Hypergraph structure-based multi-physics field transformer fault diagnosis method and system

Abstract

The invention provides a multi-physics field transformer fault diagnosis method and system based on a hypergraph structure. Multiple physical fields of the transformer in the operation process are collected, and a multi-channel data matrix is formed; establishing a transformer multi-physics field coupling simulation model, simulating a normal working condition and a plurality of fault working conditions by adjusting related parameters, and outputting simulation data; generating a hypergraph structure by adopting a proximity relation method based on the simulation data to obtain a node set, a hyperedge set and a hyperedge weight matrix, and mapping actually acquired multi-physical field signals to a hypergraph; and performing propagation calculation on the node features by using a hypergraph convolutional network to obtain low-dimensional embedded vectors of all channels, performing feature splicing to form global feature vectors, and finally outputting corresponding fault categories. According to the method, a stable hypergraph topological structure is constructed through simulation data assistance, and efficient fusion of multi-physical field features is realized in combination with real data, so that lightweight fault diagnosis of the transformer is realized under the condition of reducing sensors and computing resources.

Description

Technical Field

[0001] This invention relates to the field of multi-physics field fault diagnosis technology for power system transformers, and more specifically, to a method and system for multi-physics field transformer fault diagnosis based on a hypergraph structure. Background Technology

[0002] Transformer operating conditions involve coupling mechanisms of thermal field, dielectric electrical behavior (partial discharge), and acoustic radiation. Current online monitoring primarily relies on single physical quantities, such as temperature threshold alarms, partial discharge amplitude or pulse counts, acoustic energy, or location assessments, often depending on manual features or empirical thresholds. Due to fluctuations in operating conditions and noise, single-source signals are prone to misinterpretation; differences in data scale, time alignment, and statistical distribution across different physical fields make accurate characterization of cross-source correlations difficult, especially in scenarios with small or imbalanced samples.

[0003] Existing multi-source fusion methods often employ strategies such as feature concatenation and weighted voting, requiring numerous sensors and large-scale labeled data. Deep learning methods, such as one-dimensional convolution and recurrent networks, often treat each channel as an independent input, ignoring the structural relationships between channels. Although graph / hypergraph methods can describe many-to-many relationships, they usually require direct estimation of the topology from real data, making them susceptible to noise and sample scarcity, resulting in instability, insufficient generalization, and high computational cost.

[0004] Therefore, there is an urgent need in this field for a lightweight diagnostic technique that can robustly model multi-physics channel relationships under small sample conditions and is easy to deploy. Currently, no descriptions or reports of techniques similar to this invention have been found, and no similar materials from domestic or international sources have been collected. Summary of the Invention

[0005] To address the aforementioned shortcomings in the existing technology, this invention provides a method and system for diagnosing transformer faults based on a supergraph structure and multiphysics field.

[0006] According to one aspect of the present invention, a method for fault diagnosis of transformers based on a hypergraph structure multiphysics field is provided, characterized in that it includes: Collect multi-physics field signals of the transformer during operation to form a multi-channel real data matrix; Based on the real data matrix, a multi-physics coupling simulation model of the transformer is established. By adjusting the relevant parameters, the normal operating conditions and various fault conditions are simulated, and the simulation data matrix is ​​output. Based on the simulation data matrix, a hypergraph structure is generated using the proximity relation method, and the actual collected multiphysics field signals are mapped to the hypergraph structure to obtain hypergraph samples. The node features of the hypergraph samples are propagated and calculated using a hypergraph convolutional network to obtain low-dimensional embedding feature vectors for each channel. These feature vectors are then concatenated to form a global feature vector. Finally, the corresponding fault category is output through a classifier.

[0007] Preferably, the acquisition of multi-physics field signals from the transformer during operation to form a real data matrix includes: Multi-physics field signals of the transformer during operation are synchronously acquired and aligned on a unified time axis to form a multi-channel time series data matrix; wherein, the multi-physics field includes temperature signal, partial discharge signal and sound field signal, and the multi-channel time series data matrix includes: temperature field channel set T, partial discharge signal channel set D and sound field channel set P; The time series data matrices from each channel are normalized and then stacked to form the actual data matrix. ; wherein, the real data matrix The rows and channels correspond one-to-one.

[0008] Preferably, based on the real data matrix, a multi-physics coupling simulation model of the transformer is established. By adjusting relevant parameters, normal operating conditions and various fault conditions are simulated, and a simulation data matrix is ​​output, including: A transformer multiphysics coupled simulation model is established, which is co-structured with the channel set of the real data matrix. By adjusting the relevant parameters of the simulation model, normal operating conditions and various fault conditions are simulated, and a simulation data matrix is ​​output; wherein: Let the total number of channels in the channel set be: Let the length of the feature vector for each channel be L, then the true data matrix With simulation data matrix The i-th row corresponds to the measured and simulated representations of the i-th channel, respectively; the i-th row of the simulation data matrix S is denoted as the channel vector. ; For each channel vector Zero-mean and unit-variance normalization is performed as a distance metric for subsequent construction of the hypergraph structure and nearest neighbor retrieval.

[0009] Preferably, the step of generating a hypergraph structure based on the simulation data matrix using a proximity method, and mapping the actually acquired multiphysics field signals to the hypergraph structure to obtain hypergraph samples, includes: Based on the simulation data matrix, the proximity relationship between channels is calculated to generate a hypergraph structure; wherein, the hypergraph structure includes: an association matrix, a hyperedge weight matrix, a node degree matrix, and a hyperedge degree matrix; Using the correlation matrix as the topology and the real data matrix as the node features, the actual collected multi-physics field signals are mapped to the hypergraph structure to obtain hypergraph samples.

[0010] According to another aspect of the present invention, a multiphysics transformer fault diagnosis system based on a hypergraph structure is provided, comprising: Multi-source data acquisition module: This module is used to acquire multi-physical field signals of the transformer during operation and form a multi-channel real data matrix; The model-assisted hypergraph construction module establishes a transformer multi-physics coupling simulation model based on the real data matrix. By adjusting relevant parameters, it simulates normal operating conditions and various fault conditions, and outputs a simulation data matrix. Based on the simulation data matrix, a hypergraph structure is generated using the proximity relation method, and the actually collected multi-physics signals are mapped to the hypergraph structure to obtain hypergraph samples. The hypergraph convolutional diagnostic module uses a hypergraph convolutional network to propagate and calculate the node features of the hypergraph samples, obtains the low-dimensional embedding feature vectors of each channel, and concatenates the feature vectors to form a global feature vector. The classification and result output module is used to output the corresponding fault category through the classifier.

[0011] By adopting the above technical solution, the present invention has at least one of the following beneficial effects compared with the prior art: The present invention provides a method and system for transformer fault diagnosis based on a hypergraph structure and multiphysics field. It constructs a stable hypergraph topology structure with the assistance of simulation data and achieves efficient fusion of multiphysics field features by combining real data, thereby realizing lightweight fault diagnosis of transformers while reducing sensor and computing resources.

[0012] The present invention provides a method and system for transformer fault diagnosis based on a supergraph structure and multiphysics field. By synchronously collecting multiphysics field signals such as temperature, partial discharge, and sound field during transformer operation and standardizing them to form a real data matrix, a multiphysics field coupled simulation model co-constructed with the real data channel set is established. The parameters are adjusted to simulate normal and various fault conditions and output simulation data, thereby avoiding the problem that real data is easily affected by noise and operating condition fluctuations, and providing a stable data foundation for subsequent supergraph topology construction.

[0013] The present invention provides a method and system for multi-physics transformer fault diagnosis based on a hypergraph structure. It calculates the proximity relationship between channels based on simulation data (such as Euclidean distance to select K nearest neighbors), generates a hypergraph structure including an association matrix, a hyperedge weight matrix, a node degree matrix, and a hyperedge degree matrix, and then maps the real data matrix as node features to the hypergraph to form a hypergraph sample. This solves the problem that traditional graph / hypergraph methods rely on real data to build topology, resulting in structural instability and ambiguity of physical meaning. It enables the hypergraph to accurately capture the complex relationships between multi-physics channels.

[0014] The present invention provides a method and system for transformer fault diagnosis based on a hypergraph structure and multiphysics field. It uses a hypergraph convolutional network with only two layers to propagate and calculate the features of hypergraph sample nodes. After obtaining the low-dimensional embedding vectors of each channel, they are sequentially concatenated into a global feature vector, which is then input into a two-layer fully connected classifier to output the fault category. This significantly reduces the scale of network parameters and avoids the high computational overhead of traditional deep learning (such as 1D-CNN and Transformer). It is also suitable for the lightweight deployment requirements of transformer field edge devices.

[0015] The present invention provides a method and system for transformer fault diagnosis based on a hypergraph structure and multiphysics field. During the training phase, the hypergraph topology constructed from simulation data is fixed (the correlation matrix, hyperedge weight matrix, etc. remain unchanged). Only the trainable weights and classifier parameters of the hypergraph convolution are optimized. During the inference phase, there is no need to reconstruct the topology; only data input, convolution calculation, and classification output are performed. This not only improves the stability of parameter optimization during training but also significantly shortens the inference time, further enhancing the real-time diagnostic capability. At the same time, by efficiently fusing multiphysics field information through the hypergraph, the dependence on redundant sensors is reduced, thus ensuring diagnostic accuracy while reducing hardware costs. Attached Figure Description

[0016] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the process of a multiphysics transformer fault diagnosis method based on a hypergraph structure in a preferred embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of the constituent modules of a multiphysics transformer fault diagnosis system based on a hypergraph structure in a preferred embodiment of the present invention.

[0018] Figure 3 This is a comparison chart of the accuracy of various network models in a specific application example of the present invention.

[0019] Figure 4 This is a comparison curve of the number of iterations and the average absolute percentage error of each network model in a specific application example of the present invention.

[0020] Figure 5 This is a comparison chart of the training time of various network models after 200 iterations in a specific application example of the present invention.

[0021] Figure 6 This is a comparison chart of single-sample inference time for various network models in a specific application example of the present invention. Detailed Implementation

[0022] The embodiments of the present invention are described in detail below: These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

[0023] Existing technologies lack a lightweight transformer fault diagnosis technique that can robustly model multi-physics channel relationships under small sample conditions and is easy to deploy. To address this issue, one embodiment of the present invention provides a multi-physics transformer fault diagnosis method based on a hypergraph structure. This method uses simulation data to assist in constructing a stable hypergraph topology and combines it with real data to achieve efficient fusion of multi-physics features, thereby realizing lightweight transformer fault diagnosis while reducing sensor and computing resources.

[0024] Specifically, such as Figure 1 As shown, the multiphysics transformer fault diagnosis method based on a hypergraph structure provided in this embodiment may include: S1 collects multi-physics field signals of the transformer during operation to form a multi-channel real data matrix; S2, based on the real data matrix, establishes a multi-physics coupling simulation model of the transformer, and simulates normal working conditions and various fault conditions by adjusting relevant parameters, and outputs the simulation data matrix; S3, based on the simulation data matrix, uses the proximity relationship method to generate a hypergraph structure, and maps the actual collected multi-physics field signals to the hypergraph structure to obtain hypergraph samples; S4 utilizes a hypergraph convolutional network to propagate and calculate the node features of the hypergraph samples, obtains low-dimensional embedding feature vectors for each channel, concatenates the feature vectors to form a global feature vector, and finally outputs the corresponding fault category through a classifier.

[0025] In some preferred embodiments, S1, which involves collecting multi-physics field signals of the transformer during operation to form a real data matrix, may further include: S11, synchronously acquire and align the multi-physics field signals of the transformer during operation on a unified time axis to form a multi-channel time series data matrix; wherein, the multi-physics field includes temperature signal, partial discharge signal and sound field signal, and the multi-channel time series data matrix includes: temperature field channel set T, partial discharge signal channel set D and sound field channel set P; S12, after normalization, stacks the time series data matrices of each channel into a real data matrix. Among them, the real data matrix The rows and channels correspond one-to-one.

[0026] In some preferred embodiments, S2 above, based on a real data matrix, establishes a multi-physics coupling simulation model of the transformer, simulates normal operating conditions and various fault conditions by adjusting relevant parameters, and outputs a simulation data matrix. It may further include: S21, Establish a transformer multiphysics coupled simulation model with the channel set comorphic with the real data matrix. By adjusting the relevant parameters of the simulation model, simulate normal operating conditions and various fault conditions, and output the simulation data matrix; where: Let the total number of channels in the channel set be: Let the length of the feature vector for each channel be L, then the true data matrix With simulation data matrix The i-th row corresponds to the measured and simulated representations of the i-th channel, respectively; the i-th row of the simulation data matrix S is denoted as the channel vector. ; For each channel vector Zero-mean and unit-variance normalization is performed as a distance metric for subsequent construction of the hypergraph structure and nearest neighbor retrieval.

[0027] In some preferred embodiments, S3 above, based on the simulation data matrix, generates a hypergraph structure using the proximity relation method, and maps the actually acquired multiphysics field signals to the hypergraph structure to obtain hypergraph samples, including: S31, based on the simulation data matrix, calculate the proximity relationship between channels and generate a hypergraph structure; wherein, the hypergraph structure includes: an association matrix, a hyperedge weight matrix, a node degree matrix, and a hyperedge degree matrix; S32 uses the correlation matrix as the topology and the real data matrix as the node features to map the actual collected multi-physics field signals to the hypergraph structure, thus obtaining hypergraph samples.

[0028] In some preferred embodiments, the above-mentioned S31, which calculates the proximity relationship between channels based on the simulation data matrix and generates a hypergraph structure, may further include: S311, Let the node index of the channel set be... The reference channel index is Let the i-th row of the simulation data matrix S be denoted as the channel vector. The Euclidean distance between any two channels is defined as: S312, based on Euclidean distance, calculate the K nearest neighbor set for each reference channel r. The dimension is k; with reference channel r and set The union of sets defines a superedge. This yields a hypergraph with N nodes and N hyperedges, and the hypergraph's incidence matrix. Determine using the following formula: In the formula, This indicates that the i-th node belongs to the superedge formed with the r-th channel as the center. ; This indicates that the i-th node does not belong to the hyperedge formed with the r-th channel as the center. K is a given hyperparameter for the number of nearest neighbors; N is the total number of channels; L is the length of the feature vector for each channel. S313, the hyperedge weight matrix is ​​defined as follows: In the formula, For super-edge The weights; S314, Define the hyperboundary degree matrix With node degree matrix They are respectively: In the formula Indicates the superedge The number of nodes covered; This represents the total strength of node i contained within the weighted hyperedge.

[0029] In some preferred embodiments, the hypergraph sample in S32 above is represented as: In the formula, H is the node feature matrix; H is the correlation matrix; This is the hyperedge weight matrix. Represents the node degree matrix, This represents the hypermarginality matrix.

[0030] In some preferred embodiments, S4 above, which utilizes a hypergraph convolutional network to propagate and calculate the node features of the hypergraph samples to obtain low-dimensional embedding feature vectors for each channel, and concatenates the feature vectors to form a global feature vector, and finally outputs the corresponding fault category through a classifier, may further include: S41, a graph convolutional network is used to apply two layers of hypergraph convolutional propagation to the hypergraph samples to obtain node-level low-dimensional embedded feature vectors; S42, concatenate the rows of the low-dimensional embedded feature vectors in a predetermined order according to the channels to form a global feature vector; S43, input the global feature vector into the classifier, and output the diagnostic results corresponding to the preset fault set.

[0031] In some preferred embodiments, S41 above, which employs a graph convolutional network to apply two layers of hypergraph convolutional propagation to the hypergraph samples to obtain node-level low-dimensional embedding feature vectors, may further include: Perform two layers of hypergraph convolution forward propagation on the hypergraph samples, layer Output satisfy: In the formula, H is the incidence matrix, and W is the hyperedge weight matrix. Let the node degree matrix be... It is the hypermarginality matrix; The node feature matrix; For the first The trainable weight matrix of the layer; It is an element-wise nonlinear function.

[0032] In some preferred embodiments, the above-mentioned S42, which concatenates the rows of the low-dimensional embedded feature vector in a predetermined channel order to form a global feature vector, may further include: Record the output of the second layer By concatenating each row in a predetermined channel order, the global feature vector F is obtained as follows: In the formula, for The i-th row, This indicates a vector concatenation, where k is the output dimension of the second layer.

[0033] In some preferred embodiments, S43 above, which inputs the global feature vector into the classifier and outputs a diagnostic result corresponding to a preset fault set, may further include: Inputting the global feature vector F into a two-layer fully connected classifier yields the unnormalized class score vector z: In the formula, These are the trainable parameters. It is a nonlinear function; After obtaining the unnormalized class score vector z, calculate the class probability vector. And execute the decision: In the formula, It is a natural exponential function; The sample belongs to the category The normalized probability satisfies ; C represents the preset number of fault categories and the corresponding fault set. These include normal, overheating, partial discharge abnormalities, acoustic abnormalities, etc. The rules for determining fault categories are as follows: In the formula, For optimal category index; Output the final fault classification; This is a pre-defined bijective mapping used to map category indexes to specific fault labels.

[0034] In some preferred embodiments, S41 above, in the process of calculating the unnormalized category score vector z, further includes: S411, during the training phase, the cross-entropy loss is used to optimize the first feature vector in the global feature vector F. The trainable weight matrix of the layer and trainable parameters Preserve the topology of the hypergraph samples established by the simulation data matrix Fixed and unchanging; S412, the inference phase, only performs data acquisition, inputting the real data matrix into the graph, two-layer hypergraph convolution, and classifier output, without reconstructing the topology of the hypergraph samples to meet lightweight deployment requirements.

[0035] Based on the same inventive concept, an embodiment of the present invention also provides a multiphysics transformer fault diagnosis system based on a hypergraph structure.

[0036] Specifically, such as Figure 2 As shown, the multiphysics transformer fault diagnosis system based on a hypergraph structure provided in this embodiment may include: Multi-source data acquisition module: This module is used to acquire multi-physical field signals of the transformer during operation and form a multi-channel real data matrix; The model-assisted hypergraph construction module establishes a multi-physics coupled simulation model of a transformer based on a real data matrix. By adjusting relevant parameters, it simulates normal operating conditions and various fault conditions, and outputs a simulation data matrix. Based on the simulation data matrix, a hypergraph structure is generated using the proximity relation method, and the actual collected multi-physics signals are mapped to the hypergraph structure to obtain hypergraph samples. The hypergraph convolutional diagnostic module uses a hypergraph convolutional network to propagate and calculate the node features of the hypergraph samples, obtain the low-dimensional embedding feature vectors of each channel, and concatenate the feature vectors to form a global feature vector. The classification and result output module is used to output the corresponding fault category through the classifier.

[0037] The implementation methods of each functional module constituting the system provided in the above embodiments of the present invention are further described in detail below.

[0038] The multi-source data acquisition module includes: Multi-source data acquisition unit: acquires temperature field channel set T, partial discharge signal channel set D, and sound field channel set P, and synchronously acquires and aligns them on a unified time axis; Data matrix forming unit: Normalizes and stacks the time series data from each channel to form a real data matrix. Each row corresponds one-to-one with a channel; Multiphysics simulation unit: Establishes a simulation model that is co-structured with the channel set, and outputs simulation data matrix S under normal and several fault conditions.

[0039] The model-assisted hypergraph construction module includes: Model-aided graphing unit: Based on the simulation matrix S, it calculates the proximity relationship between channels, generates the hypergraph association matrix H and the hyperedge weight matrix W, and calculates the node degree matrix accordingly. With hypermarginal matrix ; Real data input unit: with H as the topology, and with Construct hypergraph samples based on node features.

[0040] The hypergraph convolution diagnostic module includes: Hypergraph convolutional feature extraction unit: for Applying two layers of hypergraph convolution propagation, we obtain node-level low-dimensional embeddings. ; Feature-level fusion unit: spliced ​​according to a predetermined channel order Each row of the vectors forms a global feature vector F.

[0041] Classification and Result Output Module: Input F into the classifier and output the diagnostic results corresponding to the preset fault set (including normal, overheating, partial discharge abnormality, and acoustic abnormality).

[0042] In the aforementioned multi-source data acquisition unit, data matrix forming unit, and multiphysics simulation unit: Let the total number of channels be: Let the length of the feature vector for each channel be L. Then the true data matrix... With simulation data matrix The i-th row corresponds to the measured and simulated representations of the i-th channel, respectively; the i-th row of S is denoted as the channel vector. .

[0043] Furthermore, for each channel vector Zero-mean and unit-variance normalization is performed for distance measurement and nearest neighbor retrieval in subsequent graph construction.

[0044] In the above model-aided composition unit: Let the node index of the channel set be The reference channel index is Let the i-th row of the simulation matrix S be denoted as the channel vector. Define the Euclidean distance between any two channels: The calculated Euclidean distance is primarily used to quantify the feature similarity of simulation data from various channels of the transformer's multi-physics field (temperature, partial discharge, sound field): the smaller the distance, the closer the simulation data features of the two channels are, and the higher the correlation of the physical conditions they reflect. Based on this Euclidean distance, the K nearest neighbor channels with the most similar features are selected for each reference channel, and then the reference channel and these K nearest neighbor channels are combined to form a hyperedge. Subsequently, the correlation matrix of the hypergraph is determined based on the composition of the hyperedge (clarifying whether each node belongs to a certain hyperedge), and the hyperedge weight matrix, node degree matrix, and hyperedge degree matrix are further generated by combining the correlation matrix, ultimately constructing a complete hypergraph structure. This hypergraph structure will serve as the topological basis for subsequently carrying the actual acquired multi-physics field signals, providing key structural support for the hypergraph convolutional network to propagate node features and realize transformer fault diagnosis.

[0045] Based on the calculated Euclidean distance, find the K nearest neighbor set for each reference channel r. The dimension is k; and it is connected to the set via a reference channel r. The union of sets defines a superedge. This yields a hypergraph with N nodes and N hyperedges, and its incidence matrix is... Determine using the following formula: in, This means that "the i-th node belongs to the superedge formed with the r-th channel as the center". ”; Indicates not belonging; K is the given number of nearest neighbors hyperparameter; N is the total number of channels; L is the feature vector length of each channel.

[0046] Furthermore, let the hyperedge weight matrix be: in, For super-edge The weights are defined as follows: the hyperedge degree matrix and the node degree matrix are respectively... in, Indicates the superedge The number of nodes covered; This represents the total strength of node i, which is contained by how many strip weighted superedges. It comes from the correlation matrix H; W is the diagonal matrix of the hyperedge weights, i.e., the hyperedge weight matrix.

[0047] In the above-mentioned real data plotting unit: The hypergraph sample for computation is defined as follows: in, H is the node feature matrix; H is the correlation matrix; This is the hyperedge weight matrix; In the hypergraph convolutional feature extraction unit mentioned above: Perform two layers of hypergraph convolution forward propagation on the hypergraph samples, layer The output satisfies: Where H is the correlation matrix, W is the hyperedge weight matrix, For the node degree matrix, It is the hypermarginality matrix; ; For the first The trainable weight matrix of the layer; It is an element-wise nonlinear function.

[0048] In the above feature-level fusion unit: Record the output of the second layer By concatenating each row in a predetermined channel order, the global feature is obtained: in, for The i-th row, This indicates vector concatenation, and k is the output dimension of the second layer.

[0049] In the above classification and result output module: Inputting the global feature F into a two-layer fully connected classifier yields an unnormalized class score vector: in For trainable parameters, It is a non-linear function; the training phase optimizes it using cross-entropy loss. , Preserve the topology built from simulation data The process remains fixed; during the inference phase, only data acquisition, real data input into the graph, two-layer hypergraph convolution, and classifier output are performed, without the need to rebuild the topology, in order to meet the requirements of lightweight deployment.

[0050] After obtaining z, calculate the category probability vector. And execute the decision: in It is a natural exponential function; The sample belongs to the category The normalized probability satisfies .

[0051] C represents the preset number of fault categories and the corresponding fault set. (For example, it may include "normal, overheating, partial discharge abnormality, acoustic abnormality", etc.); Fault Category Determination Rules: in: For optimal category index; Output the final fault classification; It is a fixed bijective mapping used to map category indices to specific fault labels.

[0052] It should be noted that the steps in the method provided by the present invention can be implemented using the corresponding components in the system. Those skilled in the art can refer to the technical solution of the system to implement the steps of the method, and can also refer to the technical solution of the method to implement the composition of the system. That is, the embodiments in the system and the embodiments in the method can be understood as preferred examples of each other, which will not be elaborated here.

[0053] The technical solution provided by the above embodiments of the present invention will be further described in detail below with reference to a specific verification example.

[0054] To verify the effectiveness of the present invention, in this specific verification example, the present invention is compared with traditional neural network methods: Physical Information Neural Networks (PINNs), surrogate model and reduced-order modeling (ROM), and Fourier Neural Network (FNO) physical field modeling methods.

[0055] First, perform the following operations: Multi-source data acquisition: Acquire temperature field channel set T, partial discharge signal channel set D, and sound field channel set P, and simultaneously acquire and align them on a unified time axis; Data matrix formation: The time series data from each channel are normalized and then stacked to form a real data matrix. Each row corresponds one-to-one with a channel; Multiphysics simulation: Establish a simulation model that is co-structured with the channel set, and output simulation data matrix S under normal and several fault conditions; Model-assisted graph construction: Based on the simulation matrix S, the proximity relationship between channels is calculated, generating the hypergraph association matrix H and the hyperedge weight matrix W, and the node degree matrix is ​​calculated accordingly. With hypermarginal matrix ; Real data plotted: with H as the topology, and... Construct hypergraph samples based on node features; Hypergraph convolution feature extraction: for Applying two layers of hypergraph convolution propagation, we obtain node-level low-dimensional embeddings. ; Feature-level fusion: stitching together channels in a predetermined order Each row forms a global feature vector F; Classification and Decision: Input F into the classifier and output the diagnostic results corresponding to the preset fault set (including normal, overheating, partial discharge abnormality, and acoustic abnormality).

[0056] The specific process is as described above and will not be repeated here.

[0057] To verify the effectiveness of this invention, comparative experiments were conducted under the same hardware and data partitioning conditions. Temperature field, local electrical field, and acoustic field signals from multiple sources were collected. Simulation-aided mapping and two-layer hypergraph convolution modeling were completed according to the steps of this invention, with unified window length and batch size, and an optimizer. Initialization and early stopping strategies were kept consistent across all models. Three common neural networks—1D-CNN, LSTM, and Transformer—were selected for comparison. The evaluation metrics were four key aspects: iteration count - accuracy (Acc(n)), iteration count - mean absolute percentage error (MAPE(n)), and fixed iteration training time. Compared with single-sample inference latency Each experiment was repeated 5 times, and the mean ± standard deviation was used. Comparison of results curves and bar charts is shown below. Figures 3-6As shown. Experimental results show that the present invention achieves approximately It is close to convergence at that time. The accuracy was 97.5%, and the MAPE was 0.12%; in the same period, the inference accuracy of 1DCNN, LSTM, and Transformer were 94.2%, 93.1%, and 95.8%, respectively, with MAPEs of 0.45%, 0.60%, and 0.35%. The present invention has the shortest training time. 1.2s), 1D-CNN, LSTM, and Transformer have latency of 30s, 78s, and 92s respectively; single-sample inference latency is also the lowest ( The comparison models were 2.0, 2.5, and 3.1 ms, respectively. These results demonstrate that the present invention exhibits faster convergence, lower error, and better timeliness across all four metrics.

[0058] The multi-physics transformer fault diagnosis method and system based on a hypergraph structure provided in the above embodiments of the present invention first collects multi-physics signals of the transformer during operation, including temperature signals, partial discharge signals, and acoustic field signals, and forms a multi-channel data matrix. Second, a multi-physics coupled simulation model of the transformer is established, simulating normal operating conditions and various fault conditions by adjusting relevant parameters, and outputting simulation data. Then, based on the simulation data, a hypergraph structure is generated using the proximity relation method, obtaining a set of nodes, a set of hyperedges, and a hyperedge weight matrix, and mapping the actually collected multi-physics signals to the hypergraph. Finally, a hypergraph convolutional network is used to propagate and calculate node features, obtaining low-dimensional embedding vectors for each channel, and these features are concatenated to form a global feature vector. Finally, a classifier outputs the corresponding fault category. This invention uses simulation data to assist in constructing a stable hypergraph topology structure and combines it with real data to achieve efficient fusion of multi-physics features, thereby achieving lightweight transformer fault diagnosis while reducing sensor and computing resources.

[0059] Any matters not covered in the above embodiments of the present invention are well-known in the art.

[0060] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the essence of the present invention.

Claims

1. A method for fault diagnosis of transformers based on a hypergraph structure and multiphysics, characterized in that, include: Collect multi-physics field signals of the transformer during operation to form a multi-channel real data matrix; Based on the real data matrix, a multi-physics coupling simulation model of the transformer is established. By adjusting the relevant parameters, the normal operating conditions and various fault conditions are simulated, and the simulation data matrix is ​​output. Based on the simulation data matrix, a hypergraph structure is generated using the proximity relation method, and the actual collected multiphysics field signals are mapped to the hypergraph structure to obtain hypergraph samples. The node features of the hypergraph samples are propagated and calculated using a hypergraph convolutional network to obtain low-dimensional embedding feature vectors for each channel. These feature vectors are then concatenated to form a global feature vector. Finally, the corresponding fault category is output through a classifier.

2. The method for fault diagnosis of transformers based on hypergraph structure multiphysics field according to claim 1, characterized in that, The multi-physics field signals acquired during the operation of the transformer are used to form a real data matrix, including: Multi-physics field signals of the transformer during operation are synchronously acquired and aligned on a unified time axis to form a multi-channel time series data matrix; wherein, the multi-physics field includes temperature signal, partial discharge signal and sound field signal, and the multi-channel time series data matrix includes: temperature field channel set T, partial discharge signal channel set D and sound field channel set P; The time series data matrices from each channel are normalized and then stacked to form the actual data matrix. ; wherein, the real data matrix The rows and channels correspond one-to-one.

3. The method for fault diagnosis of transformers based on hypergraph structure multiphysics as described in claim 1, characterized in that, Based on the aforementioned real data matrix, a multiphysics coupling simulation model of the transformer is established. By adjusting relevant parameters, it simulates normal operating conditions and various fault conditions, and outputs a simulation data matrix, including: A transformer multiphysics coupled simulation model is established, which is co-structured with the channel set of the real data matrix. By adjusting the relevant parameters of the simulation model, normal operating conditions and various fault conditions are simulated, and a simulation data matrix is ​​output; wherein: Let the total number of channels in the channel set be: Let the length of the feature vector for each channel be L, then the true data matrix With simulation data matrix The i-th row corresponds to the measured and simulated representations of the i-th channel, respectively; the i-th row of the simulation data matrix S is denoted as the channel vector. ; For each channel vector Zero-mean and unit-variance normalization is performed as a distance metric for subsequent construction of the hypergraph structure and nearest neighbor retrieval.

4. The method for fault diagnosis of transformers based on hypergraph structure multiphysics field according to claim 1, characterized in that, The process involves generating a hypergraph structure based on the simulated data matrix using a proximity method, and mapping the actually acquired multiphysics field signals onto the hypergraph structure to obtain hypergraph samples, including: Based on the simulation data matrix, the proximity relationship between channels is calculated to generate a hypergraph structure; wherein, the hypergraph structure includes: an association matrix, a hyperedge weight matrix, a node degree matrix, and a hyperedge degree matrix; Using the correlation matrix as the topology and the real data matrix as the node features, the actual collected multi-physics field signals are mapped to the hypergraph structure to obtain hypergraph samples.

5. The method for fault diagnosis of transformers based on hypergraph structure multiphysics field according to claim 4, characterized in that, The step of calculating the proximity relationships between channels and generating a hypergraph structure based on the simulation data matrix includes: Let the node index of the channel set be The reference channel index is Let the i-th row of the simulation data matrix S be denoted as the channel vector. The Euclidean distance between any two channels is defined as: Based on the Euclidean distance, calculate the K nearest neighbor set for each reference channel r. The dimension is k; with reference channel r and set The union of sets defines a superedge e. r This yields a hypergraph with N nodes and N hyperedges, and the incidence matrix of the hypergraph is... Determine using the following formula: In the formula, This indicates that the i-th node belongs to the hyperedge e centered on the r-th channel. r ; This indicates that the i-th node does not belong to the node with the first digit as its parent. The hyperedge e formed by the channels as the center r ; Given the number of nearest neighbors hyperparameter; Total number of channels; The length of the feature vector for each channel; Define the hyperedge weight matrix as follows: In the formula, For the superedge e r The weights; Define the hyperboundary degree matrix With node degree matrix They are respectively: In the formula Indicates the superedge The number of nodes covered; This represents the total strength of node i contained within the weighted hyperedge.

6. The method for fault diagnosis of transformers based on hypergraph structure according to claim 4, characterized in that, The hypergraph sample is represented as follows: In the formula, The node feature matrix; It is an incidence matrix; This is the hyperedge weight matrix. Represents the node degree matrix, This represents the hypermarginality matrix.

7. The method for fault diagnosis of transformers based on hypergraph structure multiphysics field according to claim 1, characterized in that, The process involves using a hypergraph convolutional network to propagate and compute the node features of the hypergraph samples, obtaining low-dimensional embedding feature vectors for each channel. These feature vectors are then concatenated to form a global feature vector. Finally, a classifier outputs the corresponding fault category, including: A graph convolutional network is used to apply two layers of hypergraph convolution propagation to the hypergraph samples to obtain node-level low-dimensional embedding feature vectors. The rows of the low-dimensional embedded feature vector are concatenated in a predetermined order according to the channels to form a global feature vector; The global feature vector is input into the classifier, which outputs a diagnostic result corresponding to a preset fault set.

8. The method for fault diagnosis of transformers based on hypergraph structure according to claim 7, characterized in that, The method employs a graph convolutional network to apply two layers of hypergraph convolution propagation to the hypergraph samples, obtaining node-level low-dimensional embedding feature vectors, including: Perform a two-layer hypergraph convolution forward propagation on the hypergraph samples, layer... Output satisfy: In the formula, It is an incidence matrix. This is the hyperedge weight matrix. Let the node degree matrix be... It is the hypermarginality matrix; The node feature matrix; For the first The trainable weight matrix of the layer; It is an element-wise nonlinear function; The process of concatenating the rows of the low-dimensional embedded feature vector in a predetermined channel order to form a global feature vector includes: Record the output of the second layer By concatenating each row in a predetermined channel order, the global feature vector F is obtained as follows: In the formula, for The i-th row, This indicates a vector concatenation, where k is the output dimension of the second layer. The step of inputting the global feature vector into the classifier and outputting a diagnostic result corresponding to a preset fault set includes: Inputting the global feature vector F into a two-layer fully connected classifier yields the unnormalized class score vector z: In the formula, These are the trainable parameters. It is a nonlinear function; After obtaining the unnormalized class score vector z, calculate the class probability vector. And execute the decision: In the formula, It is a natural exponential function; The sample belongs to the category The normalized probability satisfies ; The number of preset fault categories corresponds to the fault set. ; The rules for determining fault categories are as follows: In the formula, For optimal category index; Output the final fault classification; This is a pre-defined bijective mapping used to map category indexes to specific fault labels.

9. The method for fault diagnosis of transformers based on hypergraph structure according to claim 8, characterized in that, The process of calculating the unnormalized class score vector z also includes: During the training phase, the global feature vector is optimized using cross-entropy loss. The first in The trainable weight matrix of the layer and trainable parameters Preserve the topology of the hypergraph samples established by the simulation data matrix Fixed and unchanging; During the inference phase, only data acquisition, input of the real data matrix into the graph, two-layer hypergraph convolution and classifier output are performed; the topological structure of the hypergraph samples is no longer reconstructed.

10. A multiphysics transformer fault diagnosis system based on a hypergraph structure, characterized in that, include: Multi-source data acquisition module: This module is used to acquire multi-physical field signals of the transformer during operation and form a multi-channel real data matrix; The model-aided hypergraph construction module establishes a multi-physics coupled simulation model of the transformer based on the real data matrix. By adjusting relevant parameters, it simulates normal operating conditions and various fault conditions, and outputs a simulation data matrix. Based on the simulation data matrix, a hypergraph structure is generated using the proximity relation method, and the actual collected multiphysics field signals are mapped to the hypergraph structure to obtain hypergraph samples. The hypergraph convolutional diagnostic module uses a hypergraph convolutional network to propagate and calculate the node features of the hypergraph samples, obtains the low-dimensional embedding feature vectors of each channel, and concatenates the feature vectors to form a global feature vector. The classification and result output module is used to output the corresponding fault category through the classifier.

Images