Single-phase transformer monitoring method and system

Abstract

The invention discloses a single-phase transformer monitoring method and system, and relates to the technical field of transformer monitoring, a working condition feature set is adopted to train a preset transformer fault detection model, a corresponding target transformer fault detection model is obtained, the target transformer fault detection model comprises an extraction network and a hierarchical prediction network, and the extraction network and the hierarchical prediction network are used to predict the transformer fault. Working condition parameters of the to-be-tested single-phase transformer are obtained, multi-modal feature extraction is carried out on the working condition parameters through an extraction network, and corresponding electrical quantity feature vectors, partial discharge waveform feature vectors and physical state feature vectors are obtained; and performing fault detection on the electrical quantity feature vector, the partial discharge waveform feature vector and the physical state feature vector through a hierarchical prediction network to obtain a corresponding monitoring result. The technical problems that a traditional single-phase transformer monitoring method mostly depends on manual inspection, the requirements for monitoring real-time performance and accuracy in operation are difficult to meet, and the operation reliability of the single-phase transformer is reduced are solved.

Description

Technical Field

[0001] This invention relates to the field of transformer monitoring technology, and in particular to a method and system for monitoring single-phase transformers. Background Technology

[0002] With the continuous growth of electricity load, the large-scale integration of new energy power generation, and the increasing complexity of the power grid structure, the operating environment of single-phase transformers is becoming increasingly harsh. They operate under high loads and alternating operating conditions for extended periods, facing various potential fault risks such as insulation aging, partial discharge, and overheating losses. Once a transformer fails, it can not only damage the equipment itself but also trigger a chain reaction of large-scale power outages and production interruptions, causing significant economic losses and social impact. Therefore, real-time and accurate condition monitoring and fault diagnosis of single-phase transformers are crucial for ensuring the safe and stable operation of the power system.

[0003] Currently, traditional single-phase transformer monitoring methods mostly rely on threshold judgment of single operating condition parameters or manual inspection, but the rate of false fault diagnosis and missed fault diagnosis is high, making it difficult to meet the requirements of real-time and accurate monitoring during operation, thus reducing the reliability of single-phase transformer operation. Summary of the Invention

[0004] This invention provides a single-phase transformer monitoring method and system, which solves the technical problem that traditional single-phase transformer monitoring methods rely on threshold judgment of single operating condition parameters or manual inspection, but have a high rate of false fault diagnosis and missed fault diagnosis, making it difficult to meet the requirements of real-time and accurate monitoring during operation, and reducing the reliability of single-phase transformer operation.

[0005] The first aspect of this invention provides a method for monitoring a single-phase transformer, comprising:

[0006] Multiple training condition parameters are obtained, and each training condition parameter is preprocessed to obtain the corresponding condition feature set.

[0007] Based on a preset multi-task loss function, the preset transformer fault detection model is trained using the operating condition feature set to obtain the corresponding target transformer fault detection model, wherein the target transformer fault detection model includes an extraction network and a hierarchical prediction network.

[0008] The operating parameters of the single-phase transformer under test are obtained, and the operating parameters are subjected to multimodal feature extraction through the extraction network to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector.

[0009] The hierarchical prediction network is used to perform fault detection on the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector to obtain the corresponding monitoring results.

[0010] Optionally, the step of training a preset transformer fault detection model using the operating condition feature set based on a preset multi-task loss function to obtain a corresponding target transformer fault detection model includes:

[0011] The preset transformer fault detection model is trained using the aforementioned operating condition feature set to obtain corresponding training detection data;

[0012] The training detection data and the working condition feature set are input into a preset multi-task loss function to obtain the corresponding loss function value;

[0013] When the loss function value is greater than or equal to the preset loss threshold, the network parameters of the transformer fault detection model are adjusted until the loss function value is less than the loss threshold.

[0014] When the value of the loss function is less than the loss threshold, a corresponding target transformer fault detection model is generated.

[0015] Optionally, the extraction network includes an electrical quantity feature extraction branch, a partial discharge waveform feature extraction branch, and a physical state feature extraction branch. The operating condition parameters include electrical quantity parameters, partial discharge parameters, and physical quantity parameters. The step of performing multimodal feature extraction on the operating condition parameters through the extraction network to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector, and physical state feature vector includes:

[0016] The electrical quantity feature extraction branch is used to extract electrical quantity features from the electrical quantity parameters to obtain the corresponding electrical quantity feature vectors;

[0017] The partial discharge parameters are extracted by the partial discharge waveform feature extraction branch to obtain the corresponding partial discharge waveform feature vector.

[0018] The physical state feature extraction branch adaptively adjusts the weights of the physical quantity parameters to obtain the corresponding physical state feature vector.

[0019] Optionally, the electrical quantity feature extraction branch includes a first causal convolutional block, a second causal convolutional block, a third causal convolutional block, a global average pooling layer, and a multi-head temporal attention layer. The step of extracting electrical quantity features from the electrical quantity parameters through the electrical quantity feature extraction branch to obtain the corresponding electrical quantity feature vector includes:

[0020] Multi-scale feature extraction is performed on the electrical quantity parameters by sequentially connecting the first causal convolutional block, the second causal convolutional block, and the third causal convolutional block to obtain the corresponding multi-dimensional electrical quantity feature vector;

[0021] The multidimensional electrical quantity feature vector is reduced in dimensionality by a global average pooling layer to obtain the corresponding initial electrical quantity feature vector.

[0022] The initial electrical quantity feature vector is extracted by a multi-head timing attention layer to obtain the corresponding electrical quantity feature vector.

[0023] Optionally, the partial discharge waveform feature extraction branch includes a first convolutional layer, a second convolutional layer, a global max pooling layer, and a Transformer encoder. The step of extracting partial discharge features from the partial discharge parameters through the partial discharge waveform feature extraction branch to obtain the corresponding partial discharge waveform feature vector includes:

[0024] The local feature extraction of the partial discharge parameters is performed by the first convolutional layer to obtain the corresponding first partial discharge feature vector. The first convolutional layer includes a 3×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence.

[0025] The first partial discharge feature vector is subjected to global feature extraction through the second convolutional layer to obtain the corresponding second partial discharge feature vector. The second convolutional layer includes a 5×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence.

[0026] The second partial discharge feature vector is reduced in dimensionality by a global max pooling layer to obtain the corresponding third partial discharge feature vector.

[0027] The key features of the third partial discharge feature vector are extracted by the Transformer encoder to obtain the corresponding partial discharge waveform feature vector.

[0028] Optionally, the physical state feature extraction branch includes an adaptive weighting module and a perception module. The step of adaptively adjusting the weights of the physical quantity parameters through the physical state feature extraction branch to obtain the corresponding physical state feature vector includes:

[0029] The physical quantity parameters are subjected to feature weighting processing by the adaptive weighting module to obtain the corresponding weighted physical feature vector.

[0030] The perception module performs nonlinear feature extraction on the weighted physical feature vector to obtain the corresponding physical state feature vector. The perception module includes a first hidden layer, a second hidden layer, and a normalization layer connected in sequence.

[0031] Optionally, the hierarchical prediction network includes a multimodal feature fusion layer, a first detection head, and a second detection head. The step of performing fault detection on the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector through the hierarchical prediction network to obtain the corresponding monitoring results includes:

[0032] The electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector are weighted and fused through a multimodal feature fusion layer to obtain the corresponding fused feature vector.

[0033] The first detection head performs fault detection on the fused feature vector to obtain the corresponding fault type.

[0034] When the fault type is normal, the monitoring result is determined to be in normal operation;

[0035] When the fault type is not normal, the second detection head is used to evaluate the fused feature vector to obtain the corresponding fault level.

[0036] The fault level and the fault type are determined as the corresponding monitoring results.

[0037] A second aspect of the present invention provides a single-phase transformer monitoring system, comprising:

[0038] The acquisition module is used to acquire multiple training working condition parameters, preprocess each of the training working condition parameters, and obtain the corresponding working condition feature set.

[0039] The training module is used to train the preset transformer fault detection model based on the preset multi-task loss function and the working condition feature set to obtain the corresponding target transformer fault detection model. The target transformer fault detection model includes an extraction network and a hierarchical prediction network.

[0040] The extraction module is used to obtain the operating parameters of the single-phase transformer under test, and to extract multi-modal features from the operating parameters through the extraction network to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector.

[0041] The detection module is used to perform fault detection on the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector through the hierarchical prediction network, and obtain the corresponding monitoring results.

[0042] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the single-phase transformer monitoring method described above.

[0043] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed, implements the single-phase transformer monitoring method as described above.

[0044] As can be seen from the above technical solutions, the present invention has the following advantages:

[0045] This invention trains a pre-defined transformer fault detection model using a preset multi-task loss function and a set of operating condition features to obtain a corresponding target transformer fault detection model. The target transformer fault detection model includes an extraction network and a hierarchical prediction network. It acquires the operating condition parameters of the single-phase transformer under test, extracts multi-modal features from these parameters through the extraction network to obtain corresponding electrical quantity feature vectors, partial discharge waveform feature vectors, and physical state feature vectors, and then uses the hierarchical prediction network to perform fault detection on these vectors to obtain the corresponding monitoring results. This overcomes the technical problem that traditional single-phase transformer monitoring methods often rely on threshold judgments of single operating condition parameters or manual inspections, which are insufficient to meet the requirements for real-time and accurate monitoring during operation, thus reducing the reliability of single-phase transformer operation. Compared with traditional single-phase transformer monitoring methods, this invention significantly improves the accuracy of single-phase transformer fault detection by using a target transformer fault detection model for multi-modal feature extraction and fault detection of operating conditions. Furthermore, it can adapt to different operating conditions, improving the reliability of single-phase transformer operation. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the steps of a single-phase transformer monitoring method provided in Embodiment 1 of the present invention.

[0048] Figure 2 This is a flowchart illustrating the steps of a single-phase transformer monitoring method provided in Embodiment 2 of the present invention.

[0049] Figure 3 This is a schematic diagram of the structure of the target transformer fault detection model provided in Embodiment 2 of the present invention;

[0050] Figure 4 This is a structural block diagram of a single-phase transformer monitoring system provided in Embodiment 3 of the present invention;

[0051] Figure 5 This is a structural block diagram of an electronic device provided in Embodiment 4 of the present invention. Detailed Implementation

[0052] This invention provides a single-phase transformer monitoring method and system to address the technical problem that traditional single-phase transformer monitoring methods rely heavily on threshold judgment of single operating parameters or manual inspection, which results in high rates of false faults and missed faults, making it difficult to meet the requirements for real-time and accurate monitoring during operation and reducing the reliability of single-phase transformer operation.

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that in the optional embodiments of the present invention, the object information and other related data involved require the permission or consent of the object when the embodiments of the present invention are applied to specific products or technologies, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. That is to say, if the embodiments of the present invention involve data related to the object, it needs to be obtained with the authorization and consent of the object, the authorization and consent of the relevant departments, and in compliance with the relevant laws, regulations, and standards of the country and region. If personal information is involved in the embodiments, the acquisition of all personal information requires the consent of the individual. If sensitive information is involved, the separate consent of the information subject is required, and the embodiments also need to be implemented with the authorization and consent of the object.

[0054] Please see Figure 1 , Figure 1 This is a flowchart illustrating the steps of a single-phase transformer monitoring method provided in Embodiment 1 of the present invention.

[0055] This invention provides a single-phase transformer monitoring method, comprising:

[0056] Step 101: Obtain multiple training condition parameters, preprocess each training condition parameter, and obtain the corresponding condition feature set.

[0057] Training operating condition parameters refer to the various data sets used to train the transformer fault detection model, reflecting different operating states of a single-phase transformer, including but not limited to high-voltage side voltage / current, low-voltage side voltage / current, top oil temperature, oil level, ambient temperature / humidity, UHF partial discharge waveform, discharge amplitude, as well as operating time, load rate, and historical fault records.

[0058] The operating condition feature set refers to the set of training operating condition parameters that, after preprocessing, can effectively characterize different operating conditions of a transformer.

[0059] In this embodiment of the invention, multiple training condition parameters are obtained, and each training condition parameter is subjected to missing value completion (filling in missing data using interpolation or other methods), outlier removal (removing abnormal data that deviates from the normal range using methods such as the 3σ criterion), and normalization processing (converting parameters of different dimensions to the same numerical range to eliminate the influence of differences in dimensions) to obtain the corresponding condition feature set.

[0060] Step 102: Based on the preset multi-task loss function, the preset transformer fault detection model is trained using the working condition feature set to obtain the corresponding target transformer fault detection model. The target transformer fault detection model includes an extraction network and a hierarchical prediction network.

[0061] In this embodiment of the invention, a preset transformer fault detection model is trained using an input set of operating condition features to obtain corresponding training detection data. Based on a preset multi-task loss function, the loss function value corresponding to the operating condition feature set is determined according to the training detection data. When the loss function value is greater than or equal to a preset loss threshold, the network parameters of the transformer fault detection model are adjusted until the loss function value is less than the loss threshold. When the loss function value is less than the loss threshold, a corresponding target transformer fault detection model is generated. The target transformer fault detection model includes an extraction network and a hierarchical prediction network.

[0062] It should be noted that the multi-task loss function is as follows:

[0063]

[0064] in, The value of the loss function. Loss values ​​are categorized by fault type. The loss value is used to classify the severity of the fault. The total number of samples, The weight coefficient for class c is... Let i be the true label of the i-th sample belonging to the c-th class. Let be the probability that the i-th sample belongs to the c-th class. Let be the ordered penalty coefficient for the s-th level. Let be the weighting coefficient for the s-th level. Let the i-th sample be the true label of level s. Let c be the probability that the i-th sample belongs to level s, where i is the sample index, c is the category index, and s is the level index.

[0065] Step 103: Obtain the operating parameters of the single-phase transformer under test. Extract the multi-modal features of the operating parameters through the extraction network to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector.

[0066] The single-phase transformer under test refers to a single-phase transformer that requires fault monitoring.

[0067] Operating parameters refer to various status data generated during the real-time operation of the single-phase transformer under test, including three categories: electrical parameters, partial discharge parameters, and physical parameters.

[0068] In this embodiment of the invention, the operating parameters of the single-phase transformer under test are obtained. The electrical quantity parameters, partial discharge parameters and physical quantity parameters are extracted by extracting the electrical quantity feature extraction branch, the partial discharge waveform feature extraction branch and the physical state feature extraction branch in the network, respectively, to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector.

[0069] Step 104: Fault detection is performed on the electrical quantity feature vector, partial discharge waveform feature vector, and physical state feature vector through a hierarchical prediction network to obtain the corresponding monitoring results.

[0070] In this embodiment of the invention, a hierarchical prediction network is used to perform fault detection on electrical quantity feature vectors, partial discharge waveform feature vectors, and physical state feature vectors to obtain corresponding monitoring results.

[0071] In this embodiment of the invention, a preset transformer fault detection model is trained using a set of operating condition features based on a preset multi-task loss function to obtain a corresponding target transformer fault detection model. The target transformer fault detection model includes an extraction network and a hierarchical prediction network. It acquires the operating condition parameters of the single-phase transformer under test, extracts multimodal features from the operating condition parameters through the extraction network to obtain corresponding electrical quantity feature vectors, partial discharge waveform feature vectors, and physical state feature vectors, and performs fault detection on these vectors through the hierarchical prediction network to obtain the corresponding monitoring results. This overcomes the technical problem that traditional single-phase transformer monitoring methods often rely on threshold judgments of single operating condition parameters or manual inspections, making it difficult to meet the requirements for real-time and accurate monitoring during operation, thus reducing the reliability of single-phase transformer operation. Compared with traditional single-phase transformer monitoring methods, this invention significantly improves the accuracy of single-phase transformer fault detection by extracting multimodal features from operating condition parameters and detecting faults through the target transformer fault detection model. Furthermore, it can adapt to different operating conditions, improving the reliability of single-phase transformer operation.

[0072] Please see Figure 2 , Figure 2This is a flowchart illustrating the steps of a single-phase transformer monitoring method provided in Embodiment 2 of the present invention.

[0073] This invention provides a single-phase transformer monitoring method, comprising:

[0074] Step 201: Obtain multiple training condition parameters, preprocess each training condition parameter, and obtain the corresponding condition feature set.

[0075] In this embodiment of the invention, multiple training condition parameters are obtained, and missing values ​​are filled, outliers are removed, and normalization is performed on each training condition parameter to obtain the corresponding condition feature set.

[0076] Step 202: Train the preset transformer fault detection model using the operating condition feature set to obtain the corresponding training detection data.

[0077] Training and testing data refers to the set of fault detection results output by the model after the operating condition feature set is input into the preset transformer fault detection model, which includes the predicted fault type and predicted fault level for each sample.

[0078] In this embodiment of the invention, the operating condition feature set is input into a preset transformer fault detection model for fault detection to obtain corresponding training detection data.

[0079] Step 203: Input the training detection data and the working condition feature set into the preset multi-task loss function to obtain the corresponding loss function value.

[0080] In this embodiment of the invention, a preset multi-task loss function is used to determine the loss function value corresponding to the working condition feature set based on the training detection data.

[0081] Step 204: When the loss function value is greater than or equal to the preset loss threshold, adjust the network parameters of the transformer fault detection model until the loss function value is less than the loss threshold.

[0082] The loss threshold refers to a pre-set critical value used to determine whether the training accuracy of a model has met the target.

[0083] Network parameters refer to the variables in the transformer fault detection model that can be optimized through training, including the weights (parameters representing the importance of features) and biases (parameters that adjust the feature output baseline) of each convolutional layer and fully connected layer.

[0084] In this embodiment of the invention, when the loss function value is greater than or equal to a preset loss threshold, the gradient descent method or the adaptive momentum optimization algorithm (Adam optimizer) is used to optimize the network parameters of the transformer fault detection model until the loss function value is less than the loss threshold.

[0085] Step 205: When the loss function value is less than the loss threshold, the corresponding target transformer fault detection model is generated. The target transformer fault detection model includes an extraction network and a hierarchical prediction network.

[0086] In this embodiment of the invention, when the loss function value is less than the loss threshold, it indicates that the deviation between the prediction result of the transformer fault detection model and the actual fault state corresponding to the operating condition feature set is within an acceptable range, the model has fully learned the mapping law between the operating condition features and the fault state, and the transformer fault detection model is determined as the target transformer fault detection model.

[0087] Step 206: Obtain the operating parameters of the single-phase transformer under test. Extract the multi-modal features of the operating parameters through the extraction network to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector.

[0088] Furthermore, the extraction network includes an electrical quantity feature extraction branch, a partial discharge waveform feature extraction branch, and a physical state feature extraction branch. The operating parameters include electrical quantity parameters, partial discharge parameters, and physical quantity parameters. Step 206 includes the following sub-steps:

[0089] S11. Through the electrical quantity feature extraction branch, electrical quantity parameters are extracted to obtain the corresponding electrical quantity feature vector.

[0090] Further, see Figure 3 As shown, the electrical quantity feature extraction branch includes a first causal convolutional block, a second causal convolutional block, a third causal convolutional block, a global average pooling layer, and a multi-head temporal attention layer. S11 includes the following sub-steps:

[0091] S111. Multi-scale feature extraction of electrical quantity parameters is performed by sequentially connecting the first causal convolution block, the second causal convolution block, and the third causal convolution block to obtain the corresponding multi-dimensional electrical quantity feature vector.

[0092] Electrical parameters refer to parameters that reflect the power transmission status of the transformer within a preset sampling period, including high-voltage side voltage / current, low-voltage side voltage / current, etc.

[0093] It should be noted that the first causal convolutional block consists of a causal convolutional layer (kernel size 4×6, dilation factor 1) connected in sequence, a normalization layer, a GELU activation function layer, and a residual connection (i.e., the output of the GELU activation function layer is fused with the input of the causal convolutional layer for feature fusion). The second causal convolutional block consists of a causal convolutional layer (kernel size 4×128, dilation factor 2) connected in sequence, a normalization layer, a GELU activation function layer, and a residual connection (i.e., the output of the GELU activation function layer is fused with the input of the causal convolutional layer for feature fusion). The third causal convolutional block consists of a causal convolutional layer (kernel size 2×128, dilation factor 4) connected in sequence, a normalization layer, a GELU activation function layer, and a residual connection (i.e., the output of the GELU activation function layer is fused with the input of the causal convolutional layer for feature fusion).

[0094] In this embodiment of the invention, a first causal convolutional block is used to extract basic temporal features from electrical quantity parameters (i.e., basic temporal features are captured from electrical quantity parameters through a causal convolutional layer, then normalized to eliminate data distribution differences, a nonlinear feature expression is introduced through a GELU activation function layer, and finally the output of the GELU activation function layer and the input of the causal convolutional layer are fused through a residual connection), resulting in a corresponding basic temporal feature vector. A second causal convolutional block is used to extract mesoscale temporal correlation features from the basic temporal feature vector, resulting in a corresponding mesoscale temporal correlation feature vector. A third causal convolutional block is used to extract long-term, large-scale deep electrical quantity features from the mesoscale temporal correlation feature vector, resulting in a corresponding multidimensional electrical quantity feature vector.

[0095] It is worth mentioning that by sequentially processing and scaling the electrical quantity parameters through the first, second, and third causal convolutional blocks, the key information of the electrical quantity parameters under different time scales and feature dimensions can be fully extracted to obtain the corresponding multidimensional electrical quantity feature vectors.

[0096] It should be noted that a causal convolutional layer is a type of convolutional layer that follows temporal causal relationships. It only uses data from the current and historical moments for convolution calculations, avoiding interference from future data and ensuring the temporal rationality of feature extraction. Its key parameters include the kernel size (which determines the range of data covered by a single convolution) and the dilation coefficient (which determines the stride interval of the convolution operation and affects the size of the receptive field).

[0097] S112. The multidimensional electrical quantity feature vector is reduced in dimensionality by a global average pooling layer to obtain the corresponding initial electrical quantity feature vector.

[0098] In this embodiment of the invention, the multidimensional electrical quantity feature vector is reduced in dimensionality by a global average pooling layer (i.e., by calculating the average value of all elements in each channel, the high-dimensional multidimensional electrical quantity feature vector is mapped into a vector form with lower dimensionality and more concise information) to obtain the corresponding initial electrical quantity feature vector.

[0099] S113. The key features of the initial electrical quantity feature vector are extracted through the multi-head timing attention layer to obtain the corresponding electrical quantity feature vector.

[0100] In this embodiment of the invention, a multi-head temporal attention layer is used to extract key features from the initial electrical quantity feature vector (that is, to map the initial electrical quantity feature vector into a query vector, a key vector, and a value vector, and to calculate the temporal correlation weights between features in different dimensions through multiple parallel attention heads, highlighting the weight ratio of temporal features strongly related to the transformer fault state (such as voltage and current mutations, abnormal fluctuations, etc.), weakening the influence of irrelevant features, and then splicing and fusing the output results of each attention head to obtain the corresponding electrical quantity feature vector), thus obtaining the corresponding electrical quantity feature vector.

[0101] It should be noted that the multi-head temporal attention layer is a network layer used to enhance the key features of temporal data. By using multiple parallel attention heads to mine the temporal correlation between features, key features are given higher weights, enabling the focus and extraction of key temporal information and improving the relevance of feature representation.

[0102] S12. The partial discharge parameters are extracted by the partial discharge waveform feature extraction branch to obtain the corresponding partial discharge waveform feature vector.

[0103] Further, see Figure 3 As shown, the partial discharge waveform feature extraction branch includes a first convolutional layer, a second convolutional layer, a global max pooling layer, and a Transformer encoder. S12 includes the following sub-steps:

[0104] S121. Local feature extraction is performed on the partial discharge parameters through the first convolutional layer to obtain the corresponding first partial discharge feature vector. The first convolutional layer includes a 3×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence.

[0105] Partial discharge parameters refer to the partial discharge-related data acquired within a preset sampling period, including UHF partial discharge waveforms and discharge amplitude.

[0106] In this embodiment of the invention, local feature extraction of partial discharge parameters is performed through a first convolutional layer (i.e., a 3×1 one-dimensional convolutional layer is used to accurately capture the short-range correlation features and detailed fluctuation information of the partial discharge waveform in the partial discharge parameters using a narrow window convolution method; then, a normalization layer is used to perform distribution calibration on the feature data output by the 3×1 one-dimensional convolutional layer to eliminate the numerical differences between features of different dimensions and avoid gradient imbalance during model training; then, a LeakyReLU activation function layer is used to introduce nonlinear expressive power into the output of the normalization layer, while retaining some feature information in the negative interval to avoid gradient vanishing, thus obtaining the corresponding first partial discharge feature vector), thus obtaining the corresponding first partial discharge feature vector.

[0107] S122. Global feature extraction is performed on the first partial discharge feature vector through the second convolutional layer to obtain the corresponding second partial discharge feature vector. The second convolutional layer includes a 5×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence.

[0108] In this embodiment of the invention, the first partial discharge feature vector is extracted globally through the second convolutional layer (i.e., the first partial discharge feature vector is extracted through a 5×1 one-dimensional convolutional layer for global correlation feature extraction, and then the output of the 5×1 one-dimensional convolutional layer is normalized through a normalization layer to balance the numerical distribution of different global feature dimensions and avoid the model training effect due to the difference in feature amplitude. Then, the nonlinear expression capability of the feature is further enhanced through the LeakyReLU activation function layer, while retaining more effective feature information to avoid gradient vanishing), to obtain the corresponding second partial discharge feature vector.

[0109] It should be noted that global correlation features can characterize the overall trend of partial discharge parameters and the correlation information across time scales, complementing local features and comprehensively depicting the partial discharge state.

[0110] S123. The second partial discharge feature vector is reduced in dimensionality by a global max pooling layer to obtain the corresponding third partial discharge feature vector.

[0111] In this embodiment of the invention, the second partial discharge feature vector is reduced in dimensionality by a global max pooling layer (that is, the high-dimensional second partial discharge feature vector is mapped into a vector form with lower dimensionality and more prominent key information) to obtain the corresponding third partial discharge feature vector.

[0112] S124. The key features of the third partial discharge feature vector are extracted by the Transformer encoder to obtain the corresponding partial discharge waveform feature vector.

[0113] In this embodiment of the invention, the Transformer encoder is used to extract key features from the third partial discharge feature vector (i.e., to mine effective features strongly correlated with transformer insulation fault detection from the third partial discharge feature vector, including long-range time-series correlation, feature dependency, etc., and to strengthen effective features and weaken interference information) to obtain the corresponding partial discharge waveform feature vector.

[0114] It should be noted that the Transformer encoder is a feature extraction module based on a self-attention mechanism. Its core consists of a multi-head self-attention layer and a feedforward neural network. It is good at capturing long-range dependencies and global correlation features in sequence data, thereby improving the depth and accuracy of feature representation.

[0115] S13. The physical quantity parameters are adaptively weighted and adjusted through the physical state feature extraction branch to obtain the corresponding physical state feature vector.

[0116] Further, see Figure 3 As shown, the physical state feature extraction branch includes an adaptive weighting module and a perception module. S13 includes the following sub-steps:

[0117] S131. The physical quantity parameters are processed by the adaptive weighting module to obtain the corresponding weighted physical feature vector.

[0118] Physical parameters refer to parameters that reflect the operating status of the transformer's hardware, including top oil temperature, oil level, ambient temperature / humidity, etc.

[0119] It should be noted that the adaptive weighting module includes a condition-aware model and a 1×1 convolutional layer.

[0120] In this embodiment of the invention, normalized physical quantity parameters are input into the operating condition sensing model to obtain corresponding operating condition sensing factors. Based on the operating condition sensing factors, a weight vector adapted to the 9-dimensional physical quantity features is generated through a 1×1 convolutional layer. Based on the weight vector, the top oil temperature, oil level, ambient temperature / humidity, and other features in the physical quantity parameters are weighted respectively to obtain the corresponding weighted physical feature vector.

[0121] It should be noted that the working condition perception model is specifically as follows:

[0122]

[0123] in, For operating condition sensing factors, These are the normalized physical quantity parameters. The ambient temperature.

[0124] S132. The weighted physical feature vector is nonlinearly extracted through the perception module to obtain the corresponding physical state feature vector. The perception module includes a first hidden layer, a second hidden layer and a normalization layer connected in sequence.

[0125] It should be noted that the first hidden layer consists of a fully connected layer, a normalization layer, a ReLU activation layer, and a dropout layer connected in sequence. The second hidden layer consists of a fully connected layer, a normalization layer, a ReLU activation layer, and a dropout layer connected in sequence.

[0126] In this embodiment of the invention, the corresponding physical state feature vector is obtained by performing nonlinear feature extraction on the weighted physical feature vector through the perception module (i.e., the process of mining nonlinear correlation information in the weighted physical feature vector through activation functions, multi-layer network structures, etc.).

[0127] Step 207: Fault detection is performed on the electrical quantity feature vector, partial discharge waveform feature vector, and physical state feature vector through a hierarchical prediction network to obtain the corresponding monitoring results.

[0128] Further, see Figure 3 As shown, the hierarchical prediction network includes a multimodal feature fusion layer, a first detection head, and a second detection head. Step 207 includes the following sub-steps:

[0129] S21. The electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector are weighted and fused through the multimodal feature fusion layer to obtain the corresponding fused feature vector.

[0130] In this embodiment of the invention, a multimodal feature fusion layer performs a weighted fusion operation on the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector (i.e., through an adaptive weight calculation mechanism, dynamically adjusted weight coefficients are assigned to the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector based on their correlation with the transformer fault detection task; and element-level weighted summation and dimension alignment are performed on the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector based on each weight coefficient to obtain the corresponding fused feature vector), thus obtaining the corresponding fused feature vector.

[0131] It should be noted that the adaptive weight calculation mechanism refers to the core logic of dynamically allocating weights in the multimodal feature fusion layer. It can adjust the weight coefficients in real time according to the correlation between features and fault detection tasks, ensuring that key features receive greater attention.

[0132] S22. The first detection head performs fault detection on the fused feature vector to obtain the corresponding fault type.

[0133] It should be noted that the first detection head includes a fully connected layer and a Softmax activation function layer connected in sequence.

[0134] In this embodiment of the invention, the first detection head performs fault type classification on the fused feature vector (that is, the fused feature vector is mapped to a preset fault type label space through a fully connected layer to achieve the adaptation of feature dimension and fault type category, and then the mapped fused feature vector is subjected to probability normalization processing through a Softmax activation function layer) to obtain the corresponding fault type.

[0135] S23. When the fault type is normal, the monitoring result is determined to be normal operation.

[0136] In this embodiment of the invention, when the fault type is normal, the fault type is determined as the corresponding monitoring result.

[0137] S24. When the fault type is not normal, the second detection head is used to evaluate the fused feature vector to obtain the corresponding fault level.

[0138] It should be noted that the second detection head includes a fully connected layer and a Softmax activation function layer connected in sequence.

[0139] In this embodiment of the invention, when the fault type is winding deformation, multi-point grounding of the iron core, partial discharge in oil, poor contact of tap changer, or abnormal rise in oil temperature, the second detection head performs fault assessment on the fused feature vector (that is, the fused feature vector is mapped to the preset fault level label space through the fully connected layer to achieve the adaptation of feature dimension and fault severity classification, and then the Softmax activation function layer performs probability normalization processing on the mapped fused feature vector to obtain the corresponding fault level).

[0140] S25. Determine the fault level and fault type as the corresponding monitoring results.

[0141] In this embodiment of the invention, fault level and fault type are used as the corresponding monitoring results.

[0142] In this embodiment of the invention, a preset transformer fault detection model is trained using a set of operating condition features based on a preset multi-task loss function to obtain a corresponding target transformer fault detection model. The target transformer fault detection model includes an extraction network and a hierarchical prediction network. It acquires the operating condition parameters of the single-phase transformer under test, extracts multimodal features from the operating condition parameters through the extraction network to obtain corresponding electrical quantity feature vectors, partial discharge waveform feature vectors, and physical state feature vectors, and performs fault detection on these vectors through the hierarchical prediction network to obtain the corresponding monitoring results. This overcomes the technical problem that traditional single-phase transformer monitoring methods often rely on threshold judgments of single operating condition parameters or manual inspections, making it difficult to meet the requirements for real-time and accurate monitoring during operation, thus reducing the reliability of single-phase transformer operation. Compared with traditional single-phase transformer monitoring methods, this invention significantly improves the accuracy of single-phase transformer fault detection by extracting multimodal features from operating condition parameters and detecting faults through the target transformer fault detection model. Furthermore, it can adapt to different operating conditions, improving the reliability of single-phase transformer operation.

[0143] Please see Figure 4 , Figure 4 This is a structural block diagram of a single-phase transformer monitoring system provided in Embodiment 3 of the present invention.

[0144] This invention provides a single-phase transformer monitoring system, comprising:

[0145] The acquisition module 301 is used to acquire multiple training working condition parameters, preprocess each training working condition parameter, and obtain the corresponding working condition feature set.

[0146] Training module 302 is used to train a preset transformer fault detection model based on a preset multi-task loss function and a set of working condition features to obtain a corresponding target transformer fault detection model. The target transformer fault detection model includes an extraction network and a hierarchical prediction network.

[0147] The extraction module 303 is used to obtain the operating parameters of the single-phase transformer under test. The operating parameters are extracted in multiple modes through the extraction network to obtain the corresponding electrical quantity feature vector, partial discharge waveform feature vector and physical state feature vector.

[0148] The detection module 304 is used to perform fault detection on electrical quantity feature vectors, partial discharge waveform feature vectors and physical state feature vectors through a hierarchical prediction network, and obtain the corresponding monitoring results.

[0149] Furthermore, training module 302 includes:

[0150] The training submodule is used to train a preset transformer fault detection model by inputting a set of operating condition features, and to obtain the corresponding training detection data.

[0151] The adjustment submodule is used to input the training detection data and the working condition feature set into a preset multi-task loss function to obtain the corresponding loss function value;

[0152] When the loss function value is greater than or equal to the preset loss threshold, the network parameters of the transformer fault detection model are adjusted until the loss function value is less than the loss threshold.

[0153] When the loss function value is less than the loss threshold, the corresponding target transformer fault detection model is generated.

[0154] Furthermore, the extraction network includes an electrical quantity feature extraction branch, a partial discharge waveform feature extraction branch, and a physical state feature extraction branch. The operating parameters include electrical quantity parameters, partial discharge parameters, and physical quantity parameters. The extraction module 303 includes:

[0155] The first extraction submodule is used to extract electrical quantity features from electrical quantity parameters through the electrical quantity feature extraction branch to obtain the corresponding electrical quantity feature vector;

[0156] The second extraction submodule is used to extract partial discharge features from the partial discharge parameters through the partial discharge waveform feature extraction branch to obtain the corresponding partial discharge waveform feature vector.

[0157] The third extraction submodule is used to adaptively adjust the weights of physical quantity parameters through the physical state feature extraction branch to obtain the corresponding physical state feature vector.

[0158] Furthermore, the electrical quantity feature extraction branch includes a first causal convolutional block, a second causal convolutional block, a third causal convolutional block, a global average pooling layer, and a multi-head temporal attention layer. The first extraction submodule includes:

[0159] The multi-scale feature extraction unit is used to extract multi-scale features from electrical quantity parameters through a first causal convolutional block, a second causal convolutional block, and a third causal convolutional block connected in sequence, so as to obtain the corresponding multi-dimensional electrical quantity feature vector.

[0160] The first dimensionality reduction unit is used to reduce the dimensionality of the multidimensional electrical quantity feature vector through the global average pooling layer to obtain the corresponding initial electrical quantity feature vector.

[0161] The key feature extraction unit is used to extract key features from the initial electrical quantity feature vector through a multi-head temporal attention layer to obtain the corresponding electrical quantity feature vector.

[0162] Furthermore, the partial discharge waveform feature extraction branch includes a first convolutional layer, a second convolutional layer, a global max pooling layer, and a Transformer encoder; the second extraction submodule includes:

[0163] The first extraction unit is used to extract local features from the partial discharge parameters through the first convolutional layer to obtain the corresponding first partial discharge feature vector. The first convolutional layer includes a 3×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence.

[0164] The second extraction unit is used to perform global feature extraction on the first partial discharge feature vector through the second convolutional layer to obtain the corresponding second partial discharge feature vector. The second convolutional layer includes a 5×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence.

[0165] The third extraction unit is used to reduce the dimensionality of the second partial discharge feature vector through a global max pooling layer to obtain the corresponding third partial discharge feature vector.

[0166] The fourth extraction unit is used to extract key features from the third partial discharge feature vector using a Transformer encoder to obtain the corresponding partial discharge waveform feature vector.

[0167] Furthermore, the physical state feature extraction branch includes an adaptive weighting module and a perception module, and the third extraction sub-module includes:

[0168] The feature weighting unit is used to perform feature weighting on the physical quantity parameters through the adaptive weighting module to obtain the corresponding weighted physical feature vector;

[0169] The nonlinear feature extraction unit is used to perform nonlinear feature extraction on the weighted physical feature vector through the perception module to obtain the corresponding physical state feature vector. The perception module includes a first hidden layer, a second hidden layer and a normalization layer connected in sequence.

[0170] Furthermore, the hierarchical prediction network includes a multimodal feature fusion layer, a first detection head, and a second detection head. The detection module 304 includes:

[0171] The fusion submodule is used to perform a weighted fusion operation on the electrical quantity feature vector, the partial discharge waveform feature vector, and the physical state feature vector through the multimodal feature fusion layer to obtain the corresponding fused feature vector.

[0172] The first detection submodule is used to perform fault detection on the fused feature vector through the first detection head to obtain the corresponding fault type;

[0173] When the fault type is normal, the monitoring result is determined to be normal operation;

[0174] The second detection submodule is used to perform fault assessment on the fused feature vector through the second detection head when the fault type is not normal, and obtain the corresponding fault level.

[0175] The fault level and fault type are determined as the corresponding monitoring results.

[0176] Please see Figure 5 , Figure 5 This is a structural block diagram of an electronic device provided in Embodiment 4 of the present invention.

[0177] An electronic device according to an embodiment of the present invention includes: a memory 401 and a processor 402. The memory 401 stores a computer program. When the computer program is executed by the processor 402, the processor 402 performs a single-phase transformer monitoring method as described in any of the above embodiments.

[0178] Memory 401 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Memory 401 has storage space 403 for program code 413 for performing any of the method steps described above. For example, storage space 403 for program code may include individual program codes 413 for implementing the various steps in the methods described above. This program code may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When run by a computing processing device, this code causes the computing processing device to perform the various steps in the methods described above. This program code may be read from or written to one or more computer program products. These computer program products include program code carriers such as hard disks, CDs, memory cards, or floppy disks. The program code may be compressed, for example, in a suitable form. When this code is run by a computing device, it causes the computing device to perform the various steps in the single-phase transformer monitoring method described above.

[0179] Embodiment 5 of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the single-phase transformer monitoring method as described in any of the above embodiments.

[0180] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0181] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0182] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0183] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0184] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0185] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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. Such 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 single-phase transformer monitoring method, characterized by, The method comprises the following steps: obtaining a plurality of training working condition parameters, preprocessing each of the training working condition parameters to obtain a corresponding working condition feature set; training a preset transformer fault detection model based on a preset multi-task loss function and using the working condition feature set to obtain a corresponding target transformer fault detection model, wherein the target transformer fault detection model comprises an extraction network and a hierarchical prediction network; obtaining working condition parameters of a single-phase transformer to be measured, performing multi-modal feature extraction on the working condition parameters by using the extraction network to obtain an electrical quantity feature vector, a partial discharge waveform feature vector and a physical state feature vector; performing fault detection on the electrical quantity feature vector, the partial discharge waveform feature vector and the physical state feature vector by using the hierarchical prediction network to obtain a corresponding monitoring result.

2. The single-phase transformer monitoring method of claim 1, wherein, The step of training the preset transformer fault detection model based on the preset multi-task loss function and using the working condition feature set to obtain the corresponding target transformer fault detection model comprises the following steps: inputting the working condition feature set into the preset transformer fault detection model for training to obtain corresponding training detection data; inputting the training detection data and the working condition feature set into the preset multi-task loss function to obtain a corresponding loss function value; when the loss function value is greater than or equal to a preset loss threshold, adjusting the network parameters of the transformer fault detection model until the loss function value is less than the loss threshold; when the loss function value is less than the loss threshold, generating a corresponding target transformer fault detection model.

3. The single-phase transformer monitoring method of claim 1, wherein, The extraction network comprises an electrical quantity feature extraction branch, a partial discharge waveform feature extraction branch and a physical state feature extraction branch, the working condition parameters comprise electrical quantity parameters, partial discharge parameters and physical quantity parameters, and the step of performing multi-modal feature extraction on the working condition parameters by using the extraction network to obtain the corresponding electrical quantity feature vector, the partial discharge waveform feature vector and the physical state feature vector comprises the following steps: performing electrical quantity feature extraction on the electrical quantity parameters by using the electrical quantity feature extraction branch to obtain a corresponding electrical quantity feature vector; performing partial discharge feature extraction on the partial discharge parameters by using the partial discharge waveform feature extraction branch to obtain a corresponding partial discharge waveform feature vector; performing adaptive weight adjustment on the physical quantity parameters by using the physical state feature extraction branch to obtain a corresponding physical state feature vector.

4. The single-phase transformer monitoring method of claim 3, wherein, The electrical quantity feature extraction branch comprises a first causal convolution block, a second causal convolution block, a third causal convolution block, a global average pooling layer and a multi-head time sequence attention layer, and the step of performing electrical quantity feature extraction on the electrical quantity parameters by using the electrical quantity feature extraction branch to obtain a corresponding electrical quantity feature vector comprises the following steps: performing multi-scale feature extraction on the electrical quantity parameters by using the first causal convolution block, the second causal convolution block and the third causal convolution block connected in sequence to obtain a corresponding multi-dimensional electrical quantity feature vector; performing dimension reduction processing on the multi-dimensional electrical quantity feature vector by using the global average pooling layer to obtain a corresponding initial electrical quantity feature vector; The initial electrical quantity feature vector is subjected to key feature extraction through a multi-head time sequence attention layer to obtain a corresponding electrical quantity feature vector.

5. The single-phase transformer monitoring method of claim 3, wherein, The partial discharge waveform feature extraction branch includes a first convolutional layer, a second convolutional layer, a global maximum pooling layer and a Transformer encoder, and the step of performing partial discharge feature extraction on the partial discharge parameter through the partial discharge waveform feature extraction branch to obtain a corresponding partial discharge waveform feature vector includes: The partial discharge parameter is subjected to local feature extraction through the first convolutional layer to obtain a corresponding first partial discharge feature vector, wherein the first convolutional layer includes a 3×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence; The first partial discharge feature vector is subjected to global feature extraction through the second convolutional layer to obtain a corresponding second partial discharge feature vector, wherein the second convolutional layer includes a 5×1 one-dimensional convolutional layer, a normalization layer and a LeakyReLU activation function layer connected in sequence; The second partial discharge feature vector is subjected to dimension reduction processing through the global maximum pooling layer to obtain a corresponding third partial discharge feature vector; The third partial discharge feature vector is subjected to key feature extraction through the Transformer encoder to obtain a corresponding partial discharge waveform feature vector.

6. The single-phase transformer monitoring method of claim 3, wherein, The physical state feature extraction branch includes an adaptive weighting module and a perception module, and the step of performing adaptive weight adjustment on the physical quantity parameter through the physical state feature extraction branch to obtain a corresponding physical state feature vector includes: The physical quantity parameter is subjected to feature weighting processing through the adaptive weighting module to obtain a corresponding weighted physical feature vector; The weighted physical feature vector is subjected to nonlinear feature extraction through the perception module to obtain a corresponding physical state feature vector, wherein the perception module includes a first hidden layer, a second hidden layer and a normalization layer connected in sequence.

7. The single-phase transformer monitoring method of claim 1, wherein, The hierarchical prediction network includes a multi-modal feature fusion layer, a first detection head and a second detection head, and the step of performing fault detection on the electrical quantity feature vector, the partial discharge waveform feature vector and the physical state feature vector through the hierarchical prediction network to obtain a corresponding monitoring result includes: The electrical quantity feature vector, the partial discharge waveform feature vector and the physical state feature vector are subjected to weighted fusion operation through the multi-modal feature fusion layer to obtain a corresponding fusion feature vector; The fusion feature vector is subjected to fault detection through the first detection head to obtain a corresponding fault type; When the fault type is normal, the monitoring result is determined as normal operation; When the fault type is not normal, the fusion feature vector is subjected to fault evaluation through the second detection head to obtain a corresponding fault level; The fault level and the fault type are determined as the corresponding monitoring result.

8. A single-phase transformer monitoring system characterized by, The method includes: The acquisition module is configured to obtain a plurality of training working condition parameters, pre-process each of the training working condition parameters, and obtain a corresponding working condition feature set; The training module is configured to train a preset transformer fault detection model based on a preset multi-task loss function and the working condition feature set, to obtain a corresponding target transformer fault detection model, wherein the target transformer fault detection model comprises an extraction network and a hierarchical prediction network. The extraction module is configured to acquire working condition parameters of a single-phase transformer to be measured, perform multi-modal feature extraction on the working condition parameters through the extraction network, and obtain corresponding electrical quantity feature vectors, partial discharge waveform feature vectors, and physical state feature vectors. The detection module is configured to perform fault detection on the electrical quantity feature vectors, the partial discharge waveform feature vectors, and the physical state feature vectors through the hierarchical prediction network, and obtain corresponding monitoring results.

9. An electronic device, comprising: The computer program is executed to implement the single-phase transformer monitoring method according to any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed to implement the single-phase transformer monitoring method according to any one of claims 1-7.

Images