A voltage transformer fault diagnosis method based on improved ALIF and graph neural network

Abstract

This invention discloses a voltage transformer fault diagnosis method based on an improved ALIF and graph neural network, comprising: acquiring voltage signals of the voltage transformer under different operating conditions, preprocessing to obtain time series data; decomposing to obtain multiple intrinsic mode function components; calculating the comprehensive score of each component; reconstructing the effective components to obtain a denoised signal, converting it into graph structure data, constructing a multi-scale graph neural network model, extracting key node features, fusing multi-scale graph features, and outputting the fault diagnosis result of the voltage transformer. This invention achieves adaptive determination of the number of decomposition layers by constructing a multi-index fusion signal decomposition evaluation mechanism and a dynamic threshold control strategy, effectively avoiding the over-decomposition or under-decomposition problems in traditional decomposition algorithms; it can maintain stable feature extraction capabilities in strong noise environments, improving the accuracy and robustness of fault diagnosis, and has good engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and fault diagnosis of power systems, specifically a voltage transformer fault diagnosis method based on improved ALIF and graph neural network. Background Technology

[0002] With the rapid expansion of my country's power system and the continuous increase in the proportion of new energy integration, the safe operation and accurate monitoring of voltage transformers (VTs) in the power system are of paramount importance. Voltage transformers are used to step down high-voltage signals and provide measurement and protection signals; their operating status directly affects the accuracy of system relay protection and metering. Internal faults can lead to protection failure, measurement errors, or equipment damage, posing a serious threat to the safety and stability of the power system.

[0003] Traditional voltage transformer fault diagnosis methods mainly rely on manual experience and signal analysis based on single features, such as Fourier transform spectrum analysis, wavelet transform, or statistical feature extraction. However, due to the complex internal structure of voltage transformers, strong coupling between components, and external interference, single-feature methods often fail to accurately capture subtle abnormal signals, especially under conditions of strong non-stationary and nonlinear signals, where fault detection effectiveness is limited.

[0004] In recent years, with the development of intelligent monitoring technology, data-driven fault diagnosis methods have gradually attracted attention. Adaptive signal decomposition techniques (such as Empirical Mode Decomposition (EMD) and Local Iterative Filtering (LIF)) can decompose complex signals into intrinsic mode functions (IMFs), providing a foundation for subsequent feature extraction; Graph Neural Networks (GNNs) can model the structural relationships of signals, capture long-range dependencies between nodes, and are suitable for multi-channel, multi-scale signal feature fusion. However, existing methods still have the following problems:

[0005] 1. Insufficient decomposition accuracy for non-stationary signals, easily leading to mode aliasing or noise interference;

[0006] 2. Single-scale graphical neural networks struggle to fully utilize the correlation information between different frequency components or multi-channel signals;

[0007] 3. The signal feature selection and comprehensive evaluation lack an effective adaptive mechanism, which cannot take into account energy, entropy value and frequency domain characteristics, resulting in limited diagnostic accuracy.

[0008] Technical comparison with patent CN116243230B "An online diagnosis method for voltage transformer faults".

[0009] Patent CN116243230B describes a method for online fault diagnosis. This involves constructing an input-output mathematical model of a voltage transformer under healthy conditions, obtaining theoretical values ​​by combining them with the output of a standard voltage transformer, and then calculating the deviation between the actual output and the theoretical value and comparing it to a threshold. The core of this method relies on least squares modeling and quantitative deviation analysis. In contrast, our method employs an improved adaptive local iterative filter to adaptively decompose the voltage signal of the voltage transformer. Compared to methods relying solely on mathematical models and deviation thresholds, this approach effectively adapts to the nonlinear system characteristics of voltage transformers, overcomes the limitations of mathematical models in fitting non-stationary signals, and accurately captures subtle anomalies that are easily overlooked in model deviation analysis.

[0010] Comparison with the technology of patent CN118777962A "A method for diagnosing insulation faults in voltage transformers based on graph convolutional neural networks".

[0011] In patent CN118777962A, the error components of the secondary signals of multiple in-phase current transformers are separated by PCA algorithm. After extracting time-domain statistical features, the system identifies faults, optimizes features using a separation matrix, and then uses a graph convolutional neural network to diagnose insulation faults. Since the comparative patent relies on synchronous sampling of multiple in-phase devices for PCA decoupling, only filters fixed time-domain features, and uses a single-scale graph convolutional network, its applicability is limited to insulation faults of multiple current transformers in the same group. Our system eliminates the need for collaborative data acquisition from multiple devices, relies on adaptive filtering to suppress noise interference, and utilizes a multi-scale graph network to deeply mine multi-dimensional features of the time-series topology. This results in a wider range of applicable fault types and stronger noise resistance and diagnostic adaptability.

[0012] Technical comparison with patent CN120451737A "An Improved MobileNetV2 Voltage Transformer Fault Diagnosis Method Based on Multi-channel Feature Images".

[0013] In patent CN120451737A, an overlapping sliding window with Gaussian noise is used for data augmentation to transform one-dimensional time series into GAF, recursive graph, and wavelet time-frequency data. Figure 3 For channel images, fault classification is performed using an improved lightweight convolutional network based on MobileNetV2. After converting the temporal sequence into a two-dimensional image, local pixel features are extracted by convolution. This breaks the natural temporal topological relationship of the sampling points. We directly construct a temporal topological graph structure and rely on graph neural networks to mine long-range dependencies and temporal relationships between nodes. Combined with pre-adaptive filtering and noise reduction, the feature extraction stability and fault detail capture ability are better under strong noise conditions.

[0014] Technical comparison with patent CN121299297A "Intelligent Monitoring and Early Warning Method Based on Multi-Source Information Fusion".

[0015] In patent CN121299297A, four types of multi-source heterogeneous data—voltage, partial discharge, temperature, and vibration—are collected simultaneously. Multi-source signal preprocessing is performed using EMD and wavelet packets. Dynamic weighted fusion of multi-dimensional features is built through temporal-spatial correlation and temperature-vibration coupling. Fault identification and health scoring are completed using fixed fault thresholds and empirical formulas. This requires additional temperature, vibration, and partial discharge sensors, resulting in high hardware deployment costs. Fault determination relies on manually set thresholds, which are susceptible to interference from operating conditions and aging. This solution, however, can complete diagnosis using only existing voltage sampling data, eliminating the need for additional sensing hardware. Adaptive filtering suppresses on-site noise, and graph neural networks autonomously learn feature associations, overcoming the constraints of fixed thresholds. This results in lower engineering implementation costs and stronger environmental robustness. Summary of the Invention

[0016] To address the aforementioned technical problems, this invention proposes a voltage transformer fault diagnosis method based on an improved ALIF and graph neural network. By constructing a multi-index adaptive signal decomposition mechanism and a multi-scale topological feature learning model, it achieves high-precision identification of the operating status of voltage transformers and fault type discrimination.

[0017] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0018] A fault diagnosis method for voltage transformers based on improved ALIF and graph neural networks, the specific steps of which are as follows:

[0019] (1) Acquisition of voltage transformer operating signals;

[0020] During the operation of the voltage transformer, the voltage time series signal output from the secondary side of the voltage transformer is acquired through a data acquisition device to obtain the original signal sequence. The acquired signals are then standardized to form data inputs for subsequent signal decomposition and fault diagnosis.

[0021] (2) Signal decomposition based on improved adaptive local iterative filtering;

[0022] Original signal The input is decomposed using an adaptive local iterative filtering algorithm to obtain several intrinsic mode function components. and residual signal ,in During the decomposition process, a multi-dimensional feature evaluation system is constructed to assess the effectiveness of each component, and the number of decomposition layers is adaptively determined based on a dynamic threshold mechanism, thereby obtaining a set of signal components with physical meaning.

[0023] (3) Evaluation and screening of the effectiveness of the decomposition components;

[0024] For each decomposition component The complexity characteristic index, frequency domain structure characteristic index, and energy contribution index are calculated respectively to form a multi-dimensional evaluation vector. The weights of each indicator are calculated using the entropy weight fusion method. A comprehensive evaluation score was obtained. .

[0025] (4) Adaptive decomposition layer determination and noise reduction signal reconstruction;

[0026] According to the comprehensive scoring sequence Constructing a nonlinear dynamic threshold function This is used to determine whether each IMF component is valid; the decomposition process stops when the score of a newly generated component falls below a threshold. The set of valid components is retained. The noise-reduced signal is obtained by superimposing and reconstructing the signal. .

[0027] (5) Construction of time series graph structure;

[0028] The noise-reduced signal The sample sequence is formed by slicing the data into segments with a fixed window length, and the time series data is mapped to a graph structure data using a path graph structure construction method. , where the set of nodes Represents signal sampling points and edge sets. It represents the connection relationship between adjacent time nodes, thus forming a structured data representation suitable for graph neural network learning.

[0029] (6) Fault diagnosis of multi-scale graph neural networks;

[0030] Graph-structured data is input into a graph isomorphic network (GIN) with residual connections, and node features are encoded and learned to obtain node embedding representations. This method is used to characterize the topological relationships and local structural features between different nodes in the operating signals of voltage transformers. Based on node feature encoding, a graph U-Net network structure is introduced. Multi-scale graph representation spaces are constructed through graph downsampling and upsampling operations to extract topological information at different scales. Simultaneously, a self-attention graph pooling mechanism is used to evaluate the importance of nodes and select a set of key nodes. To enhance the model's ability to represent fault-sensitive regions, a gated readout mechanism is used to weighted aggregate node features, resulting in a graph-level representation vector. The fault status is then input into a classifier for fault condition identification, and the fault category of the voltage transformer is output. .

[0031] Furthermore, the voltage transformer operation signal acquisition in step (1) includes the following process:

[0032] Through the data acquisition system at the sampling frequency The operating signal of the voltage transformer is continuously acquired, and a length of [length missing] is obtained. The original signal sequence The signal is then normalized to eliminate the influence of dimensions.

[0033] Furthermore, step (2) of signal decomposition based on improved adaptive local iterative filtering includes the following steps:

[0034] (2-1) For the original signal Adaptive local iterative filtering decomposition is performed to obtain several intrinsic mode function components. ;

[0035] (2-2) After each layer of decomposition is completed, calculate the feature index of the current component set and determine whether to continue decomposition based on the dynamic threshold function;

[0036] (2-3) When the comprehensive evaluation index is lower than the set threshold, the decomposition is terminated and the final component set is obtained.

[0037] Furthermore, step (3) of evaluating the effectiveness of component decomposition includes the following steps:

[0038] (3-1) Calculate the complexity characteristic index of each component. ;

[0039] (3-2) Calculate the frequency domain structure characteristic index of each component. ;

[0040] (3-3) Calculate the energy contribution index of each component ;

[0041] (3-4) Calculate the weights of each indicator using the entropy weight method. And obtain a comprehensive evaluation score. .

[0042] Furthermore, step (6) of multi-scale graph neural network fault diagnosis includes the following steps:

[0043] (6-1) Graph Isomorphic Network (GIN) is used for node feature encoding, and residual connections are introduced to improve the stability of feature propagation;

[0044] (6-2) Multi-scale downsampling and upsampling of graph structures are performed through Graph U-Net to form a graph feature space that fuses local and global features;

[0045] (6-3) Neighborhood-sensitive pooling (NS-Pool) and self-attention graph pooling (SAGPool) are used to extract key nodes and suppress the influence of noisy nodes; the problem of inconsistent distribution of deep features is solved by feature alignment and normalization mechanism (Feature Alignment + GraphNorm);

[0046] The benefits of this application are:

[0047] 1. This invention proposes a fault diagnosis method for voltage transformers based on improved adaptive local iterative filtering and multi-scale graph neural networks. This method achieves adaptive determination of the number of decomposition layers by constructing a multi-index fusion signal decomposition evaluation mechanism and a dynamic threshold control strategy, effectively avoiding the over-decomposition or under-decomposition problems in traditional decomposition algorithms.

[0048] 2. Simultaneously, by constructing a multi-scale graph neural network structure and integrating graph isomorphic networks, graph U-Net, self-attention graph pooling, and gated readout mechanisms, collaborative modeling of local anomaly features and global structural information of voltage transformer operating signals was achieved.

[0049] 3. This invention can maintain stable feature extraction capability in strong noise environment, improve the accuracy and robustness of fault diagnosis, and has good engineering application value. Attached Figure Description

[0050] Figure 1 This is a flowchart of the method disclosed in this invention;

[0051] Figure 2 This is a schematic diagram of the structure of the improved graph neural network disclosed in this invention;

[0052] Figure 3 This is a schematic diagram of the forward propagation process of the improved graph neural network disclosed in this invention;

[0053] Figure 4 This is a confusion matrix diagram of the diagnostic model proposed in this invention;

[0054] Figure 5 This is a 3D t-SNE plot of the diagnostic model proposed in this invention. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0056] This invention discloses a voltage transformer fault diagnosis method based on improved adaptive local iterative filtering and multi-scale graph neural network. The implementation flowchart of the method is as follows: Figure 1 As shown, the steps include the following.

[0057] Step (1) Acquisition of voltage transformer operating signals is as follows:

[0058] (1-1) During the operation of the voltage transformer, the secondary voltage signal is continuously sampled by the data acquisition system to obtain the original time series signal:

[0059] ;

[0060] in The original acquired signal, The number of sampling points. For the first Each sample value is represented by a sliding window slice. A number of sampling points are used for subsequent decomposition and graph structure modeling.

[0061] (1-2) In order to eliminate the difference in dimensions and improve the stability of subsequent algorithms, the signal is normalized:

[0062] ;

[0063] in The mean of the signal. The standard deviation of the signal. The normalized signal yields a standardized sequence:

[0064] ;

[0065] Step (2) of adaptive local iterative filtering signal decomposition includes the following steps:

[0066] (2-1) Standardize the signal The input is an adaptive local iterative filtering algorithm, which uses a local moving average filter to calculate the local mean of the signal.

[0067] ;

[0068] in Here is the filter kernel function. Length of the filter window

[0069] (2-2) Obtain the local oscillation component by subtracting the local mean from the original signal:

[0070] ;

[0071] (2-3) Repeat the filtering iteration process continuously until the stopping condition of the intrinsic mode function (IMF) is met, and obtain the first... One component:

[0072] ;

[0073] (2-4) Update the remaining signals

[0074] ;

[0075] (2-5) Repeat the above steps to finally obtain the signal decomposition form:

[0076] ;

[0077] Step (3) The calculation of multi-index features of component decomposition includes the following steps:

[0078] (3-1) Calculate the sample entropy Used to measure signal complexity:

[0079] ;

[0080] in For matching dimensions Vector count For matching dimensions number of vectors

[0081] (3-2) Calculate spectral kurtosis Used to measure signal impulse characteristics:

[0082] ;

[0083] in For the signal at frequency Work spectral density at that location and These are the mean and standard deviation of the PSD, respectively.

[0084] (3-3) Calculate the energy ratio , used to represent the energy contribution of each component:

[0085] ;

[0086] (3-4) Calculate the comprehensive score using the entropy weight method

[0087] ;

[0088] Step (4) Dynamic threshold determination and noise reduction signal reconstruction includes the following steps:

[0089] Based on the IMF composite score sequence Constructing a nonlinear dynamic threshold function

[0090] ;

[0091] like If the decomposition stops, the effective components are retained.

[0092] ;

[0093] Step (5) Graph structure construction includes the following steps:

[0094] noise reduction signal Divided into nodes Establish a time-series path graph and an adjacency matrix. Defined as:

[0095] ;

[0096] Node feature matrix It contains information such as amplitude, frequency, and energy for each window node.

[0097] Step (6) Multi-scale graph neural network fault diagnosis, the network structure and forward propagation process are as follows: Figure 2 , 3 As shown, it includes the following steps:

[0098] (6-1) Node feature encoding using Graph Isomorphic Network with Residue (GIN):

[0099] ;

[0100] (6-2) Graph U-Net multi-scale downsampling and upsampling, combined with neighborhood-sensitive pooling (NS-Pool):

[0101] ;

[0102] in Score the importance of nodes.

[0103] (6-3) Self-attention graph pooling (SAGPool) further filters key nodes:

[0104] ;

[0105] (6-4) Gated Readout Aggregation Node Features

[0106] .

[0107] As shown in Figures 4 and 5, Figure 4 is the confusion matrix of the diagnostic model, which shows that the proposed method has high accuracy in identifying various operating states and fault types of voltage transformers, with very few misclassifications and excellent classification performance. Figure 5 is a 3D t-SNE visualization of the model features, showing good clustering effects and clear boundary distinctions for different categories of sample features, proving that the fault features extracted by this invention have strong discriminative power, and the overall algorithm demonstrates good diagnostic accuracy and feature representation ability.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A voltage transformer fault diagnosis method based on improved ALIF and graph neural network, characterized in that: The specific steps are as follows: S1. Acquisition of voltage transformer operation signals; The voltage signals of the voltage transformer under different operating conditions are collected, and the collected raw voltage signals are preprocessed to obtain the time series data to be analyzed. S2, adaptive signal decomposition; The adaptive local iterative filtering algorithm ALIF is used to decompose the time series data to be analyzed, and multiple intrinsic mode functions (IMF) components are obtained. S3, component decomposition evaluation; A three-index evaluation system consisting of sample entropy (SE), spectral peak kurtosis (SK), and energy ratio (ER) is constructed to quantitatively evaluate each IMF component. The entropy weight method is then used to adaptively fuse the three indices to obtain a comprehensive score for each IMF component. S4. Adaptive decomposition level determination; A nonlinear dynamic threshold function is constructed based on the comprehensive scoring sequence of IMF components. The number of signal decomposition layers is adaptively determined, effective IMF components are screened and reconstructed to obtain the denoised voltage signal. S5. Graph structure construction; The denoised voltage signal is converted into graph structure data, and node feature matrix and adjacency relationship are constructed. S6. Fault diagnosis of graph neural networks; The graph structure data is input into a multi-scale graph neural network model, node features are encoded through a graph isomorphic network, and multi-scale structural features are extracted by combining Graph U-Net, neighborhood-sensitive pooling, self-attention graph pooling and gated readout mechanism, and finally the fault diagnosis results of the voltage transformer are output.

2. The voltage transformer fault diagnosis method based on improved ALIF and graph neural network according to claim 1, characterized in that: In step S1, the voltage signal acquisition method of the voltage transformer under different operating states is as follows: the voltage signal of the secondary side of the voltage transformer is sampled in real time by a voltage acquisition device at a sampling frequency of 1kHz. The acquired continuous signal is sliced ​​using a fixed-length sliding window, and each sample window contains 1024 sampling points for subsequent signal processing and model training.

3. The voltage transformer fault diagnosis method based on improved ALIF and graph neural network according to claim 1, characterized in that: Step S2 involves using the Adaptive Local Iterative Filtering (ALIF) to decompose the time series data to be analyzed, specifically including the following steps: S21. Input the original voltage signal as the signal to be decomposed; S22. Construct a local filtering kernel function using the adaptive local iterative filtering algorithm ALIF, and perform a local weighted average operation on the signal to obtain the local mean of the signal; S23. By iteratively stripping away the oscillation components from the signal layer by layer, multiple intrinsic mode function (IMF) components and residual signals are obtained.

4. The voltage transformer fault diagnosis method based on improved ALIF and graph neural network according to claim 1, characterized in that: The step S3, component decomposition evaluation, includes the following steps: S31. Calculate the sample entropy SE of each IMF component to describe the signal complexity characteristics; S32. Calculate the spectral peak kurtosis SK of each IMF component to describe the frequency domain abrupt change characteristics of the signal; S33. Calculate the energy ratio ER of each IMF component to reflect the energy contribution of the component to the original signal. S34. Normalize the three indicators and calculate the weight of each indicator using the entropy weight method. S35. The three indicators are weighted and integrated according to their respective weights to obtain a comprehensive score for each IMF component.

5. The voltage transformer fault diagnosis method based on improved ALIF and graph neural network according to claim 1, characterized in that: The adaptive decomposition level determination step S4 includes the following steps: S41. Calculate the mean and standard deviation of the top k tier scores based on the obtained IMF composite scores; S42. Construct a nonlinear dynamic threshold function to analyze the trend of score changes; S43. When the overall score of the newly generated IMF component is lower than the dynamic threshold, stop the signal decomposition process; S44. Retain the IMF components that meet the conditions and perform superposition and reconstruction to obtain the denoised voltage signal.

6. The voltage transformer fault diagnosis method based on improved ALIF and graph neural network according to claim 1, characterized in that: The graph structure construction in step S5 includes the following steps: S51. Divide the noise-reduced voltage signal into multiple nodes; S52. According to the graph construction method PathGraph, construct the path graph structure in chronological order and establish edges to connect adjacent time nodes; S53. Construct the node feature matrix and adjacency matrix to form the input data for the graph neural network.

7. The voltage transformer fault diagnosis method based on improved ALIF and graph neural network according to claim 1, characterized in that: Step S6, graph neural network fault diagnosis, includes the following steps: S61. Use graph isomorphic network (GIN) to perform neighborhood aggregation and feature encoding on node features; S62. Introduce residual connection structures to enhance feature propagation stability; S63. Perform multi-scale downsampling and upsampling operations on graph structures using Graph U-Net; S64. Calculate node importance and filter key nodes using the neighborhood-sensitive pooling mechanism; S65. Further compress node features using the self-attention graph pooling module SAGPool; S66. The node features are weighted and aggregated using the gated readout mechanism to obtain a graph-level feature representation. S67. The fault category of the voltage transformer is output through the classification layer.

Images