A method for diagnosing a state of a mutual inductor

Abstract

The application is particularly a state diagnosis method of a mutual inductor, and relates to the technical field of power equipment fault diagnosis, comprising: data cleaning and filtering and denoising processing on the multi-modal original data set, feature extraction on each modal data after preprocessing, and generation of a multi-modal feature vector set.In the application, multi-physical field data such as electrical quantity and magnetic field quantity are synchronously collected, covering multiple physical dimensions of the operating state of the mutual inductor, and reducing the information blind area of single physical quantity monitoring; by constructing an electromagnetic induction chain type propagation graph and embedding the electromagnetic induction law and the magnetic flux conservation equation in the form of a physical constraint function and a loss function regularization term into the feature propagation process of the graph convolutional neural network, the fused features conform to the physical law of electromagnetic propagation, the risk of false correlation caused by feature splicing is reduced, and the accuracy of the fault diagnosis result is improved.

Description

Technical Field

[0001] This invention relates to the field of power equipment fault diagnosis technology, and in particular to a condition diagnosis method for current transformers. Background Technology

[0002] Instrument transformers are crucial devices in power systems, responsible for voltage and current transformation. Their operating status directly impacts the metering accuracy and safe operation of the power system. With the expansion of power grids and the increasing service life of equipment, the need for fault diagnosis of instrument transformers is becoming increasingly urgent.

[0003] Currently, methods for fault diagnosis of instrument transformers are mainly divided into two categories: one is the threshold judgment method based on a single physical quantity, which judges the equipment status by collecting electrical parameters and comparing them with a preset threshold; the other is the multi-modal fusion diagnosis method that introduces multi-source data, which inputs feature data from different sources into a machine learning model for classification and judgment.

[0004] The aforementioned prior art has the following drawbacks:

[0005] Methods based on a single physical quantity have information blind spots, and relying solely on electrical parameters makes it difficult to identify early non-electrical anomalies such as localized overheating of the winding or mechanical deformation of the iron core.

[0006] Existing multimodal fusion methods only involve feature stitching at the data level and do not incorporate the physical laws of electromagnetic propagation into the fusion framework. This makes the model prone to learning false correlations and may result in misjudgment or missed judgment when operating conditions change.

[0007] Furthermore, during the long-term operation of the instrument transformer, the static diagnostic model will gradually decrease in diagnostic accuracy due to changes in data distribution caused by factors such as load fluctuations, environmental changes, and equipment aging, making it difficult to maintain stable diagnostic performance over the long term. Summary of the Invention

[0008] The purpose of this invention is to provide a method for diagnosing the condition of a current transformer in order to solve the above-mentioned problems.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] A method for diagnosing the condition of a current transformer includes:

[0011] Acquire electrical and magnetic field data during the operation of the current transformer, perform time alignment processing on the raw sampling data of each sensor channel, and generate a time-synchronized multimodal raw dataset.

[0012] The original multimodal dataset is cleaned and filtered for noise reduction. Features are extracted from each of the preprocessed modal data to generate a set of multimodal feature vectors.

[0013] An electromagnetic induction chain propagation graph is constructed based on the physical structure of the mutual inductor. The electromagnetic induction chain propagation graph is a directed graph, where nodes represent physical components of the mutual inductor, and directed edges represent propagation paths of electromagnetic quantities between components, with directions following the causal relationship of electromagnetic induction propagation. Each directed edge is appended with a mutual inductance coefficient and a magnetoresistance parameter. Each modal feature vector in the multimodal feature vector set is assigned to the corresponding node of the directed graph according to the physical correspondence as initial features. A graph convolutional neural network with embedded physical constraints is used to propagate and aggregate features along the directed graph, and global pooling is performed on the final feature vector of each node to generate a fused feature vector.

[0014] The fused feature vector is input into a multi-layer classification network, which outputs the fault diagnosis results of the mutual inductor.

[0015] In this context, the graph convolutional neural network with embedded physical constraints performs a linear transformation by concatenating the feature vectors of neighboring nodes with the physical parameter encoding vectors on the propagation path during each layer of graph convolution. The physical parameter encoding vectors are composed of normalized mutual inductance coefficients and magnetoresistance parameters.

[0016] Preferably, the total loss function of the graph convolutional neural network embedded with physical constraints includes a weighted sum of classification cross-entropy loss based on training sample labels and the norm of the physical equation residuals. The physical equation residuals include the magnetic flux conservation equation residuals and the electromagnetic induction law residuals. The magnetic flux conservation equation residual is the absolute value of the sum of the predicted magnetic flux values ​​of each branch at the core node. The electromagnetic induction law residual is the absolute value of the difference between the network predicted value of the secondary side induced electromotive force and the theoretical value calculated from the rate of change of magnetic flux according to Faraday's law.

[0017] Preferably, the predicted magnetic flux of each branch is obtained by projecting the hidden state feature vector of the core node onto a scalar through a linear mapping layer, the network predicted value of the secondary side induced electromotive force is obtained by projecting the hidden state feature vector of the secondary side winding node onto a scalar through a linear mapping layer, and the rate of change of magnetic flux is calculated by dividing the sequence of predicted magnetic flux values ​​of the core node by the sampling time interval after differentiating it in the time dimension.

[0018] Preferably, the feature extraction of each modal data after preprocessing includes: calculating time-domain statistical features and frequency-domain spectral features of the preprocessed electrical quantity data and concatenating them to generate electrical quantity feature vectors; arranging the three-dimensional magnetic field intensity vectors of each measuring point according to their spatial positions to generate three-dimensional magnetic field distribution data for the preprocessed magnetic field quantity data; performing singular value decomposition on the three-dimensional magnetic field distribution data and taking the first few singular values ​​and their corresponding left singular vectors as magnetic field feature vectors; mapping each modal feature vector to the same feature dimension through an independent fully connected linear transformation layer to generate the multimodal feature vector set.

[0019] Preferably, the number of singular values ​​retained in the singular value decomposition is determined by the cumulative energy ratio: the squares of each singular value are arranged in descending order and then summed sequentially. When the cumulative sum of squares reaches the preset energy retention ratio for the first time, the corresponding number of singular values ​​is the number of retained values.

[0020] Preferably, the method further includes: performing cross-modal attention calculation on the feature vectors of each node after multi-layer graph convolution propagation using a multi-head self-attention mechanism, taking the feature vectors of each node as inputs to the query vector, key vector, and value vector, calculating the attention weights between the features of each node, and outputting the attention-weighted node feature vector; and performing multi-level fusion processing on the attention-weighted node feature vector using a gated neural network, wherein the gated neural network includes three parallel branches: in-phase feature fusion gate, inter-phase feature fusion gate, and inter-group feature fusion gate, and the outputs of the three branches are concatenated to generate a multi-level fused feature vector as the fused feature vector.

[0021] Preferably, the method further includes: calculating information entropy and permutation entropy based on the fused feature vector; normalizing the information entropy and permutation entropy respectively and then weighting and combining them; mapping the weighted combination result to the interval of zero to one hundred through linear transformation to generate a health score; and triggering a warning signal of the corresponding level when the health score is lower than a preset threshold. The preset threshold is determined by subtracting a preset multiple standard deviation from the mean of the historical health scores of the current transformer under normal operating conditions.

[0022] Preferably, when the fault diagnosis result is determined to be an abnormal state, the method further includes: using backpropagation to calculate the gradient value of the diagnostic output on the features of each node in the directed graph, sorting the gradient values ​​from largest to smallest, determining the node that contributes the most to the abnormal diagnosis result and the location of its corresponding physical component, and generating a fault tracing and location result.

[0023] Preferably, the method further includes a concept drift detection and incremental update step: for continuously arriving multimodal raw data, a parameter-free statistical detection method is used to calculate the statistical distance between the current data distribution and the historical modeling data distribution; when the statistical distance exceeds a preset drift threshold, concept drift is determined to have occurred and a drift trigger signal is generated; based on the drift trigger signal, representative samples are extracted from the newly arrived data, and the parameters of the graph convolutional neural network with embedded physical constraints and the multi-layer classification network are incrementally learned and updated, and the threshold and confidence interval for anomaly determination are dynamically adjusted.

[0024] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0025] 1. This invention covers multiple physical dimensions of the transformer's operating status by simultaneously collecting multi-physical field data such as electrical and magnetic field quantities, reducing the information blind spots of single physical quantity monitoring; by constructing an electromagnetic induction chain propagation graph and embedding the electromagnetic induction law and magnetic flux conservation equation into the feature propagation process of the graph convolutional neural network in the form of physical constraint functions and loss function regularization terms, the fused features conform to the physical laws of electromagnetic propagation, reducing the risk of false correlations generated by feature splicing and improving the accuracy of fault diagnosis results.

[0026] 2. This invention uses gradient tracing along the electromagnetic induction chain propagation graph to enable the diagnostic results to be traced back to the corresponding physical component location, thus possessing physical interpretability. By introducing online concept drift recognition and incremental learning update mechanisms, the diagnostic model can adaptively adjust parameters and thresholds when the data distribution changes, which helps maintain the long-term diagnostic stability of the model. Attached Figure Description

[0027] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0028] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0029] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0030] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0031] Example 1

[0032] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0033] In this embodiment, it includes:

[0034] Step 1: Synchronous acquisition of multiphysics sensor data

[0035] The system acquires electrical and magnetic field data during the operation of the instrument transformer to generate a multi-mode raw dataset. The electrical data includes time-series data of three-phase voltage and three-phase current synchronously acquired from the primary and secondary sides of the transformer. The magnetic field data includes the three orthogonal directional components of the magnetic field intensity vector at each measuring point, acquired by a distributed fluxgate sensor network. , , And its spatial distribution gradient data. The raw sampling data of each sensor channel is time-aligned, and the data of each channel are aligned to the same time granularity with timestamps under a unified time base, generating a time-synchronized multimodal raw dataset.

[0036] The aforementioned time alignment process refers to addressing the differences in sampling frequencies between different sensor channels by upsampling low-frequency data to a time resolution consistent with the highest sampling rate using cubic spline interpolation, and downsampling high-frequency data to a uniform time granularity using sliding window mean aggregation, thereby ensuring strict alignment of each modal data in the time domain.

[0037] In this embodiment, to obtain more comprehensive information on the operating status of the current transformer, the following steps are also included: Based on step 1, temperature field data and mechanical stress data of the current transformer are acquired simultaneously. Temperature field data is collected by a distributed fiber optic sensor after activating the fiber grating by injecting a laser pulse of a specific wavelength. The temperature value at each measuring point is calculated based on the Bragg equation to obtain the temperature field distribution inside the current transformer. Mechanical stress data is calculated by the same distributed fiber optic sensing system after obtaining the transmission time and attenuation characteristics of the reflected laser pulse through frequency domain analysis to obtain the stress distribution along the fiber length. The temperature field data and mechanical stress data are then incorporated into the multimodal raw dataset after the above time alignment processing.

[0038] In this embodiment of the application, for oil-filled current transformers, the following steps are also included: based on the above-mentioned multimodal raw dataset, partial discharge signals are collected by a high-frequency current transformer, and the composition and concentration data of dissolved gases in oil are collected by an online oil chromatography monitoring device, and then incorporated into the multimodal raw dataset after time alignment processing.

[0039] Step 2: Preprocessing and Feature Extraction of Multimodal Data

[0040] The original multimodal dataset is cleaned and filtered for noise reduction to generate preprocessed multimodal data. Feature extraction is then performed on the preprocessed multimodal data to generate a set of multimodal feature vectors.

[0041] Specifically, step 2 includes the following sub-steps:

[0042] Step 201: Outlier removal and missing value imputation are performed on the time series data of each channel in the original multimodal dataset. Outlier removal uses the 3σ criterion, and missing value imputation uses a linear interpolation algorithm to generate pre-cleaned channel data. Wavelet threshold denoising is then applied to the pre-cleaned channel data to remove environmental noise and electromagnetic interference, generating pre-processed multimodal data. Wavelet threshold denoising involves multi-level wavelet decomposition of each channel data, applying soft thresholding to high-frequency wavelet coefficients to suppress noise components, and then reconstructing the signal through inverse wavelet transform. For the fluxgate sensor channel data, power frequency notch filtering is performed before wavelet threshold denoising, using a band-stop filter with a center frequency of 50Hz and its integer multiples of harmonic frequencies to eliminate power frequency electromagnetic interference.

[0043] Step 202: Perform joint time-domain and frequency-domain feature extraction on the preprocessed electrical quantity data:

[0044] In the time domain, the statistical characteristics of each channel's data, such as mean, variance, peak value, skewness, kurtosis, and waveform factor, are calculated.

[0045] In the frequency domain, a Fast Fourier Transform is performed on the data from each channel to extract spectral features such as the fundamental amplitude, harmonic content, and total harmonic distortion rate. The time-domain and frequency-domain features are then concatenated to generate an electrical quantity feature vector.

[0046] The formula for calculating the total harmonic distortion rate is:

[0047] ;

[0048] in, The fundamental amplitude, For the first Second harmonic amplitude For harmonic order, This represents the highest harmonic order calculated. Since it is a dimensionless ratio, the numerator and denominator have the same dimensions, and can be directly calculated.

[0049] Step 203: Extract magnetic field distribution features from the preprocessed magnetic field data. Arrange the three-dimensional magnetic field intensity vectors of each measuring point according to their spatial positions to generate three-dimensional magnetic field distribution data. Perform singular value decomposition on the three-dimensional magnetic field distribution data, extract the first few singular values ​​and their corresponding left singular vectors as modal features of the magnetic field distribution, and generate magnetic field feature vectors. Singular value decomposition reconstructs the three-dimensional magnetic field distribution data into a matrix. Then, the decomposition was performed, and the decomposition result was: ,in It is a left singular matrix. It is a diagonal matrix. It is a right singular matrix. Indicates transpose, take Center front A singular value and The corresponding front in the middle Column vectors serve as modal characteristics of magnetic field distribution.

[0050] The number of singular values ​​retained in the above singular value decomposition The energy retention ratio is determined by calculating the cumulative energy percentage: The squares of each singular value are arranged in descending order and then summed sequentially. The number of singular values ​​is determined when the cumulative sum of squares first reaches the preset energy retention ratio. This allows for the compression of feature dimensions while preserving the main modal information of the magnetic field distribution.

[0051] In this embodiment of the application, in order to capture the joint time-frequency features of electrical signals, in addition to step 202, the following steps are also included: performing continuous wavelet transform on the preprocessed electrical quantity data to generate two-dimensional time-frequency images for each channel, which serve as inputs for subsequent spatial feature extraction. The continuous wavelet transform uses the Morlet wavelet as the mother wavelet, with the input being the time series of electrical quantities for each channel, and the output being a coefficient matrix in both time and frequency dimensions. This coefficient matrix is ​​then converted into a two-dimensional time-frequency image.

[0052] In this embodiment, when the multimodal raw dataset includes temperature field data and mechanical stress data, the following steps are further included: For the preprocessed temperature field data, extract features such as core hot spot temperature, winding hot spot temperature, maximum value of temperature gradient distribution, and temperature rise rate to generate a temperature field feature vector; for the preprocessed mechanical stress data, extract features such as stress peak location, stress distribution uniformity (characterized by the coefficient of variation of stress values), and stress change rate to generate a mechanical stress feature vector. Since the temperature field features (unit: °C or K), mechanical stress features (unit: Pa), electrical quantity features, and magnetic field quantity features differ in dimensions and numerical range from these features, before incorporating the above modal feature vectors into the multimodal feature vector set, all modal feature vectors are uniformly standardized using Z-score to eliminate the influence of dimensional differences on subsequent fusion calculations. The standardized temperature field feature vector, mechanical stress feature vector, electrical quantity feature vector, and magnetic field feature vector are then incorporated into the multimodal feature vector set.

[0053] In this embodiment of the application, in order to utilize the interactive information between the temperature field and mechanical stress to assist in the characterization of the magnetic field distribution characteristics, in addition to step 203, the following steps are also included: converting the temperature change data into a temperature transformation vector as a row vector, converting the mechanical stress distribution data into a mechanical distribution vector as a column vector, performing an outer product operation on the temperature transformation vector and the mechanical distribution vector to obtain a coupling matrix, performing singular value decomposition on the coupling matrix, extracting the first few singular values ​​and their corresponding singular vectors as thermo-mechanical coupling mode features, and concatenating the thermo-mechanical coupling mode features with the magnetic field feature vector to update the magnetic field feature vector.

[0054] Step 204: Vectorize the feature vectors of each modality to form a set of multimodal feature vectors with a unified dimension.

[0055] The aforementioned vectorization encoding refers to mapping each modality feature vector to the same feature dimension through an independent fully connected linear transformation layer. This ensures that feature vectors from different modalities with inconsistent original dimensions have a uniform vector length before entering the subsequent fusion step, thereby satisfying the requirement of graph convolutional neural networks for consistency of the initial feature dimensions of nodes.

[0056] Step 3: Physically Constrained Modeling and Feature Fusion of Electromagnetic Propagation Chain

[0057] An electromagnetic induction chain propagation graph is constructed based on the physical structure and electromagnetic propagation laws of mutual inductors. Each modal feature vector in the multimodal feature vector set is assigned to the corresponding node of the electromagnetic induction chain propagation graph. A graph convolutional neural network with embedded physical constraints is used to propagate and aggregate features along the electromagnetic induction chain propagation graph to generate a fused feature vector.

[0058] Specifically, step 3 includes the following sub-steps:

[0059] Step 301: Construct a directed graph based on the physical structure of the mutual inductors. , where the set of nodes Each node in the diagram represents a key physical component of the current transformer, including primary side binding nodes, core nodes, and secondary side binding nodes; the set of directed edges... Each directed edge in the diagram represents the propagation path of electromagnetic quantities between components, and the set of directed edges represents the path of electromagnetic quantities propagation between components. The direction of each directed edge follows the causal relationship of electromagnetic induction propagation, that is, from the primary side-binding node through the core node to the secondary side-binding node. Each directed edge is appended with the physical equation parameters that the propagation path follows, including the mutual inductance coefficient in Faraday's law of electromagnetic induction and the magnetic reluctance parameter in the magnetic flux conservation equation.

[0060] The above directed graph The set of nodes in It can also include auxiliary physical component nodes such as magnetic shielding nodes and insulating structure nodes, and a set of directed edges. The directed edges in the network correspondingly increase the electromagnetic propagation paths between these auxiliary nodes and the main node.

[0061] Step 302: Assign each modal feature vector in the multimodal feature vector set to the directed graph according to the physical correspondence. Each node is used as its initial feature. Among them, the primary-side related features in the electrical quantity feature vector are assigned to the primary-side binding nodes, the secondary-side related features in the electrical quantity feature vector are assigned to the secondary-side binding nodes, and the magnetic field feature vector is assigned to the core nodes.

[0062] When the multimodal feature vector set includes temperature field feature vectors and mechanical stress feature vectors, the temperature field feature vectors and mechanical stress feature vectors are assigned to the corresponding nodes according to the correspondence between the sensor deployment location and the physical component. For example, the core temperature feature is assigned to the core node, and the winding temperature feature is assigned to the corresponding winding node.

[0063] Step 303: Apply graph convolutional neural networks with embedded physical constraints to directed graphs. The features of each node are propagated and aggregated layer by layer. The graph convolutional neural network with embedded physical constraints contains multiple graph convolutional layers, and the propagation rule for each graph convolutional layer is as follows:

[0064] ;

[0065] in, For nodes In the The feature vector of the layer, For nodes neighboring nodes In the The feature vector of the layer, For nodes The set of neighboring nodes, For the first The learnable weight matrix of the layer, For the node To the node The output vector of the physical constraint function, This indicates a feature concatenation operation. It is a non-linear activation function.

[0066] The above physical constraint functions The output vector is based on the node and nodes The physical propagation relationship between them is determined, and the specific construction method is as follows: For directed edges From directed graph Read the pre-attached mutual inductance coefficient on the directed edge. and magnetoresistive parameters Mutual inductance coefficient and magnetoresistive parameters Concatenate into a fixed-length physical parameter encoding vector The physical parameter encoding vector is The output of . Among them, the mutual inductance coefficient According to Faraday's law of electromagnetic induction, characterizing from node Corresponding components to nodes Magnetic flux coupling strength of corresponding components; magnetoresistive parameters Based on the magnetic flux conservation equation, the magnetic circuit impedance along this propagation path is characterized. Due to the mutual inductance coefficient... (Unit: H) and magnetoresistive parameters (Unit: A / Wb) Due to the different dimensions, before constructing the physical parameter encoding vector, the mutual inductance coefficients were respectively... and magnetoresistive parameters The mutual inductance coefficient is normalized by mean normalization based on the range. and magnetoresistive parameters The data is uniformly mapped to a dimensionless numerical range before being concatenated. By concatenating the physical parameter encoding vector with the feature vectors of neighboring nodes and having them participate in a linear transformation, the feature propagation process is numerically constrained by the laws of electromagnetic induction and the magnetic flux conservation equation.

[0067] The above mutual inductance coefficient and magnetoresistive parameters The values ​​are derived from the parameters on the instrument transformer's nameplate or the design values ​​recorded in the equipment file, and are pre-written into the directed graph during the model building phase. The attribute fields of each corresponding directed edge remain unchanged as fixed physical prior parameters during subsequent training and inference.

[0068] The aforementioned graph convolutional neural network with embedded physical constraints takes the initial feature vector of each node as input and the hidden state feature vector of each node after propagation through multiple layers of graph convolution as output. It uses labeled samples for supervised training. During training, a fully connected output layer is connected after the graph convolutional layer to map the hidden state feature vector of each node to the classification prediction probability distribution of each fault category. The training loss is calculated and backpropagation is used to update the network parameters. After training, the fully connected output layer is removed, and the hidden state feature vector output by the graph convolutional layer is retained for use in subsequent steps.

[0069] The physical constraints of the aforementioned graph convolutional neural network with embedded physical constraints are injected into the training process as soft constraints. Specifically, the total loss function of the graph convolutional neural network with embedded physical constraints is defined as:

[0070] ;

[0071] in, For classification cross-entropy loss based on training sample labels, Let the norm of the residuals of the physical equations be denoted as . To balance the hyperparameters of the two losses, the residuals of the above physical equations include the residuals of the magnetic flux conservation equation and the electromagnetic induction law. The magnetic flux conservation equation residuals are used to calculate the sum of the magnetic flux in each branch at the core node. The deviation from zero, where For the first core node The magnetic flux of the branch circuit, For branch index, i.e., it is required The degree to which the law of electromagnetic induction holds true; the deviation between the network-predicted value of the induced electromotive force on the secondary side calculated from the residual of the law of electromagnetic induction and the theoretical value calculated from the rate of change of magnetic flux according to Faraday's law, that is, the requirement that the network-predicted induced electromotive force is consistent with... The difference between them approaches zero, where This refers to the number of turns in the secondary winding. The magnetic flux passing through the iron core, This represents the derivative of the magnetic flux with respect to time. The Adam optimization algorithm is used to update the network parameters during the training process.

[0072] The above The specific calculation method is as follows: For each sample in the current batch, calculate the residuals of the magnetic flux conservation equation and the electromagnetic induction law. The residual of the magnetic flux conservation equation is taken as the absolute value of the sum of the predicted magnetic flux values ​​of each branch at the core node. The residual of the electromagnetic induction law is taken as the sum of the predicted values ​​of the induced electromotive force network on the secondary side and... The absolute value of the difference between the theoretical values ​​is obtained by adding the two residuals and taking the mean of all samples within the batch. The predicted magnetic flux values ​​for each branch are obtained by projecting the hidden state feature vectors corresponding to the core nodes in the graph convolutional neural network with embedded physical constraints onto a scalar through a linear mapping layer. Similarly, the predicted values ​​for the secondary-side induced electromotive force network are obtained by projecting the hidden state feature vectors corresponding to the secondary-side winding nodes onto a scalar through a linear mapping layer. The predicted magnetic flux sequence of the iron core nodes is differentially divided by the sampling time interval in the time dimension. The calculation yielded, where The time interval between two adjacent sampling times. The value is equal to the reciprocal of the sensor's sampling frequency.

[0073] The aforementioned physical constraints can also be injected into the network structure in the form of hard constraints. Hard constraints involve embedding the physical equations into the network structure itself, ensuring that the network output strictly satisfies the physical equations. For example, the magnetic flux conservation condition... The normalization layer, which acts on the graph convolutional layer, projects the magnetic flux-related features output by the graph convolutional layer, forcing them to satisfy the constraint that the sum of magnetic fluxes is zero.

[0074] Step 304: Perform global pooling on the feature vectors of each node after multi-layer graph convolution propagation, and aggregate the final feature vectors of each node into a global fusion feature vector, which is then output as the fusion feature vector.

[0075] In this embodiment, to adaptively allocate importance weights for different modal features during the fusion process, the following step is included between steps 303 and 304: cross-modal attention calculation is performed on the feature vectors of each node after multi-layer graph convolution propagation using a multi-head self-attention mechanism. The multi-head self-attention mechanism takes the feature vectors of each node as inputs to the query vector, key vector, and value vector, calculates the attention weights between the features of each node, automatically focuses on the feature dimension most relevant to the current diagnostic task, and outputs the attention-weighted node feature vector.

[0076] In this embodiment, to learn multi-level fusion representations, after the aforementioned multi-head self-attention mechanism, the following steps are further included: using a gated neural network to perform multi-level fusion processing on the attention-weighted node feature vectors. The gated neural network includes three parallel branches: an in-phase feature fusion gate, an inter-phase feature fusion gate, and an inter-group feature fusion gate. These branches respectively handle feature interactions within the same physical component, feature interactions between adjacent physical components, and feature interactions between different physical quantity groups. The outputs of the three branches are concatenated to generate a multi-level fusion feature vector, which replaces the global pooling processing result in step 304 as the fusion feature vector output.

[0077] The aforementioned gated neural network takes attention-weighted node feature vectors as input and outputs a multi-level fused feature vector resulting from the concatenation of three parallel branches. Each parallel branch consists of a gating unit, which processes the input feature vector... First, a linear transformation is performed on the learnable weight matrix, and then the gated weight vector is calculated using the Sigmoid activation function. Then the gating weight vector With input feature vector Element-wise multiplication is performed to filter out effective feature interaction information at the corresponding levels; the outputs of the three branches are concatenated along the feature dimension to generate a multi-level fused feature vector. The gated neural network is trained under supervision using labeled samples. After the outputs of the three parallel branches are concatenated, a fully connected output layer is connected to map the multi-level fused feature vector to the classification prediction probability distribution of each fault category. The loss function is cross-entropy loss, and the Adam optimization algorithm is used to update the network parameters.

[0078] Step 4: Fault diagnosis decision based on fusion features

[0079] The fused feature vectors are input into a multi-layer classification network, which outputs the fault diagnosis results of the mutual inductor.

[0080] The aforementioned multi-layer classification network uses fused feature vectors as input to its input layer and a Softmax classification layer as its output layer. It outputs the probability distribution of each classification label and selects the label with the highest probability as the diagnostic result. The classification labels include four categories: healthy state, attentive state, warning state, and fault state, corresponding to four operating states: all parameters within the normal range, some parameters deviating from the baseline value, multiple parameters exhibiting abnormalities, and a definite fault occurring. The multi-layer classification network is trained with labeled samples under supervised supervision, using cross-entropy loss as the loss function and the Adam optimization algorithm to update the network parameters.

[0081] The aforementioned multi-layer classification network can also contain multiple dedicated classifier branches, each classifying and determining different fault types, including inter-turn short circuits, core saturation, insulation degradation, and poor neutral contact. The outputs of each dedicated classifier branch are combined and processed by a fusion decision process to determine the fault type. The fusion decision process uses a weighted vote of the probabilities of each branch's outputs to generate the final fault type determination result.

[0082] In this embodiment of the application, in order to quantitatively assess the health status of the current transformer, in addition to step 4, the following steps are also included: calculating the health score of the current transformer based on the fused feature vector; and calculating the information entropy of the fused feature vector. and permutation entropy Information entropy is defined as:

[0083] ;

[0084] in, For the first The output probability of each category label The total number of category labels, This represents the index of the category label. Permutation entropy. The complexity of the feature sequence is measured by statistically analyzing the permutation patterns of the components of the fused feature vector. Specifically, the components of the fused feature vector are arranged according to a set embedding dimension. Divided into lengths of For each subsequence, calculate the relative frequency of each permutation pattern. Then the permutation entropy is defined as:

[0085] ;

[0086] in, For the first The relative frequency of each permutation pattern in all subsequences. This is the sequence number of the arrangement pattern. For embedded dimensions The total number of all possible permutations. Due to information entropy. With permutation entropy The numerical ranges are different, so the information entropy is evaluated separately before weighted combination. and permutation entropy The mean normalization based on the range is used to reduce the information entropy. and permutation entropy After being uniformly mapped to the same numerical range, the values ​​are weighted and combined. The weighted combination result is then linearly transformed and mapped to the interval between 0 and 100 to generate a health score. When the health score falls below a preset threshold, a warning signal of the corresponding level is triggered.

[0087] The aforementioned preset threshold is determined based on the historical health score distribution of the current transformer under normal operating conditions. Specifically, the lower limit threshold for triggering an early warning is obtained by subtracting the preset multiple standard deviation from the mean of the health scores corresponding to the historical normal operating samples. When the current health score is lower than the lower limit threshold, it is judged as abnormal and an early warning signal of the corresponding level is triggered. Different early warning levels correspond to different standard deviation multiples, and the larger the deviation, the higher the early warning level.

[0088] In this embodiment of the application, in order to locate the specific physical location of the fault, in addition to step 4, the following step is also included: when the diagnostic result determines that the state is abnormal, along the directed graph Backward gradient tracing is performed along the directed edges to calculate the gradient contribution of each node's features to the diagnostic results. For the final feature vector of each node in the graph convolutional neural network with embedded physical constraints, backpropagation is used to calculate the gradient value of the diagnostic output with respect to each node's features. The gradient values ​​are sorted from largest to smallest to determine the node that contributes the most to the abnormal diagnosis results and its corresponding physical component location, generating fault tracing and localization results.

[0089] In this embodiment of the application, in order to maintain the accuracy of the diagnostic model during long-term operation, the following steps are also included:

[0090] Step 501: For continuously arriving multimodal raw data, use a parameter-free statistical detection method to monitor changes in data distribution in real time, calculate the statistical distance between the current data distribution and the historical modeled data distribution, and determine that concept drift has occurred when the statistical distance exceeds a preset drift threshold, and generate a drift trigger signal.

[0091] The above parameter-free statistical detection method uses the Kolmogorov-Smirnov test to perform a two-sample test on the edge distribution of each modality feature within the current sliding window and the distribution of the corresponding features in the historical modeling stage. After calculating the test statistics of each dimension, the maximum value is taken as the statistical distance output at the current time.

[0092] The aforementioned preset drift threshold is determined based on the statistical characteristics of the historical data distribution of the current transformer during the stable operation phase. Specifically, the upper limit threshold for drift determination is obtained by adding the mean of the statistical distance sequence calculated during the historical stable operation phase to a preset multiple of the standard deviation. When the current statistical distance exceeds the upper limit threshold, a concept drift is determined to have occurred and a drift trigger signal is generated.

[0093] Step 502: Based on the drift trigger signal, extract representative samples from the newly arrived data, and perform mini-batch incremental learning to update the parameters of the graph convolutional neural network and multi-layer classification network with embedded physical constraints, generating updated diagnostic model parameters. Simultaneously, dynamically adjust the anomaly detection threshold and confidence interval according to changes in data distribution, generating updated diagnostic threshold configurations.

[0094] The above-mentioned extraction of representative samples from newly arrived data refers to stratified sampling of newly arrived data after drift triggering according to each fault category label. Under the premise that the number of samples in each category meets the minimum batch requirement, the samples with the lowest cosine similarity to historical modeling data in the feature space are selected first to maximize the information increment of the new samples to the existing training set. The selected samples are then used to form a small batch dataset for incremental learning and updating.

[0095] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0096] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0097] In addition, the functional units in the various embodiments of this application 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.

[0098] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for diagnosing the condition of a current transformer, characterized in that, include: Acquire electrical and magnetic field data during the operation of the current transformer, perform time alignment processing on the raw sampling data of each sensor channel, and generate a time-synchronized multimodal raw dataset. The original multimodal dataset is cleaned and filtered for noise reduction. Features are extracted from each of the preprocessed modal data to generate a set of multimodal feature vectors. An electromagnetic induction chain propagation graph is constructed based on the physical structure of the mutual inductor. The electromagnetic induction chain propagation graph is a directed graph, where nodes represent physical components of the mutual inductor, and directed edges represent propagation paths of electromagnetic quantities between components, with directions following the causal relationship of electromagnetic induction propagation. Each directed edge is appended with a mutual inductance coefficient and a magnetoresistance parameter. Each modal feature vector in the multimodal feature vector set is assigned to the corresponding node of the directed graph according to the physical correspondence as initial features. A graph convolutional neural network with embedded physical constraints is used to propagate and aggregate features along the directed graph, and global pooling is performed on the final feature vector of each node to generate a fused feature vector. The fused feature vector is input into a multi-layer classification network, which outputs the fault diagnosis results of the mutual inductor. In this context, the graph convolutional neural network with embedded physical constraints performs a linear transformation by concatenating the feature vectors of neighboring nodes with the physical parameter encoding vectors on the propagation path during each layer of graph convolution. The physical parameter encoding vectors are composed of normalized mutual inductance coefficients and magnetoresistance parameters.

2. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, The total loss function of the graph convolutional neural network with embedded physical constraints includes a weighted sum of classification cross-entropy loss based on training sample labels and the norm of the physical equation residuals. The physical equation residuals include the magnetic flux conservation equation residuals and the electromagnetic induction law residuals. The magnetic flux conservation equation residual is the absolute value of the sum of the predicted magnetic flux values ​​of each branch at the core node. The electromagnetic induction law residual is the absolute value of the difference between the network predicted value of the secondary side induced electromotive force and the theoretical value calculated from the rate of change of magnetic flux according to Faraday's law.

3. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, The predicted values ​​of magnetic flux in each branch are obtained by projecting the hidden state feature vector of the core node onto a scalar through a linear mapping layer. The predicted values ​​of the induced electromotive force on the secondary side are obtained by projecting the hidden state feature vector of the secondary winding node onto a scalar through a linear mapping layer. The rate of change of magnetic flux is calculated by dividing the sequence of predicted magnetic flux values ​​of the core node by the sampling time interval after differentiating it in the time dimension.

4. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, Feature extraction is performed on each of the preprocessed modal data, including: calculating time-domain statistical features and frequency-domain spectral features of the preprocessed electrical quantity data and concatenating them to generate electrical quantity feature vectors; arranging the three-dimensional magnetic field intensity vectors of each measuring point according to their spatial positions to generate three-dimensional magnetic field distribution data for the preprocessed magnetic field quantity data; performing singular value decomposition on the three-dimensional magnetic field distribution data and taking the first few singular values ​​and their corresponding left singular vectors as magnetic field feature vectors; mapping each modal feature vector to the same feature dimension through an independent fully connected linear transformation layer to generate the multimodal feature vector set.

5. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, The number of singular values ​​retained in singular value decomposition is determined by the cumulative energy ratio: the squares of each singular value are arranged in descending order and then summed up. When the cumulative sum of squares reaches the preset energy retention ratio for the first time, the corresponding number of singular values ​​is the number of retained singular values.

6. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, Before the global pooling process, the method further includes: performing cross-modal attention calculation on the feature vectors of each node after multi-layer graph convolution propagation using a multi-head self-attention mechanism, taking the feature vectors of each node as inputs to the query vector, key vector, and value vector, calculating the attention weights between the features of each node, and outputting the attention-weighted node feature vector; and performing multi-level fusion processing on the attention-weighted node feature vector using a gated neural network, wherein the gated neural network includes three parallel branches: in-phase feature fusion gate, inter-phase feature fusion gate, and inter-group feature fusion gate, and the outputs of the three branches are concatenated to generate a multi-level fused feature vector as the fused feature vector.

7. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, Also includes: Information entropy and permutation entropy are calculated based on the fused feature vector. The information entropy and permutation entropy are normalized and then weighted and combined. The weighted combination result is linearly transformed and mapped to the interval of zero to one hundred to generate a health score. When the health score is lower than a preset threshold, a warning signal of the corresponding level is triggered. The preset threshold is determined by subtracting a preset multiple standard deviation from the mean of the historical health scores of the current transformer under normal operating conditions.

8. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, When the fault diagnosis result is determined to be an abnormal state, the method further includes: using backpropagation to calculate the gradient value of the diagnostic output on the features of each node in the directed graph, sorting the gradient values ​​from largest to smallest, determining the node that contributes the most to the abnormal diagnosis result and the location of its corresponding physical component, and generating a fault tracing and location result.

9. The method for diagnosing the condition of a current transformer according to claim 1, characterized in that, It also includes concept drift detection and incremental update steps: for continuously arriving multimodal raw data, a parameter-free statistical detection method is used to calculate the statistical distance between the current data distribution and the historical modeling data distribution. When the statistical distance exceeds a preset drift threshold, concept drift is determined to have occurred and a drift trigger signal is generated. Based on the drift trigger signal, representative samples are extracted from the newly arrived data, and the parameters of the graph convolutional neural network with embedded physical constraints and the multi-layer classification network are incrementally learned and updated, and the threshold and confidence interval for anomaly determination are dynamically adjusted.

Images