A Method and System for Fault Diagnosis of Tap Changers Based on Mechanistic Dual-Graph and Time-Frequency Hierarchical Fusion

Abstract

This invention discloses a method and system for fault diagnosis of tap changers based on mechanistic dual-graph fusion and time-frequency hierarchical fusion. The method includes: collecting time-series data of dissolved gas in oil from on-load tap changers of transformers and preprocessing it to obtain a standardized multi-source time-series dataset; based on the gas generation mechanism of on-load tap changers, performing hierarchical extraction of time-frequency features from the standardized multi-source time-series dataset, fusing the extracted time-frequency features to obtain a mechanistic multi-scale feature set; constructing a feature association graph that integrates the gas generation mechanism weights and insulation state constraints to form a mechanistic dynamic dual-graph structure data; training an optimal fault diagnosis model; inputting the dissolved gas data into the optimal fault diagnosis model, and outputting the fault diagnosis result. This invention closely aligns with the physical nature of UHV transformer faults, significantly improves robustness in small-sample, high-interference scenarios, and adapts to the online monitoring needs of complex engineering scenarios of UHV transformers.

Description

Technical Field

[0001] This invention mainly relates to the field of power equipment condition monitoring and intelligent fault diagnosis technology, specifically to a method and system for fault diagnosis of tap changers based on mechanistic dual-map and time-frequency hierarchical fusion. Background Technology

[0002] The on-load tap changer (OLTC) of an ultra-high voltage transformer is the core component with the most frequent operation in the ultra-high voltage transmission system, and dissolved gas analysis (DGA) is an effective means of online monitoring of its internal faults. However, the "operational gas generation" generated by the normal voltage regulation operation of the OLTC and the fault-generated gas are superimposed. In addition, the electromagnetic interference in the ultra-high voltage scenario is strong and effective fault samples are scarce, making fault diagnosis difficult.

[0003] Among the existing diagnostic methods, empirical rule methods such as the IEC three ratio method have low accuracy, traditional machine learning methods such as SVM are difficult to capture the dynamic characteristics of time series, and deep learning methods such as CNN / LSTM have improved accuracy, but have the following shortcomings: (1) They do not combine the physical mechanism of gas generation in OLTC faults, the model lacks interpretability and has insufficient generalization for small samples; (2) The time-frequency feature fusion adopts fixed weights and cannot be adaptively adjusted; (3) The graph structure is static and cannot adapt to the dynamic evolution of faults; (4) There is a lack of interference suppression mechanism for gas generation in OLTC actions, and the accuracy drops significantly under strong interference conditions. Summary of the Invention

[0004] To address the technical problems existing in the prior art, this invention provides a tap changer fault diagnosis method and system based on mechanistic dual-graph and time-frequency hierarchical fusion, which significantly improves robustness in small sample and high interference scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A method for fault diagnosis of tap changers based on mechanistic dual-map and time-frequency hierarchical fusion includes the following steps: S1. Collect time-series data of dissolved gas in oil of transformer on-load tap changer and preprocess it to obtain a standardized multi-source time-series dataset. S2. Based on the fault gas generation mechanism of transformer on-load tap changer, time-frequency features are extracted hierarchically from standardized multi-source time-series datasets. Combined with a dynamic weighted fusion strategy based on fault identification, the extracted time-frequency features are fused to obtain a mechanistic multi-scale feature set. S3. Using the mechanistic multi-scale feature set as input, construct a feature correlation graph that integrates the gas production mechanism weight and insulation state constraint, as well as a time-series evolution graph driven by the fault evolution rate, to form a mechanistic dynamic dual-graph structure data. S4. Construct a dual-graph attention network diagnostic model. The diagnostic model takes the mechanistic dynamic dual-graph structure data as input and optimizes the hyperparameters of the diagnostic model by improving the sine and cosine algorithms to obtain the optimal fault diagnosis model after training. S5. After processing the dissolved gas data in the oil of the on-load tap changer of the transformer to be tested through steps S1 to S3, the data is input into the optimal fault diagnosis model, and the optimal fault diagnosis model outputs the fault diagnosis results. The fault diagnosis results include the fault type, identification confidence level, and contribution of key nodes.

[0006] Preferably, the preprocessing in step S1 includes adaptively eliminating interference from on-load tap changer operation and performing multi-stage preprocessing on the data after interference elimination. The adaptive elimination of interference from on-load tap changer operation includes: identifying the gas generation interference period caused by normal operation of on-load tap changer through a time-series mutation point detection algorithm of dissolved gas data in oil, and applying a time-frequency feature weakening coefficient λ to the data during the gas generation interference period to achieve adaptive suppression of interference features. The multi-stage preprocessing includes: removing abnormal time-series samples using an improved Laida criterion that combines sliding window threshold adaptive adjustment; eliminating data dimension differences through Z-score standardization; supplementing missing data using an adaptive interpolation method based on time-series trends; and suppressing environmental noise interference through wavelet threshold denoising.

[0007] Preferably, the time-frequency feature hierarchical extraction in step S2 specifically includes: The preprocessed data is decomposed by discrete wavelet transform using the db4 wavelet basis to obtain a low-frequency component and multiple high-frequency components. The low-frequency component and some of the high-frequency components are retained. For overheating faults, the trend-type time-frequency features of the low-frequency components are extracted, and the trend-type time-frequency features include low-frequency energy entropy and trend slope; For discharge faults, abrupt time-frequency features based on Hilbert-Huang transform decomposition are extracted, including instantaneous amplitude abrupt change rate and high-frequency energy proportion. For complex faults, cross-scale coupling features are extracted, including the high-low frequency energy ratio and the time-frequency feature covariance.

[0008] Preferably, the dynamic weighted fusion strategy based on fault identification in step S2 specifically includes: Calculate the identification degree of various faults of on-load tap changer of transformer for different time-frequency characteristics. The degree of recognition Determined based on the ratio of inter-class distance to intra-class distance; based on the discriminative power of each time-frequency feature. Calculate the corresponding dynamic fusion weights Dynamic fusion weights are adopted. Weighted fusion of trend-type time-frequency features, abrupt time-frequency features, and cross-scale coupled features; The fused mechanistic time-frequency features are further fused with the statistical features and gas ratio features of the original data to obtain a multi-dimensional feature set. Redundant features in the multi-dimensional feature set are then removed by principal component analysis, and core features with a cumulative contribution rate reaching a preset ratio are retained to obtain a mechanistic multi-scale feature set.

[0009] Preferably, in step S3, constructing a feature correlation graph that integrates the weights of the gas generation mechanism and the insulation state constraints specifically includes: Using the core features extracted in step S2 as the base nodes, insulation parameters are added as insulation state constraint nodes to form the node set of the feature association graph; Edge weights between nodes Integrating data correlation and fault gas generation mechanisms, the calculation formula is as follows:

[0010] in, The Pearson correlation coefficient is used. For mutual information value, For mechanism weighting coefficients, The prior weights are based on the gas generation mechanism of UHV OLTC faults; Cross-validation is used to adaptively determine weight thresholds, and edges with association strength lower than the weight threshold are removed to form a feature association graph.

[0011] Preferably, in step S3, the time-series evolution diagram driven by the fault evolution rate specifically includes: Each time step is defined as a set of mechanistic multi-scale features, and each node feature is the core feature vector of the corresponding time step. An edge is defined as the temporal transition relationship between adjacent time step nodes, and the edge weight is... Based on the dynamic calculation of fault evolution rate and characteristic changes, the calculation formula is as follows: ,in This is the fault evolution rate coefficient. The characteristic changes at adjacent time steps are used to enhance the temporal characteristics of faults with different evolution rates. The feature association graph and the temporal evolution graph support online dynamic updates: when the change in feature association degree exceeds a preset association degree change threshold, or the temporal evolution rate exceeds a preset evolution rate threshold, the node weights and / or edge weights of the feature association graph and the temporal evolution graph are automatically updated.

[0012] Preferably, in step S4, the dual-graph attention network diagnostic model includes a spatial attention module, a temporal attention module, a feature fusion layer, and an adaptive node selection pooling layer; The spatial attention module takes the feature association graph as input and adaptively learns the attention weight of each node through a multi-head attention mechanism to enhance the contribution of core fault feature nodes and insulation state constraint nodes, and suppress redundant features and interference nodes. The time attention module takes the time evolution graph as input and focuses on the key time step nodes of fault evolution through a multi-head attention mechanism to capture the long-term temporal dependencies of faults at different rates. The feature fusion layer uses a weighted concatenation strategy to fuse the output features of the spatial attention module and the temporal attention module, and maps the fused features into a high-dimensional feature vector through a fully connected layer to provide support for fault classification. The adaptive node selection pooling layer filters key nodes and removes noisy nodes based on the contribution score of nodes to fault classification, and extracts multi-scale graph features by combining local max pooling and global mean pooling.

[0013] Preferably, in step S4, the formula for calculating the attention weights of the spatial attention module and the temporal attention module is as follows:

[0014] The formula for calculating node feature updates is:

[0015] in, Let be the attention weight of node i to node j, be the attention weight vector, and be the learnable weight matrix. and Let be the feature vectors of node i and node j, respectively. Let i be the neighborhood set of node i The feature vector of any neighbor node k in the traversal, where k is the feature vector of the traversal. The index variable of all neighboring nodes; || represents the feature concatenation operation. Let σ be the set of neighboring nodes of node i, and let σ be the activation function. This refers to the updated node features.

[0016] The present invention also discloses a computer-readable storage medium having a computer program stored thereon, the computer program performing the steps of the method described above when run by a processor.

[0017] This invention further discloses a tap changer fault diagnosis system based on mechanistic dual-graph fusion and time-frequency hierarchical fusion, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0018] Compared with the prior art, the advantages of the present invention are as follows: This invention improves anti-interference capabilities under strong interference scenarios in ultra-high voltage (UHV) environments by adaptively suppressing OLTC (Online Continuous Voltage Control) gas generation interference at the data level and combining an improved Laida criterion with wavelet denoising to eliminate multi-source noise. At the feature level, it extracts three types of time-frequency features—trend-type, abrupt-type, and cross-scale coupling—based on the fault gas generation mechanism to enhance fault identification. Design dynamic fusion weights Instead of a fixed-weight strategy, this approach enhances the differentiated sensitivity of various fault types. At the graph structure level, it constructs a feature association graph that integrates gas generation mechanism weights and insulation state constraint nodes, as well as a temporal evolution graph that drives dynamic edge weights with fault evolution rate coefficients. A dual-graph online update mechanism adapts to the complete dynamic evolution process of faults from early stages to degradation. At the model level, DualGAT's spatial and temporal attention modules model the spatial relationships and temporal dependencies of nodes, respectively. Weighted concatenation achieves deep synergistic fusion of spatiotemporal features, and ISCA global optimization of hyperparameters overcomes the limitations of a single model. Furthermore, the contribution of core feature nodes and key time step nodes is quantified based on attention weights, making the diagnostic results physically interpretable. Physical priors such as gas generation mechanism, insulation constraints, and evolution rate are embedded into each learning stage, significantly improving the model's generalization ability under the condition of scarce UHV samples. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the transformer on-load tap changer fault diagnosis method provided in an embodiment of the present invention.

[0020] Figure 2 This is a flowchart illustrating the construction process of the mechanistic dynamic dual-graph structure in an embodiment of the present invention.

[0021] Figure 3 This is a flowchart illustrating the construction process of the dual-graph attention network diagnostic model in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, the on-load tap changer fault diagnosis method for UHV transformers based on mechanistic dual-graph construction and hierarchical time-frequency feature fusion provided by this embodiment of the invention includes the following steps: Step S1: Collect time-series data of dissolved gas in oil from the on-load tap changer of the UHV transformer, adaptively eliminate OLTC operation interference, and perform multi-stage preprocessing to obtain a standardized multi-source time-series dataset; the specific process is as follows: S11: Collect time-series sample data of dissolved gases in UHV OLTC oil, covering normal samples and five types of fault samples: low temperature overheating, high temperature overheating, partial discharge, arc discharge, and insulation aging. Each sample contains characteristic gas content data of H2, CH4, C2H6, C2H4, and C2H2, oil micro-water / dielectric loss insulation parameters, and load and ambient temperature operating parameters for 15-30 consecutive time steps. S12: The time-series abrupt change point detection algorithm based on DGA data identifies gas-generating interference periods during normal switch operation. A time-frequency feature weakening coefficient λ (0 < λ ≤ 0.3) is applied to the data during these periods to achieve adaptive removal of OLTC operation interference. The time-series abrupt change point detection algorithm is based on the first-order difference sequence of gas content change rate in DGA data, and locates the start and end positions of gas-generating interference periods by setting an adaptive abrupt change threshold. The value of the weakening coefficient λ can be dynamically adjusted according to the OLTC operation frequency and interference intensity: when the OLTC operation frequency is high per unit time, λ is taken as a smaller value to enhance the interference suppression effect; when the operation frequency is low, λ is taken as a larger value to retain more effective data.

[0024] S13: An improved Laida criterion combining adaptive adjustment of sliding window threshold is used to remove abnormal time series samples; the size of the sliding window is set to w, where w is a positive integer; the mean and standard deviation of the data within the window are calculated, and the adaptive threshold interval is set as mean ± k times standard deviation (the value of k is dynamically adjusted according to the distribution characteristics of the data within the window). Data points exceeding the threshold interval are marked as outliers and removed.

[0025] S14: Eliminate data dimension differences through Z-score standardization; the Z-score standardization calculation formula is:

[0026] in, For the original characteristic gas data, This is the mean of this type of data. The standard deviation of this type of data. The data is standardized. Meanwhile, an adaptive interpolation method based on time-series trends is used to supplement missing data to ensure sequence integrity; S15: Environmental noise interference is initially suppressed through wavelet threshold denoising to obtain a preprocessed multi-source time series dataset.

[0027] Step S2: Based on the gas generation mechanism of on-load tap changer faults in UHV transformers, time-frequency features are extracted hierarchically, and a mechanistic multi-scale feature set is obtained by combining the fault identification degree with a dynamic weighted fusion strategy; specifically, it includes the following sub-steps: S21: The db4 wavelet basis is used to perform discrete wavelet transform decomposition on the preprocessed data to obtain one low-frequency component and three high-frequency components. The low-frequency component and the first two high-frequency components are retained. The low-frequency component reflects the slow pyrolysis trend of the overheating fault, and the high-frequency component reflects the instantaneous gas production change of the discharge fault. S22: Based on the differentiated gas production characteristics of different faults in UHV OLTC, the following three types of time-frequency features are extracted in a hierarchical manner: For overheating faults (slowly changing), trend-type time-frequency features of wavelet low-frequency components are extracted, including low-frequency energy entropy and trend slope. For arc / partial discharge faults (instantaneous), the abrupt time-frequency features of HHT decomposition are extracted, including the instantaneous amplitude abrupt change rate and the proportion of high-frequency energy; For complex faults, cross-scale coupling features are extracted, including the high-low frequency energy ratio and time-frequency feature covariance. S23: Perform Hilbert-Huang transform analysis on each retained component. First, obtain the intrinsic mode functions (IMFs) through empirical mode decomposition (EMD), and calculate the instantaneous frequency, instantaneous amplitude, and energy entropy of each IMF. The formula for calculating the energy entropy is:

[0028] in The value represents the ratio of the energy of the i-th intrinsic mode function to the total energy, and n represents the number of intrinsic mode functions. This enables the accurate capture of local features in nonlinear and non-stationary time series data. S24: Calculate the identification degree D of typical faults in UHV OLTC with different time-frequency characteristics. f The formula for calculating the recognition rate is:

[0029] Where d f(k1,k2) Let d be the inter-class distance of feature f between fault types k1 and k2. f(k) The intra-class distance of feature f within fault type k; Dynamic weighted fusion weights are designed based on recognition. m represents the total number of features, replacing the traditional fixed weights to complete the fusion of trend-type, mutation-type, and cross-scale coupled features, and constructing a mechanistic multi-scale feature set; S25: Further integrate the fused mechanistic time-frequency features with the statistical features (mean, variance, peak value) and gas ratio features (C2H2 / C2H4, CH4 / H2) of the original data to construct a multi-dimensional mechanistic multi-scale feature set; S26: Principal component analysis (PCA) is used to reduce the dimensionality of the multidimensional mechanistic multiscale feature set, eliminate redundant features, and retain the core features with a cumulative contribution rate of 95%, thereby obtaining an optimized mechanistic multiscale feature set, reducing the feature dimensionality and model learning difficulty.

[0030] Step S3: Based on the spatiotemporal characteristics and physical mechanisms of dissolved gas data in the on-load tap changer oil of UHV transformers, construct a feature correlation graph and a time-series evolution graph that integrate gas generation mechanisms, insulation state constraints, and support online dynamic updates, forming a mechanistic dynamic dual-graph structure data; specifically including the following sub-steps: S31: Constructing a correlation diagram of mechanistic features 1. Using the core features extracted in step S2 as the base nodes, add trace water in the oil as insulation state constraint points to form a node set of the feature association graph; 2. The correlation between node weighted fusion data and the gas generation mechanism of UHV on-load tap changer faults are calculated using the following formula:

[0031] in The Pearson correlation coefficient is used. For mutual information value, Mechanism weighting coefficients (strong correlation characteristics between C2H2 and arc discharge, CH4 / C2H4 and overheating mechanisms) =0.3~0.5, weak correlation feature =0~0.1), The prior weights are based on the gas generation mechanism of UHV OLTC faults; 3. The weight threshold is adaptively determined through cross-validation, and edges with association strength lower than the weight threshold are removed to form an initial mechanistic feature association graph; S32: Constructing a time-series evolution graph driven by the fault evolution rate 1. The mechanistic multi-scale feature set of each time step is used as a node, and the node feature is the core feature vector of the corresponding time step; 2. An edge is defined as the temporal transition relationship between adjacent time step nodes. The edge weight is dynamically calculated based on the fault evolution rate and feature changes, as shown in the formula. , where k v For fault evolution rate coefficient (arc fault k) v =0.8 1.0, Insulation aging k v =0.2~0.4, low temperature / high temperature overheating fault k v =0.5~0.7), ΔF is the characteristic change amount between adjacent time steps, realizing the time-series characteristic differentiation enhancement of faults with different evolution rates; S33: Online dynamic update of dual-graph structure: Set a feature correlation degree change threshold of 20% and a fault evolution rate threshold. Based on the real-time DGA data and insulation parameters of UHV OLTC, when the feature correlation degree change > 20% or the time-series evolution rate > the preset threshold, the node / edge weights of the feature correlation graph and the time-series evolution graph are automatically updated to adapt to the dynamic evolution process of the fault and replace the traditional static graph structure.

[0032] Step S4: Construct and optimize the dual-graph attention network diagnostic model. The model hyperparameters are optimized using an improved sine and cosine algorithm to obtain the optimal fault diagnosis model after training. The dual-graph attention network diagnostic model includes a spatial attention module (DisGAT), a temporal attention module (TempGAT), a feature fusion layer, and an adaptive node selection pooling layer, such as... Figure 2 As shown, the specific construction and optimization process includes: S41: The spatial attention module inputs a feature association graph and adaptively learns node weights through a multi-head attention mechanism. The formula for calculating the attention weights is:

[0033] in, Let be the attention weight of node i to node j, be the attention weight vector, and be the learnable weight matrix. and Let be the feature vectors of node i and node j, respectively. This indicates a feature concatenation operation. Let i be the set of neighboring nodes of node i; The contribution of core fault feature nodes and insulation state constraint nodes is enhanced, while redundant features and interference nodes are suppressed. The time attention module takes into account the time evolution graph driven by the fault evolution rate and focuses on the key time step nodes of fault evolution through the above attention weight formula to capture the long-term temporal dependencies of faults at different rates, thus solving the problem of insufficient temporal capture capability of traditional models. S42: Through Complete the node feature update for the dual attention module, and select the activation function σ. The function; the feature fusion layer uses a weighted concatenation strategy to fuse the output features of the dual attention module. The weights are dynamically adjusted based on the recognizability of the two types of features for UHV OLTC faults, achieving deep collaborative fusion of spatiotemporal features; then, a fully connected layer maps the fused features into a high-dimensional feature vector, providing support for fault classification output; the fused feature vector z∈R d ( d (To fuse feature dimensions) it enters the fault type discrimination layer. This layer first maps z into a 5-dimensional logits vector o∈R through a fully connected classification head. 5 The calculation formula is:

[0034] in Wc ∈R 5×d 、 b c ∈R 5 These are learnable parameters. There are 5 typical fault types for UHV transformers with logits dimensions k=0,1,…,4: k=0 Normal operation, k=1 Low temperature overheating, k=2 High temperature overheating, k=3 Partial discharge, and k=4 Arc discharge.

[0035] The logits are then converted into posterior probabilities for each fault type using the Softmax function. ; The final fault type determination result is: ; The confidence level is the corresponding probability value. .

[0036] The model training uses a multi-class cross-entropy loss function: ; Where y represents the one-hot true label.

[0037] In addition, attention weights based on space and time It can directly quantify the contribution of strongly correlated nodes of gas generation mechanism in the feature correlation diagram (such as the correlation between C2H2 and arc discharge), insulation state constraint nodes, and key time steps with high evolution rate in the time series evolution diagram to specific fault types, providing operation and maintenance personnel with interpretable fault mechanism basis.

[0038] S43: The adaptive node selection pooling layer selects key nodes and removes noisy nodes based on the contribution score of nodes to the classification of UHV OLTC faults; it combines local max pooling and global mean pooling to extract multi-scale graph features, further improving the model's anti-interference ability and adaptability in UHV high-interference and small-sample scenarios. S44: The improved sine cosine algorithm (ISCA) is adopted, with the 5-fold cross-validation accuracy as the objective function. Key hyperparameters such as the number of attention heads (2-8), hidden layer dimension (64-256), learning rate (0.001-0.01), and pooling threshold are optimized, and the optimal combination of hyperparameters is obtained iteratively. S45: The Adam optimizer is used to train the model. An early stopping strategy is introduced to prevent overfitting. If the accuracy of the validation set does not improve for 5 consecutive rounds, training is stopped to obtain the optimal fault diagnosis model after training.

[0039] Step S5: Input the data to be detected, after processing in steps S1-S3, into the optimal fault diagnosis model trained in step S4, and output the fault diagnosis results, including fault type, identification confidence and contribution of key nodes.

[0040] Specifically, after the DGA data and insulation parameters of the UHV OLTC to be tested are processed through OLTC action interference removal and preprocessing in step S1, mechanistic time-frequency feature hierarchical extraction and dynamic weighted fusion in step S2, and mechanistic dynamic dual-graph structure construction in step S3, they are input into the trained dual-graph attention network diagnostic model. The model outputs the fault type, identification confidence, and contribution of key nodes. The contribution of key nodes includes the contribution of core fault feature nodes and insulation state constraint nodes in the feature association graph, as well as the contribution of key time step nodes in the fault development graph in the time series evolution graph, providing quantitative basis for fault location, troubleshooting, and evolution trend analysis for operation and maintenance personnel.

[0041] This invention improves anti-interference capabilities under strong interference scenarios in ultra-high voltage (UHV) power transmission by adaptively suppressing OLTC (Online Continuous Voltage Control) gas generation interference at the data level and combining an improved Laida criterion with wavelet denoising to eliminate multi-source noise. At the feature level, it extracts three types of time-frequency features—trend-type, abrupt-type, and cross-scale coupling—based on the fault gas generation mechanism. It replaces the fixed-weight strategy with a dynamic fusion weight ωf designed using the fault identification degree Df, achieving differentiated sensitivity enhancement for multiple fault types. At the graph structure level, it constructs a feature association graph between the fusion gas generation mechanism weight βg and the insulation state constraint nodes, and a time-series evolution of the dynamic edge weights driven by the fault evolution rate coefficient kv. The graph is illustrated, and a dual-graph online update mechanism is used to adapt to the complete dynamic evolution process of the fault from its early stage to its deterioration. At the model level, the spatial and temporal attention modules of DualGAT model the spatial relationships and temporal dependencies of nodes respectively, and achieve deep synergistic fusion of spatiotemporal features through weighted splicing. Combined with ISCA global optimization of hyperparameters, it breaks through the limitations of a single model. In addition, the contribution of core feature nodes and key time step nodes is quantified based on attention weights, which makes the diagnostic results physically interpretable. Furthermore, physical priors such as gas generation mechanism, insulation constraints, and evolution rate are embedded into each learning stage, which significantly improves the generalization ability of the model under the condition of scarce UHV samples.

[0042] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0043] S1. Multi-source DGA timing data acquisition, OLTC motion interference removal and preprocessing 1000 DGA time-series samples were collected from the OLTC (Optical Linear Transformer) of an ultra-high voltage transformer of a power grid company. Each sample contains characteristic gas data (H2, CH4, C2H6, C2H4, C2H2), oil micro-water parameters, load, and ambient temperature operating parameters for 20 consecutive time steps. Among these, 500 samples are normal, 160 are low-temperature overheating samples, 140 are high-temperature overheating samples, 110 are partial discharge samples, and 90 are arc discharge samples. A time-series abrupt change detection algorithm identified 42 gas-generating interference samples indicating normal switch operation, and these were removed by applying a feature weakening coefficient λ=0.2. An improved Laida criterion combining sliding window threshold adaptive adjustment was used to remove 48 abnormal samples, leaving 952 samples. Z-score standardization was then applied. To eliminate dimensional differences, the mean μ and standard deviation σ of each type of characteristic gas and insulation parameter were calculated. A small amount of missing data was supplemented by an adaptive interpolation method based on time-series trends. After wavelet threshold denoising, the dataset was divided into a training set (700 sets) and a test set (300 sets) in a 7:3 ratio to obtain the preprocessed multi-source time-series dataset.

[0044] S2. Mechanistic Time-Frequency Feature Hierarchical Extraction and Dynamic Weighted Fusion like Figure 3 As shown, the db4 wavelet basis is used to perform DWT decomposition on the preprocessed data, resulting in one low-frequency component and three high-frequency components. The low-frequency component and the first two high-frequency components are retained. EMD decomposition is then performed on each retained component, resulting in five IMF components. The instantaneous frequency and instantaneous amplitude are calculated for each IMF component, and then... Calculate the energy entropy, where p i Let be the ratio of the energy of the i-th IMF component to the total energy of that decomposed component. Calculate the fault identification degree of each time-frequency characteristic for UHV OLTC6 state, and obtain the instantaneous amplitude mutation rate identification degree D of C2H2. f =8.7 (highest), the low-frequency trend slope discrimination of CH4 D f =7.2, dynamic fusion weight ω calculated based on recognition. f The process involves fusing trend-based, mutation-based, and cross-scale coupling features; then fusing the statistical features of the original data with the ratio features of the four types of gases to construct a multidimensional mechanistic multiscale feature set; and finally, through PCA dimensionality reduction, retaining the core features with a cumulative contribution rate of 95%, resulting in an optimized mechanistic multiscale feature set.

[0045] S3. Construction of Mechanistic Dynamic Dual-Graph Structure Mechanistic feature correlation diagram: Based on core features as base nodes, new nodes for oil micro-water and insulation state constraints are added, for a total of 17 nodes; the edge weights between base nodes are as follows: Calculate the weighting coefficient β of C2H2 and the mechanism of arc discharge. g=0.4, CH4 / C2H4 and β of overheating fault g =0.3, the remaining weakly correlated features β g =0.05, the mechanism prior weight Wm is set according to the UHV OLTC fault gas generation standard; the edge weight between insulation constraint nodes and basic nodes is calculated based on the correlation between insulation parameters and gas characteristics, and an adaptive threshold of 0.28 is set to remove low correlation edges and form an initial mechanism feature correlation graph.

[0046] Fault evolution rate-driven time-series evolution graph: using core features at 20 time steps as nodes, edge weights are assigned according to... Calculate, where the arc discharge k v =0.9, partial discharge k v =0.8, high temperature overheating k v =0.7, low temperature overheating k v =0.6, insulation aging k v =0.3, ΔF is the feature Euclidean distance between adjacent time steps; the temporal evolution rate threshold is set to 0.6, and the feature correlation change threshold is set to 20%. When the conditions are met, the node weights of the two graphs are automatically updated. Finally, a mechanistic dynamic two-graph structure data is formed.

[0047] S4. Construction and Optimization of Dual-Graph Attention Network Model A DualGAT model is constructed, comprising a spatial attention module (DisGAT), a temporal attention module (TempGAT), a feature fusion layer, and an adaptive node selection pooling layer. Each DisGAT and TempGAT module has four attention heads, with a hidden layer dimension of 128. The adaptive node selection threshold is set to 0.3, and the pooling layer employs a combination of local max pooling and global mean pooling. Both attention modules are implemented using... Calculate the attention weights, where the attention weight vector 'a' has dimensions [64, 1] and the learnable weight matrix 'W' has dimensions [64, 128]. After completing node feature updates, the ReLU activation function was selected. An improved sine-cosine algorithm (ISCA) was used to optimize hyperparameters, with 5-fold cross-validation accuracy as the objective function. The number of iterations was 60, and the population size was 35. The optimal parameters were obtained after optimization: learning rate 0.007, batch size 32, and pooling threshold 0.35. The Adam optimizer was used to train the model, with an early stopping policy: training was stopped if the validation set accuracy showed no improvement for five consecutive rounds. This resulted in the optimal fault diagnosis model after training.

[0048] S5. Fault Identification and Result Output The 513 test datasets were processed using the OLTC action interference removal, preprocessing, mechanistic feature extraction, and dual-graph construction described above, and then input into the optimal fault diagnosis model. The model learns and updates node features through an attention module, and outputs the fault type, identification confidence, and contribution of key nodes after feature fusion and pooling. Experimental results show that the overall identification accuracy of this method is significantly better than the traditional IEC three-ratio method, the single-graph attention network model, and the basic David's triangle model.

[0049] For the dissolved gas data and insulation parameters in the UHV OLTC oil to be tested, after processing according to the above steps S1-S3, input them into the trained model, and the fault diagnosis results can be quickly output. Among them, the contribution of key nodes can clearly identify the contribution of core fault feature nodes and insulation state constraint nodes in the feature association diagram, as well as the key time step nodes of fault development in the time series evolution diagram, providing accurate quantitative basis for the precise location, investigation and evolution trend analysis of UHV OLTC faults.

[0050] This invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0051] This invention further provides a tap changer fault diagnosis system based on mechanistic dual-map and time-frequency hierarchical fusion, including a memory and a processor connected to each other. The memory stores a computer program, which executes the steps of the method described above when run by the processor.

[0052] The medium and system of the present invention, corresponding to the methods described above, also have the advantages described above.

[0053] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0054] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for fault diagnosis of tap changers based on mechanistic dual-graph fusion and time-frequency hierarchical fusion, characterized in that, Includes the following steps: S1. Collect time-series data of dissolved gas in oil of transformer on-load tap changer and preprocess it to obtain a standardized multi-source time-series dataset. S2. Based on the fault gas generation mechanism of transformer on-load tap changer, time-frequency features are extracted hierarchically from standardized multi-source time-series datasets. Combined with a dynamic weighted fusion strategy based on fault identification, the extracted time-frequency features are fused to obtain a mechanistic multi-scale feature set. S3. Using the mechanistic multi-scale feature set as input, construct a feature correlation graph that integrates the gas production mechanism weight and insulation state constraint, as well as a time-series evolution graph driven by the fault evolution rate, to form a mechanistic dynamic dual-graph structure data. S4. Construct a dual-graph attention network diagnostic model. The diagnostic model takes the mechanistic dynamic dual-graph structure data as input and optimizes the hyperparameters of the diagnostic model by improving the sine and cosine algorithms to obtain the optimal fault diagnosis model after training. S5. After processing the dissolved gas data in the oil of the on-load tap changer of the transformer to be tested through steps S1 to S3, the data is input into the optimal fault diagnosis model, and the optimal fault diagnosis model outputs the fault diagnosis results. The fault diagnosis results include the fault type, identification confidence level, and contribution of key nodes.

2. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to claim 1, characterized in that, The preprocessing in step S1 includes adaptively eliminating interference from on-load tap changer operation and performing multi-stage preprocessing on the data after interference elimination. The adaptive elimination of interference from on-load tap changer operation includes: identifying the gas generation interference period caused by normal operation of on-load tap changer through a time-series mutation point detection algorithm of dissolved gas data in oil, and applying a time-frequency feature weakening coefficient λ to the data during the gas generation interference period to achieve adaptive suppression of interference features. The multi-stage preprocessing includes: removing abnormal time-series samples using an improved Laida criterion that combines sliding window threshold adaptive adjustment; eliminating data dimension differences through Z-score standardization; supplementing missing data using an adaptive interpolation method based on time-series trends; and suppressing environmental noise interference through wavelet threshold denoising.

3. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to claim 1, characterized in that, Step S2, the hierarchical extraction of time-frequency features, specifically includes: The preprocessed data is decomposed by discrete wavelet transform using the db4 wavelet basis to obtain a low-frequency component and multiple high-frequency components. The low-frequency component and some of the high-frequency components are retained. For overheating faults, the trend-type time-frequency features of the low-frequency components are extracted, and the trend-type time-frequency features include low-frequency energy entropy and trend slope; For discharge faults, abrupt time-frequency features based on Hilbert-Huang transform decomposition are extracted, including instantaneous amplitude abrupt change rate and high-frequency energy proportion. For complex faults, cross-scale coupling features are extracted, including the high-low frequency energy ratio and the time-frequency feature covariance.

4. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to claim 3, characterized in that, The dynamic weighted fusion strategy based on fault identification in step S2 specifically includes: Calculate the identification degree of various faults of on-load tap changer of transformer for different time-frequency characteristics. The degree of recognition Determined based on the ratio of inter-class distance to intra-class distance; based on the discriminative power of each time-frequency feature. Calculate the corresponding dynamic fusion weights Dynamic fusion weights are adopted. Weighted fusion of trend-type time-frequency features, abrupt time-frequency features, and cross-scale coupled features; The fused mechanistic time-frequency features are further fused with the statistical features and gas ratio features of the original data to obtain a multi-dimensional feature set. Redundant features in the multi-dimensional feature set are then removed by principal component analysis, and core features with a cumulative contribution rate reaching a preset ratio are retained to obtain a mechanistic multi-scale feature set.

5. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to any one of claims 1-4, characterized in that, In step S3, a feature correlation graph is constructed that integrates the weights of the gas generation mechanism and the insulation state constraints, specifically including: Using the core features extracted in step S2 as the base nodes, insulation parameters are added as insulation state constraint nodes to form the node set of the feature association graph; Edge weights between nodes Integrating data correlation and fault gas generation mechanisms, the calculation formula is as follows: in, The Pearson correlation coefficient is used. For mutual information value, For mechanism weighting coefficients, The prior weights are based on the gas generation mechanism of UHV OLTC faults; Cross-validation is used to adaptively determine weight thresholds, and edges with association strength lower than the weight threshold are removed to form a feature association graph.

6. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to claim 5, characterized in that, In step S3, the time-series evolution diagram driven by the fault evolution rate specifically includes: Each time step is defined as a set of mechanistic multi-scale features, and each node feature is the core feature vector of the corresponding time step. An edge is defined as the temporal transition relationship between adjacent time step nodes, and the edge weight is... Based on the dynamic calculation of fault evolution rate and characteristic changes, the calculation formula is as follows: ,in This is the fault evolution rate coefficient. The characteristic changes at adjacent time steps are used to enhance the temporal characteristics of faults with different evolution rates. The feature association graph and the temporal evolution graph support online dynamic updates: when the change in feature association degree exceeds a preset association degree change threshold, or the temporal evolution rate exceeds a preset evolution rate threshold, the node weights and / or edge weights of the feature association graph and the temporal evolution graph are automatically updated.

7. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to any one of claims 1-4, characterized in that, In step S4, the dual-graph attention network diagnostic model includes a spatial attention module, a temporal attention module, a feature fusion layer, and an adaptive node selection pooling layer. The spatial attention module takes the feature association graph as input and adaptively learns the attention weight of each node through a multi-head attention mechanism to enhance the contribution of core fault feature nodes and insulation state constraint nodes, and suppress redundant features and interference nodes. The time attention module takes the time evolution graph as input and focuses on the key time step nodes of fault evolution through a multi-head attention mechanism to capture the long-term temporal dependencies of faults at different rates. The feature fusion layer uses a weighted concatenation strategy to fuse the output features of the spatial attention module and the temporal attention module, and maps the fused features into a high-dimensional feature vector through a fully connected layer to provide support for fault classification. The adaptive node selection pooling layer filters key nodes and removes noisy nodes based on the contribution score of nodes to fault classification, and extracts multi-scale graph features by combining local max pooling and global mean pooling.

8. The tap changer fault diagnosis method based on mechanistic dual-graph and time-frequency hierarchical fusion according to claim 7, characterized in that, In step S4, the formulas for calculating the attention weights of the spatial attention module and the temporal attention module are as follows: The formula for calculating node feature updates is: in, Let be the attention weight of node i to node j, be the attention weight vector, and be the learnable weight matrix. and Let be the feature vectors of node i and node j, respectively. Let i be the neighborhood set of node i The feature vector of any neighbor node k in the traversal, where k is the feature vector of the traversal. The index variable of all neighboring nodes; || represents the feature concatenation operation. Let σ be the set of neighboring nodes of node i, and let σ be the activation function. This refers to the updated node features.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.

10. A tap changer fault diagnosis system based on mechanistic dual-graph and time-frequency hierarchical fusion, comprising an interconnected memory and a processor, wherein the memory stores a computer program, characterized in that, The computer program, when run by a processor, performs the steps of the method as described in any one of claims 1-8.

Images