An online credible fault identification method for oil-immersed transformer
Patent Information
- Application Number
- CN202610993387.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-06
- Publication Date
- 2026-09-29
AI Technical Summary
现有研究通常采用固定邻接矩阵表征气体关联关系，但不同故障类型下气体关联模式存在显著差异，单一固定图结构无法反映这种差异性；而且不同工况和故障阶段的具体样本，其气体关联特征可能偏离整体统计规律，固定图结构无法针对单个样本自适应调整
[0053]本发明通过标准化完成原始DGA时序数据预处理，统一不同特征气体的数据量纲与数值分布，消除各类气体浓度跨度悬殊带来的特征权重失衡问题，保障后续模型能够均等关键故障气体的浓度变化规律，从数据源头提升特征利用有效性。搭建多类故障原型初始邻接矩阵并绑定可学习键向量，摒弃传统固定不变的气体关联拓扑，借助注意力机制依据单样本实际产气特征自适应加权融合各类原型矩阵生成动态邻接矩阵，有效解决固定拓扑无法适配过渡故障、复合型故障的技术短板，大幅提升气体空间关联特征提取的贴合度。图卷积网络（Graph Convolutional Network, GCN）依托动态邻接矩阵挖掘气体间空间耦合特征，时序卷积网络（Temporal Convolutional Network, TCN）采用带残差的因果膨胀卷积提取长时序演化特征，二者融合实现故障时空特征的全方位挖掘，既捕捉短时浓度突变细节，又挖掘数月尺度缓慢劣化的长期变化规律，弥补单一卷积网络只能提取局部特征的缺陷。运维人员可依据不确定性数值筛选疑难样本开展人工复核，从机制上降低误判、漏判概率。整体方案兼顾先验机理与数据驱动优势，适配变压器在线运维场景，有效提升油浸式变压器故障在线辨识的可靠性与工程实用性。
Smart Images
Figure CN122839110A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of transformer technology, and more specifically, to a reliable online fault identification method for oil-immersed transformers. Background Technology
[0002] Oil-immersed transformers are core equipment in power systems, and their safe and stable operation directly affects the reliable power supply and economical operation of the power grid. During operation, the internal insulation materials of the transformer gradually deteriorate and decompose under thermal stress, electrical stress, and chemical action, producing characteristic gases such as hydrogen, methane, and acetylene. Dissolved gas analysis (DG) in the oil is crucial for this process. Detecting the concentration and changing trends of these gases to determine the type and severity of internal transformer faults is one of the most widely used methods for transformer fault detection.
[0003] Traditional based on DG Transformer fault identification methods such as the IEC three-ratio method and the Rogers ratio method rely on the gas concentration ratio or interval threshold rules on a single time section for judgment. On the one hand, it is difficult to capture the coupled evolution of multiple characteristic gases in the time dimension, and the identification accuracy is insufficient in compound fault or early fault scenarios. On the other hand, the judgment boundary is fixed, and it is easy to make misjudgments when faced with continuous changes in gas concentration and complex operating conditions.
[0004] With the development of deep learning technology, the introduction of models such as CNN and LSTM has improved the performance of DG-based models to some extent. While existing methods improve fault identification accuracy, they often treat each gas channel as an independent parallel sequence, lacking explicit modeling of the topological relationships between gases. Current research typically uses a fixed adjacency matrix to represent gas relationships, but gas correlation patterns differ significantly under different fault types, and a single fixed graph structure cannot reflect these differences. Furthermore, the gas correlation characteristics of specific samples under different operating conditions and fault stages may deviate from the overall statistical regularity, and a fixed graph structure cannot adaptively adjust for individual samples.
[0005] Furthermore, existing deep learning models typically output failure category probabilities, but these probabilities do not accurately reflect the model's cognitive uncertainty. When encountering out-of-distribution anomalous samples or difficult-to-distinguish boundary samples, the model may still output high-confidence false identifications, posing safety risks to operational decisions. Summary of the Invention
[0006] To address the shortcomings of existing technologies, the present invention aims to provide a reliable online fault identification method for oil-immersed transformers.
[0007] To achieve the above objectives, the present invention provides the following technical solution:
[0008] A reliable online fault identification method for oil-immersed transformers includes the following steps:
[0009] The concentration time series data of dissolved gases in transformer oil were obtained, and the time series data were preprocessed to obtain time series processed data.
[0010] Construct a fault prototype adjacency matrix library, with each type of fault corresponding to an initial adjacency matrix and a learnable key vector;
[0011] A query vector is generated based on the global features of the sample to be identified. The attention weights of the query vector and the key vectors corresponding to various faults are calculated. The initial adjacency matrix is weighted and fused using the attention weights to obtain the adaptive dynamic adjacency matrix of the current sample.
[0012] Based on the adaptive dynamic adjacency matrix, graph convolution operation is performed on the gas features at each time step in the time series data to obtain a spatially correlated feature sequence;
[0013] Spatial correlation feature sequences are input into a temporal convolutional network to extract long-term evolution features, resulting in spatiotemporal fusion features;
[0014] Based on the spatiotemporal fusion characteristics, the fault category prediction results and corresponding identification uncertainties are output.
[0015] Preferably, the time-series data is preprocessed to obtain time-series processed data, specifically as follows:
[0016] Time series data are categorized by feature dimensions. Standardization processes yield time-series processed data.
[0017] Preferably, for the first For each type of fault, an initial adjacency matrix is constructed based on the statistical characteristics of gas concentration in the fault samples. ; where the matrix's first ; The element represents the first element. Class of failure modes The gas and the first The correlation strength between the gases; =1,2,..., , Total number of fault categories; These are the characteristic numbers of the gas.
[0018] Associate a learnable key vector with the initial adjacency matrix for each type of fault. This is used for subsequent matching calculations with the query vector of the sample to be identified; where, The key vector dimension is used as a learnable parameter and is jointly optimized with other network parameters during model training. Specifically, each key vector is initialized uniformly using Glorot and iteratively updated during end-to-end training by obtaining gradient signals from the final identification loss function through backpropagation. After sufficient training, each key vector will converge to the feature anchor position of the corresponding fault category in the attention space, maximizing the response of the sample query vector belonging to that fault category to its dot product.
[0019] Preferably, a query vector is generated based on the global features of the sample to be identified, specifically as follows:
[0020] Using feature extractor Extract the global state features of the sample to be identified;
[0021] Global features are projected onto the key vector semantic space through linear mapping to obtain the sample to be identified. query vector The calculation formula is: ;
[0022] in For feature extractors, The weight matrix is the linear mapping. This is the bias vector. and As learnable parameters, they are jointly optimized with the network with the aim of projecting the global statistical features of the samples into the same semantic space as the key vector, so that the query vector can accurately reflect the direction of the fault mode matched by the current sample.
[0023] Preferably, the attention weights between the query vector and the key vectors corresponding to various fault types are calculated, specifically as follows:
[0024] Calculate the query vector of the sample to be identified Key vectors of the initial adjacency matrices of various faults Attention score between scaled dot products: ;
[0025] in, This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to excessively large dot product values;
[0026] Softmax normalization is applied to the attention score to obtain the attention weights of the sample and each initial adjacency matrix of the faults:
[0027]
[0028] Preferably, the initial adjacency matrix is weighted and fused using attention weights to obtain the adaptive dynamic adjacency matrix of the current sample to be identified, specifically as follows:
[0029] Using attention weights The initial adjacency matrix of the fault We obtain the adaptive dynamic adjacency matrix by weighted summation. ;
[0030] in, It is an adaptive dynamic adjacency matrix that can reflect the gas correlation characteristics of the fault mode matched by the current sample.
[0031] Preferably, graph convolution operations are performed on the gas features at each time step in the time-series processed data based on an adaptive dynamic adjacency matrix to obtain a spatially correlated feature sequence, specifically:
[0032] Treatment of identification samples At time step Gas eigenvectors The graph convolution operation is defined as: ;
[0033] in, The normalized adjacency matrix, Add a self-loop to the adjacency matrix of this sample. for The degree matrix; The learnable weight matrix for the graph convolutional layer; It is a non-linear activation function. After graph convolution, This reflects the time step of each gas node. Spatial dependencies.
[0034] Sample to be identified Within the time window Perform graph convolution operations on the gas features at each time step, and concatenate the outputs of each time step along the time dimension to obtain a spatially correlated feature sequence. .
[0035] Preferably, the spatial correlation feature sequence is input into a temporal convolutional network to extract long-term evolution features, resulting in spatiotemporal fusion features, specifically:
[0036] Spatial correlation feature sequences As the input of the first layer of the temporal convolutional network , its first The layer output formula is:
[0037] ;
[0038] in, Indicates sample In the temporal convolutional network, the... Features of the layer output; For the first Layer convolution kernel parameters; For the first Layer bias term; For the first Layer expansion coefficient; Indicates the expansion factor as The causal dilation convolution operation; For activation functions; This indicates a residual connection.
[0039] Preferably, it further includes:
[0040] For the first time sequence convolutional network Layer output Global average pooling is performed in the time dimension to obtain spatiotemporal fusion features:
[0041] ;in For the first Layer in time step The output vector.
[0042] Preferably, the fault category prediction result and the corresponding identification uncertainty are output based on the spatiotemporal fusion features, specifically:
[0043] The feature vector output by the temporal convolutional network Evidence values are mapped to each fault category through a fully connected layer:
[0044] in, and These are the parameters for the fully connected layer; The activation function ensures that the evidence values are non-negative. Indicates sample Belongs to the The amount of evidence for this type of fault;
[0045] Transform the evidence values into the concentration parameters of a Dirichlet distribution:
[0046]
[0047] Based on the Dirichle distribution parameters, the prediction probability and identification uncertainty for each fault category are calculated as follows:
[0048]
[0049]
[0050] in, The total strength of evidence for the Dirichlet distribution; For the sample Belongs to the Predicted probability of type of failure; For the sample The uncertainty in identification. The more evidence the model collects, the better. The larger the uncertainty The smaller the value, the lower the threshold; when the model lacks knowledge of the input samples and all types of evidence are low. Approaching 1.
[0051] The model ultimately uses the category corresponding to the highest predicted probability as the identification result, i.e. and with This serves as a reliability indicator for the identification result. For uncertainties exceeding a preset threshold... The sample (i.e.) The system automatically marks these samples as high-uncertainty samples that require manual verification, thus balancing identification efficiency and reliability in practical engineering applications.
[0052] Compared with the prior art, the present invention has the following beneficial effects:
[0053] This invention is achieved through Standardized preprocessing of raw DGA time-series data unifies the data dimensions and numerical distribution of different characteristic gases, eliminating the feature weight imbalance caused by the large differences in the concentration range of various gases. This ensures that subsequent models can equally reflect the concentration change patterns of key fault gases, improving the effectiveness of feature utilization from the data source. Initial adjacency matrices for multiple fault prototypes are constructed and bound to learnable key vectors. Abandoning the traditional fixed gas association topology, an attention mechanism is used to adaptively weight and fuse various prototype matrices based on the actual gas production characteristics of a single sample to generate a dynamic adjacency matrix. This effectively solves the technical shortcomings of fixed topologies in adapting to transitional and complex faults, significantly improving the fit of gas spatial association feature extraction. The Graph Convolutional Network (GCN) relies on the dynamic adjacency matrix to mine spatial coupling features between gases, while the Temporal Convolutional Network (TCN) uses residual causal dilation convolution to extract long-term temporal evolution features. The fusion of these two achieves comprehensive mining of fault spatiotemporal features, capturing both short-term concentration abrupt changes and long-term changes in slow degradation over several months, overcoming the limitation of single convolutional networks that can only extract local features. Maintenance personnel can manually review complex samples based on uncertain values, thereby reducing the probability of misjudgments and omissions. The overall solution combines the advantages of prior mechanisms and data-driven approaches, adapting to online transformer maintenance scenarios and effectively improving the reliability and engineering practicality of online fault identification for oil-immersed transformers. Attached Figure Description
[0054] Figure 1This is a general flowchart of an embodiment of the present invention;
[0055] Figure 2 This is a schematic diagram illustrating the construction of the fault prototype adjacency matrix library according to an embodiment of the present invention;
[0056] Figure 3 This is a schematic diagram illustrating the adaptive dynamic adjacency matrix generation according to an embodiment of the present invention;
[0057] Figure 4 This is a schematic diagram of time-step graph convolution feature extraction based on a dynamic adjacency matrix according to an embodiment of the present invention;
[0058] Figure 5 This is a schematic diagram of the spatiotemporal feature network fusion structure according to an embodiment of the present invention; Detailed Implementation
[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0060] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0061] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.
[0062] Reference Figures 1-5 As shown.
[0063] The embodiments further illustrate the online reliable fault identification method for oil-immersed transformers proposed in this invention.
[0064] A reliable online fault identification method for oil-immersed transformers includes the following steps:
[0065] Time-series data of dissolved gas concentrations in transformer oil were obtained, and the time-series data were preprocessed to obtain processed time-series data, specifically as follows:
[0066] Data Acquisition and Preprocessing: Dissolved gas analysis (DGA) data in oil-immersed transformer oil was acquired, containing a total of 3000 samples. Each sample has a dimension of 420×4, where the time window length is... =420, corresponding to 420 equally spaced gas concentration records, with a sampling interval of 12 hours between adjacent records; the number of gas features N=4, namely H2, CO, C2H4, and C2H2. Each sample corresponds to a fault category label y∈{1,2,3,4}, for a total of C=4 fault categories: normal mode, partial discharge, low-energy discharge, and low-temperature overheating. The dataset is divided into a training set (2100 samples) and a test set (900 samples) in a 7:3 ratio. The time-series data is then processed according to the feature dimension. Standardization processes yield time-series processed data;
[0067] To introduce prior knowledge of fault types into graph convolution, a fault prototype adjacency matrix library is constructed. In this embodiment, each fault type corresponds to an initial adjacency matrix and a learnable key vector, forming a fault prototype graph library containing four prototype graphs. Each fault type corresponds to an initial adjacency matrix and a learnable key vector.
[0068] For the For each type of fault, an initial adjacency matrix is constructed based on the statistical characteristics of gas concentration in the fault samples. ; where the matrix's first ; The element represents the first element. Class 1 Fault Mode The gas and the first The correlation strength between the gases; =1,2,..., , Total number of fault categories; These are the characteristic numbers of the gas.
[0069] Associate a learnable key vector with the initial adjacency matrix for each type of fault. This is used for subsequent matching calculations with the query vector of the sample to be identified; a learnable key vector is associated with the initial adjacency matrix of each type of fault. Key vector dimension ;in, The key vector dimension is used as a learnable parameter and is jointly optimized with other network parameters during model training. Specifically, each key vector is initialized uniformly using Glorot and iteratively updated during end-to-end training by obtaining gradient signals from the final identification loss function through backpropagation. After sufficient training, each key vector will converge to the feature anchor position of the corresponding fault category in the attention space, maximizing the response of the sample query vector belonging to that fault category to its dot product.
[0070] A query vector is generated based on the global features of the sample to be identified, specifically:
[0071] Using feature extractor Extract global state features of the sample to be identified
[0072] For the current sample Through feature extractor Extracting global state features. In this embodiment, the feature extractor... Using the sample matrix in the time dimension ( Global average pooling is performed on (=420) to obtain the global state features of the sample to be identified, and then a query vector is obtained through linear mapping. :
[0073] Global features are projected onto the key vector semantic space through linear mapping to obtain the sample to be identified. query vector The calculation formula is: ;
[0074] in For feature extractors, The weight matrix is the linear mapping. This is the bias vector. and As learnable parameters, they are jointly optimized with the network with the aim of projecting the global statistical features of the samples into the same semantic space as the key vector, so that the query vector can accurately reflect the direction of the fault mode matched by the current sample.
[0075] The attention weights for the query vector and the corresponding key vectors for each type of fault are calculated as follows:
[0076] Calculate the query vector of the sample to be identified Key vectors of the initial adjacency matrices of various faults Attention score between scaled dot products: ;
[0077] in, This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to excessively large dot product values;
[0078] Softmax normalization is applied to the attention score to obtain the attention weights of the sample and each initial adjacency matrix of the faults:
[0079] .
[0080] The initial adjacency matrix is weighted and fused using attention weights to obtain the adaptive dynamic adjacency matrix for the current sample, specifically:
[0081] Using attention weights The initial adjacency matrix of the fault We obtain the adaptive dynamic adjacency matrix by weighted summation. ;
[0082] in, It is an adaptive dynamic adjacency matrix that can reflect the gas correlation characteristics in the fault mode matched by the current sample;
[0083] Based on an adaptive dynamic adjacency matrix, graph convolution operations are performed on the gas features at each time step in the time-series data to obtain a spatially correlated feature sequence, specifically:
[0084] Using the generated sample adaptive dynamic adjacency matrix, graph convolution operation is performed on the gas feature vector of each time step within the time window of the sample to capture the spatial correlation features between gases under the current fault mode.
[0085] Treatment of identification samples At time step Gas eigenvectors For the sample At time step Gas eigenvectors The graph convolution operation is defined as: ;
[0086] in, The normalized adjacency matrix, Add a self-loop to the adjacency matrix of this sample. for The degree matrix; The learnable weight matrix for the graph convolutional layer; The ReLU activation function is used. After graph convolution, This reflects the time step of each gas node. Spatial dependencies;
[0087] Sample to be identified Within the time window Graph convolution operations are performed on the gas features at each time step. The outputs of each time step are concatenated along the time dimension to obtain a spatially correlated feature sequence. .
[0088] Spatial correlation feature sequences are input into a temporal convolutional network to extract long-term evolution features, resulting in spatiotemporal fusion features, specifically:
[0089] Long-term gas evolution feature extraction based on temporal convolutional networks. Spatial correlation feature sequences are fed into a temporal convolutional network (TCN), and the long-term temporal evolution features of the DGA signal are extracted using a causal dilated convolution mechanism. In this embodiment, the TCN is configured with... The convolutional layers have a uniform kernel size of 3, and the dilation coefficients of each layer are as follows: , , , .
[0090] Spatial correlation feature sequences As the input of the first layer of the temporal convolutional network , its first The layer output formula is:
[0091] ;
[0092] in, Indicates sample In the temporal convolutional network, the... Features of the layer output; For the first Layer convolution kernel parameters; For the first Layer bias term; For the first Layer expansion coefficient; Indicates the expansion factor as The causal dilation convolution operation; Use the ReLU activation function; Indicates residual connection;
[0093] By stacking four layers of causal dilated convolutions, TCN can exponentially expand its receptive field without significantly increasing the number of parameters, thereby simultaneously capturing both short-term gas fluctuation characteristics and long-term gas evolution trends. This is further demonstrated by the first layer of temporal convolutional networks. Layer output Global average pooling is performed in the time dimension to obtain spatiotemporal fusion features: ;in For the first Layer in time step The output vector.
[0094] Suppose The output after the TCN layer is Global average pooling is performed on the samples over the time dimension to obtain the results. The final spatiotemporal fusion characteristics:
[0095]
[0096] in, For the first Layer in time step The output vector.
[0097] Based on the spatiotemporal fusion features, the fault category prediction results and corresponding identification uncertainties are output, specifically:
[0098] The feature vector output by the temporal convolutional network Evidence values are mapped to each fault category through a fully connected layer:
[0099] in, and These are the parameters for the fully connected layer; The activation function ensures that the evidence values are non-negative. Indicates sample Belongs to the The amount of evidence for this type of fault;
[0100] Transform the evidence values into the concentration parameters of a Dirichlet distribution:
[0101]
[0102] Based on the Dirichle distribution parameters, the prediction probability and identification uncertainty for each fault category are calculated as follows:
[0103]
[0104]
[0105] in, The total strength of evidence for the Dirichlet distribution; For the sample Belongs to the Predicted probability of type of failure; For the sample The uncertainty in identification. The more evidence the model collects, the better. The larger the uncertainty The smaller the value, the lower the threshold; when the model lacks knowledge of the input samples and all types of evidence are low. Approaching 1.
[0106] The model ultimately uses the category corresponding to the highest predicted probability as the identification result, i.e. and with This serves as a reliability indicator for the identification result. For uncertainties exceeding a preset threshold... The sample (i.e.) The system automatically marks these samples as high-uncertainty samples that require manual verification, thus balancing identification efficiency and reliability in practical engineering applications.
[0107] The model was optimized using the Adam optimizer with an initial learning rate of 0.001 and a batch size of 64. Training lasted for 150 epochs, and an early stopping strategy (patience set to 15 epochs) was employed to prevent overfitting. Model performance metrics included accuracy, precision, recall, and F1 score. Experiments demonstrated that the proposed method significantly outperformed traditional DGA-based fault identification methods on the test set in terms of accuracy and precision. Furthermore, the model effectively identified high-uncertainty samples in the test set, and excluding these samples further improved identification accuracy, validating the effectiveness of the uncertainty quantification mechanism in enhancing identification reliability.
[0108] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A reliable online fault identification method for oil-immersed transformers, characterized in that, Includes the following steps: The concentration time series data of dissolved gases in transformer oil were obtained, and the time series data were preprocessed to obtain time series processed data. Construct a fault prototype adjacency matrix library, with each type of fault corresponding to an initial adjacency matrix and a learnable key vector; A query vector is generated based on the global features of the sample to be identified. The attention weights of the query vector and the key vectors corresponding to various faults are calculated. The initial adjacency matrix is weighted and fused using the attention weights to obtain the adaptive dynamic adjacency matrix of the current sample. Based on the adaptive dynamic adjacency matrix, graph convolution operation is performed on the gas features at each time step in the time series data to obtain a spatially correlated feature sequence; Spatial correlation feature sequences are input into a temporal convolutional network to extract long-term evolution features, resulting in spatiotemporal fusion features; Based on the spatiotemporal fusion characteristics, the fault category prediction results and corresponding identification uncertainties are output.
2. The online reliable fault identification method for oil-immersed transformers according to claim 1, characterized in that, Preprocessing the time series data yields the processed time series data, specifically: Time series data are processed according to feature dimensions. Standardization processes yield time-series processed data.
3. The online reliable fault identification method for oil-immersed transformers according to claim 1, characterized in that, For the For each type of fault, an initial adjacency matrix is constructed based on the statistical characteristics of gas concentration in the fault samples. ; where the matrix's first ; The element represents the first element. Class of failure modes The gas and the first The correlation strength between the gases; =1,2,..., , Total number of fault categories; These are the characteristic numbers of the gas. Associate a learnable key vector with the initial adjacency matrix for each type of fault. ;in, The dimension of the key vector.
4. The online reliable fault identification method for oil-immersed transformers according to claim 3, characterized in that, A query vector is generated based on the global features of the sample to be identified, specifically: Using feature extractor Extract the global state features of the sample to be identified; Global features are projected onto the key vector semantic space through linear mapping to obtain the sample to be identified. query vector The calculation formula is: ; in, For feature extractors, The weight matrix is the linear mapping. This is the bias vector.
5. The online reliable fault identification method for oil-immersed transformers according to claim 4, characterized in that, The attention weights for the query vector and the corresponding key vectors for each type of fault are calculated as follows: Calculate the query vector of the sample to be identified Key vectors of the initial adjacency matrices of various faults Attention score between scaled dot products: ; in, This is a scaling factor used to prevent the gradient of the Softmax function from vanishing due to excessively large dot product values; Softmax normalization is applied to the attention score to obtain the attention weights of the sample and each initial adjacency matrix of the faults:
6. The online reliable fault identification method for oil-immersed transformers according to claim 5, characterized in that, The initial adjacency matrix is weighted and fused using attention weights to obtain the adaptive dynamic adjacency matrix of the current sample to be identified, specifically as follows: Using attention weights The initial adjacency matrix of the fault We obtain the adaptive dynamic adjacency matrix by weighted summation. ; in, It is an adaptive dynamic adjacency matrix that can reflect the gas correlation characteristics of the fault mode matched by the current sample.
7. The online reliable fault identification method for oil-immersed transformers according to claim 6, characterized in that, Based on an adaptive dynamic adjacency matrix, graph convolution operations are performed on the gas features at each time step in the time-series data to obtain a spatially correlated feature sequence, specifically: Treatment of identification samples At time step Gas eigenvectors The graph convolution operation is defined as follows: ; in, The normalized adjacency matrix, Add a self-loop to the adjacency matrix of this sample. for The degree matrix; The learnable weight matrix for the graph convolutional layer; It is a non-linear activation function; Sample to be identified Within the time window Perform graph convolution operations on the gas features at each time step, and concatenate the outputs of each time step along the time dimension to obtain a spatially correlated feature sequence. .
8. The online reliable fault identification method for oil-immersed transformers according to claim 7, characterized in that, Spatial correlation feature sequences are input into a temporal convolutional network to extract long-term evolution features, resulting in spatiotemporal fusion features, specifically: Spatial correlation feature sequences As the input of the first layer of the temporal convolutional network , its first The layer output formula is: ； in, Indicates sample In the temporal convolutional network, the... Features of the layer output; For the first Layer convolution kernel parameters; For the first Layer bias term; For the first Layer expansion coefficient; Indicates the expansion factor as The causal dilation convolution operation; For activation functions; This indicates a residual connection.
9. The online reliable fault identification method for oil-immersed transformers according to claim 8, characterized in that, Also includes: For the first time-series convolutional network Layer output Global average pooling is performed in the time dimension to obtain spatiotemporal fusion features: ;in For the first Layer in time step The output vector.
10. The online reliable fault identification method for oil-immersed transformers according to claim 9, characterized in that, Based on the spatiotemporal fusion features, the fault category prediction results and corresponding identification uncertainties are output, specifically: The feature vector output by the temporal convolutional network Evidence values are mapped to each fault category through a fully connected layer: in, and These are the parameters for the fully connected layer; The activation function ensures that the evidence values are non-negative. Indicates sample Belongs to the The amount of evidence for this type of fault; Transform the evidence values into the concentration parameters of a Dirichlet distribution: Based on the Dirichlet distribution parameters, the prediction probability and identification uncertainty for each fault category are calculated as follows: in, The total strength of evidence for the Dirichlet distribution; For the sample Belongs to the Predicted probability of type of failure; For the sample Uncertainty in identification.