation conditions The larger the value, the greater the weight of the indirect association between components. The relationship attribute is the reciprocal of the shortest distance between each vertex and the physical connection attributes of all vertices. Specifically: Step 1.3.2.1: Calculate the shortest distance of physical connectivity between each vertex and all other vertices using Dijkstra's algorithm. Based on these shortest distances, calculate the relation attributes, which are the reciprocals of the shortest distances of the physical connectivity attributes. Use these relation attributes to obtain the updated relation matrix. ; Step 1.3.2.2: Sort the relational attributes of each vertex from smallest to largest, and select the vertex. x The former The set of neighboring nodes is denoted as _____. and each vertex The minimum relational attribute composition of the order neighbor nodes Order submatrix .

[0036] Step 1.3.2.3, Calculate the kernel radius : ; Step 1.3.2.4: Calculate the weighted relation matrix for each vertex. : like , ; like , ; Step 1.3.2.5: If the weighted relation matrix is ​​asymmetric, then convert it to a symmetric matrix. Specifically, if... and ,but ; Step 1.4: Construct an edge set E based on the weighted relation matrix obtained in Step 1.3, denoted as E. ,in, This represents the first edge. Indicates the second edge. Indicates the m-th edge; when When, the edge exists, when At that time, the edge does not exist.

[0037] Step 1.5: Obtain monitoring data by monitoring or simulating typical systems, and construct a fault feature matrix based on the monitoring data of each node. ,in, This represents the fault characteristic data of the nth vertex; Step 1.6, the completed system diagram data is denoted as... Where V is the set of vertices, each vertex representing a system component; E is the set of edges, each edge representing the relationship between two components; and A is the weighted relation matrix, i.e., the numerical values ​​of the association relationships between components. This is the fault characteristic matrix, which consists of the monitoring parameters of each component.

[0038] The system diagram data enables the modeling of indirect relationships between typical system components, effectively representing the information propagation relationships between components. The larger the relationship value, the stronger the indirect relationship and information propagation relationship between components.

[0039] Step 2: Using the system graph data constructed in Step 1 as input, build a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network, which includes an input module, a multi-scale Chebyshev convolutional fusion module, a fully connected module, and a graph classification module; specifically: To address the conditions of random node missingness and fault node missingness, a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network is constructed. This model utilizes multi-scale Chebyshev graph convolutional kernels to learn the feature information of multiple neighboring nodes. The multi-scale neighboring node information of missing data nodes is weighted and aggregated to obtain fault information of the missing nodes. Furthermore, global fault features are extracted to achieve fault pattern recognition under conditions of random data missingness. The model construction process is as follows: Step 2.1, Input module: The input module receives the completed system diagram data, where the fault feature matrix of the monitoring data is denoted as... ,in n The number of vertices. This refers to the fault characteristic dimension of the monitoring data.

[0040] Step 2.2, Multi-scale Chebyshev Convolution Fusion Module: The multi-scale Chebyshev convolution fusion module includes multiple Chebyshev graph convolution kernels. Specifically, the degree of the Chebyshev polynomial in each Chebyshev convolution... k The size of the convolution kernel determines the size of the Chebyshev graph convolution, indicating that the convolution of the central node will aggregate the values ​​of the convolution kernel. k The feature information of the neighborhood is limited. Traditional Chebyshev graph convolution kernels typically only include one scale, thus restricting the scope of neighborhood node feature information learned by graph convolution and limiting the learning process. Specifically: The multi-scale Chebyshev convolutional fusion module is constructed by stacking multiple multi-scale Chebyshev convolutional fusion layers, each layer including Chebyshev graph convolution kernels at multiple scales. Multi-scale Chebyshev graph convolution operations are performed in each layer, weighted and aggregated fault feature information from different scale neighborhoods. The final output is a two-dimensional enhanced feature matrix that fuses fault information from multiple scale neighborhoods. ,in n The number of vertices. To enhance the feature dimension of the feature matrix.

[0041] Single-layer multi-scale Chebyshev convolutional fusion layer weighted aggregation i Chebyshev plot convolution kernels of different scales. j The multi-scale Chebyshev convolutional fusion layer includes i Each Chebyshev plot convolution kernel has a corresponding kernel. The Chebyshev plot convolution kernel scales are respectively ,different The Chebyshev plot convolution kernel propagates fault information in different ranges of the neighborhood.

[0042] Regarding the first j The input of the multi-scale Chebyshev convolutional fusion layer is processed by Chebyshev graph convolution kernels. Then, its output is recorded as We utilize weighted aggregation to achieve weighted aggregation of Chebyshev diagram convolution kernels at different scales, thereby obtaining enhanced fault features that fuse neighborhood information at different scales, denoted as... The weighted aggregation process is as follows: (1) in, The weighted aggregation operator controls the fusion weights of Chebyshev graph convolution kernels at different scales.

[0043] Step 2.3, Fully Connected Module: The fully connected module expands the obtained enhanced feature matrix into a one-dimensional fault feature vector, and then extracts global fault features through a fully connected layer. Specifically, the fully connected module converts the two-dimensional enhanced feature matrix output by the multi-scale Chebyshev convolution fusion module into a one-dimensional fault feature vector. Expanded into a one-dimensional fault feature vector, the length of the one-dimensional fault feature vector is After processing through multiple fully connected layers, global fault characteristics of a typical system are extracted.

[0044] Step 2.4, Graph Classification Module: Global fault features are fed into the graph classification module, which uses a Log-Softmax classifier to output the fault modes of typical systems. In this embodiment, the multi-scale Chebyshev convolutional fusion module includes three multi-scale Chebyshev convolutional fusion layers, denoted as follows: , , Each multi-scale Chebyshev convolutional fusion layer consists of three Chebyshev graph convolutional kernels with scales K of 3, 4, and 5, and a total of 64 kernels. The weighted aggregation operator of the weighted aggregation function h is also described. The fully connected module consists of three fully connected layers, with 2048, 1024, and 512 neurons in each layer, respectively. The graph classification module has 260 neurons. The fault diagnosis algorithm model structure parameters are set as follows: Figure 2 As shown.

[0045] Table 2: Fault Diagnosis Algorithm Model Structure Parameter Settings Based on MOF-ChebyNet

[0046] Step 3: Train the fault diagnosis model constructed in Step 2 to obtain the trained fault diagnosis model; specifically: A training set is established based on typical system data without missing data to train a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network. Specifically: monitoring data without missing data from typical systems is collected to establish a training set. System graph data without missing data is constructed to train the fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network.

[0047] In this embodiment, the simulation data for 260 fault modes (including the normal state) undergoes the above processing and then normalization. Each fault mode generates 20 sample points, totaling 5200 sample points. A training set is established based on typical system data without missing data, using 2600 samples as the training set to generate a label set with 260 corresponding categories. The normal state is treated as one label, denoted as label=0; each fault mode is treated as one fault label, denoted as label=1,2,3,…,259 respectively.

[0048] Step 4: Based on the fault diagnosis model trained in Step 3, perform fault diagnosis on typical systems under conditions of random data loss; specifically: Random data loss includes missing random nodes and missing faulty nodes, as defined below: (1) Random node missing; Random node missing refers to the absence of all sensor measurement data points of random components in a typical system. The randomness of this data missing situation is reflected in the randomness of the system components and their number. Therefore, the impact of random node missing data conditions on the accuracy of fault diagnosis is also random.

[0049] (2) The fault node is missing; Missing fault nodes refer to the absence of all sensor measurement data for a component in a typical system that has failed. This data loss signifies a loss of fault information, significantly impacting the accuracy of fault diagnosis. Furthermore, missing fault nodes also include the absence of sensor measurement data for all random components other than the failed component. This situation, on top of the loss of fault information, further increases the randomness of the impact on fault diagnosis accuracy.

[0050] To address the issues of missing data at random nodes and missing data at faulty nodes, a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network, trained beforehand, is used to diagnose typical system faults. Specifically, during the testing phase, a test set is established to accommodate these conditions. System graph data under missing data conditions is constructed. The trained fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network feeds this missing data into the input module. The multi-scale Chebyshev convolutional fusion module obtains a two-dimensional enhanced feature matrix under these conditions. This matrix is ​​then processed by a fully connected module to extract global fault features under these missing data conditions. Finally, a graph classification module is used to diagnose typical system faults under these missing data conditions.

[0051] In this embodiment, test sets are constructed for the conditions of missing data at random nodes and missing data at faulty nodes, as shown in Table 3. The other 2600 samples are subjected to six types of random missing data processing to generate six test sets, denoted as ModeA~F respectively.

[0052] Table 3: Random Missing Data in 6 Categories

[0053] The method was tested on both the training set with no missing data and the test sets under six different missing data conditions. The fault diagnosis accuracy was calculated by dividing the number of correctly diagnosed samples by the total number of samples in the test set. Considering the randomness of missing data, 100 tests were performed on each test set to fully validate the method's effectiveness. The final average diagnostic accuracy is shown in Table 5. Specifically, in the 100 tests conducted on the ModeA, ModeC, ModeD, ModeE, and ModeF test sets, the randomly missing nodes were varied in each test, ensuring the randomness of the experiment.

[0054] Table 4: Fault diagnosis results of the IEEE123 system based on MOF-ChebyNet

[0055] The fault diagnosis results are shown in Table 4. For the case where each sample is missing one random node with six parameter values ​​(Mode A), the model achieved a fault diagnosis accuracy of 94.96%, compared to 97.12% accuracy under the condition of no random data loss (None). This preliminarily verifies the model's fault diagnosis performance under the condition of random data loss. However, when the sample is exactly missing all six parameter values ​​of the fault node (Mode B), the model's diagnostic accuracy drops to 89.73%. This shows that missing information about the fault node affects the fault diagnosis effect. Nevertheless, the proposed method still achieves nearly 90% diagnostic accuracy, demonstrating its advantage in using graph representation to represent fault propagation information. As the number of missing nodes increases, the diagnostic difficulty increases, and the model's performance gradually declines. However, when three or fewer nodes are missing, the model consistently demonstrates good diagnostic performance.

[0056] Furthermore, when information about the faulty node is missing, fault diagnosis is extremely difficult. The proposed diagnostic algorithm model can still accurately identify the faulty node and its fault mode. The above specific implementation methods further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above are merely specific implementation methods of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made based on the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A fault diagnosis method based on a multi-scale Chebyshev convolution fusion network, characterized in that, The fault diagnosis method includes the following steps: Step 1: Construct system diagram data based on relational attributes and physical connection attributes between components; The system graph data method based on the relationship attribute takes each component of a typical system as a vertex, and establishes an initial relationship matrix through the physical connection attribute between components to represent the relationship between components. For all components, the relationship information between a component and its neighbor components within a certain order is selected, and a new weighted relationship matrix is calculated using a first-order Gaussian kernel function. The system graph data construction is completed, denoted as wherein, , , , are a vertex set, an edge set, a weighted relationship matrix, and a fault feature matrix, respectively. Step 2: Using the system graph data constructed in Step 1 as input, construct a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network, which includes an input module, a multi-scale Chebyshev convolutional fusion module, a fully connected module, and a graph classification module. To address the conditions of random node missing and fault node missing, a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network is constructed. The model utilizes multi-scale Chebyshev graph convolutional kernels to learn the feature information of multiple neighboring nodes, performs weighted aggregation of the multi-scale neighboring node information of missing data nodes, obtains the fault information of missing nodes, extracts global fault features, and realizes fault mode recognition under the condition of random data missing. Step 3: Train the fault diagnosis model constructed in Step 2 to obtain the trained fault diagnosis model; specifically: Training sets are established based on typical system data without missing data to complete the training of a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network; monitoring data without missing data from typical systems are collected to establish a training set; system graph data without missing data is constructed to complete the training of a fault diagnosis model based on a multi-scale Chebyshev convolutional fusion network. Step 4: Using the fault diagnosis model trained in Step 3, perform fault diagnosis on typical systems under conditions of random data loss.

2. The fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network according to claim 1, characterized in that, In the first step, the process of constructing the system diagram data is as follows: Step 1.1: Using each component of a typical system as a vertex, denote it as the vertex set. ; Step 1.2: Establish an initial relationship matrix based on the physical connection attributes of each component, denoted as... , where R represents a matrix and n represents the number of vertices; definition As vertices i and vertex j The physical connection attribute value, and ; Step 1.3, Determine the physical connection attribute value The correlation between the size and indirect relationship of components is determined by using a first-order Gaussian kernel function to calculate the relationship attributes and update the initial relationship matrix, resulting in an updated weighted relationship matrix, denoted as . ; Step 1.4: Construct an edge set E based on the weighted relation matrix obtained in Step 1.3, denoted as E. ,in, This represents the first edge. Indicates the second edge. Indicates the m-th edge; when When, the edge exists, when At that time, the edge does not exist; Step 1.5: Obtain monitoring data by monitoring or simulating typical systems, and construct a fault feature matrix based on the monitoring data of each node. ,in, This represents the fault characteristic data of the nth vertex; Step 1.6, the completed system diagram data is denoted as... Where V is the set of vertices, each vertex representing a system component; E is the set of edges, each edge representing the relationship between two components; and A is the weighted relation matrix, i.e., the numerical values ​​of the association relationships between components. This is the fault characteristic matrix, which consists of the monitoring parameters of each component; The system diagram data enables the modeling of indirect relationships between typical system components and characterizes the information propagation relationships between components.

3. The fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network according to claim 2, characterized in that, Step 1.3 specifically refers to: Step 1.3.1, Physical connection attribute value under negative correlation conditions The smaller the value, the greater the weight of the indirect association between components; combined with Dijkstra's algorithm, the relationship attribute is the shortest distance between each vertex and the physical connection attributes of all vertices; specifically: Step 1.3.1.1: Calculate the relation attributes using Dijkstra's algorithm. The relation attributes are the shortest distances between each vertex and all other vertices in terms of physical connectivity. Obtain the updated relation matrix using these attributes. ; Step 1.3.1.2: Sort the relational attributes of each vertex in ascending order, and take the first few vertices of vertex x. The set of neighboring nodes is denoted as _____. and each vertex The minimum relational attribute composition of the order neighbor nodes Order submatrix ; Step 1.3.1.3: Calculate the kernel radius using a first-order Gaussian kernel function. : ,in, Indicates the distance from vertex i to... The distance between neighboring nodes, where i represents the vertex index and n represents the total number of vertices; Step 1.3.1.4: Calculate the weighted relation matrix for each vertex. : like , ; like , ; in, This represents the distance between vertex i and vertex j; The attribute value represents the relationship between vertex i and vertex j; Step 1.3.1.5: If the weighted relation matrix obtained in step 1.3.1.3 is an asymmetric matrix, then convert it into a symmetric matrix; specifically, if... and ,but ; Step 1.3.2, Physical connection attribute values ​​under positive correlation conditions The larger the value, the greater the weight of the indirect association between components; the relationship attribute is the reciprocal of the shortest distance between each vertex and the physical connection attributes of all vertices; specifically: Step 1.3.2.1: Calculate the shortest distance of the physical connectivity attribute between each vertex and all other vertices using Dijkstra's algorithm. Calculate the relation attribute based on the shortest distance; the relation attribute is the reciprocal of the shortest distance of the physical connectivity attribute. Obtain the updated relation matrix using the relation attribute. ; Step 1.3.2.2: Sort the relational attributes of each vertex from smallest to largest, and select the vertex. x The former The set of neighboring nodes is denoted as _____. and each vertex The minimum relational attribute composition of the order neighbor nodes Order submatrix ; Step 1.3.2.3, Calculate the kernel radius: ; Step 1.3.2.4: Calculate the weighted relation matrix for each vertex. : like , ; like , ; Step 1.3.2.5: If the weighted relation matrix is ​​asymmetric, then convert it into a symmetric matrix; specifically, if... and ,but .

4. The fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network according to claim 3, characterized in that, The third step is specifically as follows: Step 2.1, Input module: The input module receives the completed system diagram data, where the fault feature matrix of the monitoring data is denoted as... ,in n The number of vertices. For monitoring the fault characteristics dimension of the data; Step 2.2, Multi-scale Chebyshev Convolution Fusion Module: The multi-scale Chebyshev convolutional fusion module is constructed by stacking multiple multi-scale Chebyshev convolutional fusion layers, each layer including Chebyshev graph convolution kernels of multiple scales; multi-scale Chebyshev graph convolution operations are performed in each layer to weighted aggregate fault feature information from different scale neighborhoods; finally, a two-dimensional enhanced feature matrix that fuses fault information from multiple scale neighborhoods is output. ,in n The number of vertices. To enhance the feature dimension of the feature matrix; Step 2.3, Fully Connected Module: The fully connected module integrates the two-dimensional enhanced feature matrix output by the multi-scale Chebyshev convolution fusion module. Expanded into a one-dimensional fault feature vector, the length of the one-dimensional fault feature vector is After processing through multiple fully connected layers, global fault characteristics of typical systems are extracted. Step 2.4, Graph Classification Module: The global fault features are fed into the graph classification module, which uses the Log-Softmax classifier to output the fault modes of typical systems.

5. The fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network according to claim 4, characterized in that, In step 2.2: Single-layer multi-scale Chebyshev convolutional fusion layer weighted aggregation i Chebyshev plot convolution kernels of different scales; the first j The multi-scale Chebyshev convolutional fusion layer includes i Each Chebyshev plot convolution kernel has a corresponding kernel. The Chebyshev plot convolution kernel scales are respectively ,different Chebyshev plot convolution kernels propagate fault information in different ranges of neighborhood; Regarding the first j The input of the multi-scale Chebyshev convolutional fusion layer is processed by Chebyshev graph convolution kernels. Then, its output is recorded as ; Weighted aggregation is used to perform weighted aggregation of Chebyshev graph convolution kernels at different scales, thereby obtaining enhanced fault features that fuse neighborhood information at different scales, denoted as... The weighted aggregation process is as follows: （1） in, The weighted aggregation operator controls the fusion weights of Chebyshev graph convolution kernels at different scales.

6. The fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network according to claim 5, characterized in that, In the fourth step, the cases of random data loss include the loss of random nodes and the loss of faulty nodes.

7. The fault diagnosis method based on a multi-scale Chebyshev convolutional fusion network according to claim 6, characterized in that, In the fourth step, the system graph data under missing data conditions is sent to the input module of the fault diagnosis model. The two-dimensional enhanced feature matrix under missing data conditions is obtained by the multi-scale Chebyshev convolution fusion module. Then, the global fault features under missing data conditions are extracted by the fully connected module. Finally, the graph classification module is used to realize the diagnosis of typical system faults under missing data conditions.

Images