A mutual inductor test abnormal data automatic filtering analysis method and system

Abstract

The present application relates to the technical field of mutual inductor test, and provides a mutual inductor test abnormal data automatic filtering analysis method and system, which comprises the following steps: acquiring magnetic field data, electric field data and thermal field data, respectively performing feature extraction and weighted fusion to obtain a multi-physical field fusion feature vector; constructing an electromagnetic induction chain type propagation graph, calculating the weight of the edges in the graph, constructing a magnetic flux conservation constraint graph convolutional neural network, learning the propagation and evolution mode of the mutual inductor test abnormality, and obtaining a coupled abnormal feature vector; classifying the coupled abnormal feature vector, identifying normal data, single physical field abnormal data and multi-physical field coupled abnormal data, determining the corresponding data as abnormal data and filtering; and performing physical mechanism analysis on the filtered abnormal data to obtain an abnormality detection report. The present application realizes high-precision automatic identification, filtering analysis and physical mechanism tracing diagnosis of mutual inductor test abnormal data.

Description

Technical Field

[0001] This invention relates to the field of instrument transformer testing technology, and in particular to an automatic filtering and analysis method and system for abnormal data in instrument transformer testing. Background Technology

[0002] Automatic filtering and analysis of abnormal data from instrument transformer tests refers to a technical system that intelligently detects, classifies, identifies, and automatically filters anomalies in multi-physics field data generated by instrument transformers under various test conditions in power systems. Compared with traditional methods of monitoring single physical quantities or manual judgment, modern instrument transformer testing involves complex data from multiple physical domains such as magnetic fields, electric fields, and thermal fields. These data vary significantly in terms of test frequency, load conditions, ambient temperature, and equipment aging. Through collaborative analysis of multi-physics field data, the system can identify diverse anomaly modes of instrument transformers to the greatest extent possible. For example, it can simultaneously detect different types of fault symptoms such as hysteresis nonlinearity anomalies, dielectric loss anomalies, and temperature rise anomalies under complex operating conditions, exhibiting stronger detection accuracy, higher anomaly identification accuracy, and a wider range of applications. However, multi-physics field anomaly data analysis also brings additional challenges, the key being how to handle the coupling relationships between different physical fields, accurately identify progressive insulation degradation processes, and achieve intelligent anomaly filtering.

[0003] In existing technologies, the analysis of abnormal data in instrument transformer tests mainly adopts independent analysis of a single physical field. By establishing independent anomaly detection models for magnetic field, electric field, and thermal field data, basic anomaly identification and data filtering functions are achieved. However, this independent analysis method ignores the complex coupling and propagation relationships between multiple physical fields and cannot accurately capture cross-domain coupling anomalies caused by physical mechanisms such as electromagnetic induction, dielectric loss heating, and hysteresis loss. This results in serious misjudgments and omissions of coupling anomalies and low accuracy in filtering abnormal data. Summary of the Invention

[0004] In view of this, the present invention proposes an automatic filtering and analysis method and system for abnormal data of current transformer tests, which solves the problem that the existing technology ignores the complex coupling and propagation relationship between multiple physical fields, cannot accurately capture cross-domain coupling anomalies caused by physical mechanisms such as electromagnetic induction, dielectric loss heating, and hysteresis loss, resulting in serious misjudgment and omission of coupling anomalies and low accuracy of abnormal data filtering.

[0005] The technical solution of this invention is implemented as follows: On the one hand, this invention provides an automatic filtering and analysis method for abnormal data from current transformer tests, comprising the following steps:

[0006] Magnetic field data, electric field data, and thermal field data of the current transformer test are acquired. Feature extraction is performed on the magnetic field data, electric field data, and thermal field data respectively, and weighted fusion is performed to obtain a multi-physics field fusion feature vector.

[0007] An electromagnetic induction chain propagation graph is constructed, with magnetic field, electric field, and thermal field as graph nodes. Chain propagation edges and feedback coupling edges are set, and the weights of the edges in the electromagnetic induction chain propagation graph are calculated. A magnetic flux conservation constraint graph convolutional neural network is constructed. The propagation and evolution modes of transformer test anomalies in multiple physical fields are learned through the magnetic flux conservation constraint graph convolutional neural network, and the coupling anomaly feature vector is obtained.

[0008] An insulation degradation progressive perception classifier and a multi-level anomaly identification network for oil-paper insulation are constructed to classify the coupled anomaly feature vectors, identify normal data, single physical field anomaly data and multi-physical field coupled anomaly data. When the classification probability exceeds the dynamic adjustment threshold, the corresponding data is judged as anomaly data and filtered.

[0009] Physical mechanism analysis is performed on the filtered abnormal data to establish a mapping relationship between abnormal data categories and physical faults, thereby obtaining an anomaly detection report.

[0010] Based on the above technical solutions, preferably, the construction of the electromagnetic induction chain propagation graph involves using magnetic field, electric field, and thermal field as graph nodes, setting chain propagation edges and feedback coupling edges, calculating the weights of the edges in the electromagnetic induction chain propagation graph, constructing a magnetic flux conservation constraint graph convolutional neural network, and learning the propagation and evolution patterns of transformer test anomalies among multiple physical fields through the magnetic flux conservation constraint graph convolutional neural network to obtain coupled anomaly feature vectors, including:

[0011] Construct an electromagnetic induction chain propagation graph structure, using magnetic field, electric field, and thermal field as graph nodes, and setting chain propagation edges and feedback coupling edges. The chain propagation edges are used to represent abnormal unidirectional propagation paths, and the feedback coupling edges are used to represent bidirectional feedback relationships of electromagnetic induction.

[0012] Calculate the weights of the edges in the electromagnetic induction chain propagation graph, including the magneto-electric coupling weight based on the law of electromagnetic induction, the electro-thermal coupling weight based on the dielectric loss heating mechanism, and the magneto-thermal coupling weight based on the hysteresis loss heating mechanism.

[0013] A magnetic flux conservation-constrained graph convolutional neural network is constructed. During the graph convolution process, the magnetic flux conservation law is used as a constraint condition. A variable ratio error propagation matrix is ​​established to simulate the propagation of error among multiple test items.

[0014] The magnetic flux conservation constraint graph convolutional neural network learns the propagation and evolution patterns of anomalies in multiple physics fields and outputs coupled anomaly feature vectors.

[0015] Based on the above technical solutions, preferably, the construction of the electromagnetic induction chain propagation diagram structure includes:

[0016] Magnetic field nodes, electric field nodes, and thermal field nodes are used as the basic nodes of the graph, and each node carries the characteristic information of the corresponding physical field.

[0017] Chain-like propagation edges are established between nodes to simulate the unidirectional propagation process of anomalies from the primary side to the secondary side, including chain paths for the propagation of magnetic field anomalies to the electric field and the propagation of electric field anomalies to the thermal field.

[0018] Feedback coupling edges are established between nodes to simulate the bidirectional feedback relationship of electromagnetic induction, including the feedback effect of secondary load changes on the primary side.

[0019] Based on the above technical solutions, preferably, the chain propagation edge includes a primary side magnetic field anomaly propagation edge, a secondary side electric field anomaly propagation edge, and a hysteresis loss propagation edge, wherein:

[0020] The primary side magnetic field anomaly propagation edge is established by creating a directed edge from the magnetic field node to the electric field node to simulate the process of primary side excitation anomaly propagating to the secondary side through electromagnetic induction.

[0021] The abnormal propagation edge of the secondary electric field is established by creating a directed edge from the electric field node to the thermal field node to simulate the process of increased dielectric loss and subsequent temperature rise caused by abnormal secondary insulation.

[0022] Hysteresis loss propagation edge: Establish a directed edge from the magnetic field node to the thermal field node to simulate the process by which hysteresis loss directly causes the iron core to heat up.

[0023] Based on the above technical solutions, preferably, the construction of the insulation degradation progressive perception classifier and the oil-paper insulation multi-level anomaly identification network classifies the coupled anomaly feature vectors to identify normal data, single-physical-field anomaly data, and multi-physical-field coupled anomaly data. When the classification probability exceeds a dynamically adjusted threshold, the corresponding data is judged as anomaly data and filtered, including:

[0024] An insulation degradation progressive sensing classifier is constructed, and a degradation sensing neuron is designed. The bias term of the degradation sensing neuron is dynamically adjusted based on the equipment operating time of the current transformer to adapt to the insulation progressive degradation process.

[0025] A multi-level anomaly identification network for oil-paper insulation was established, including an oil quality deterioration detection layer, a paper insulation deterioration detection layer, and a systemic deterioration assessment layer, to handle insulation anomalies at different levels respectively;

[0026] A load history adaptive threshold adjustment mechanism is constructed to dynamically adjust the anomaly judgment threshold based on the historical load of the current transformer, and to give higher weight to recent loads.

[0027] The coupling anomaly feature vectors are classified to identify normal data, single-physics-field anomaly data, and multi-physics-field coupling anomaly data. When the classification probability exceeds the dynamically adjusted threshold, the corresponding data is judged as anomaly data and filtered.

[0028] Based on the above technical solutions, preferably, the insulation degradation progressive perception classifier adopts a time-varying weighted neural network structure, including:

[0029] A time-varying weight layer is constructed. The weight matrix of the time-varying weight layer varies exponentially according to the insulation aging time. The network structure of the time-varying weight layer includes three sub-modules connected in series: an aging time encoder, a weight modulator, and a classification decision unit.

[0030] An adaptive branch network for the degradation stage is constructed, which includes three parallel branches corresponding to the initial, middle and late degradation states of insulation, respectively. The active branch is dynamically selected through a degradation degree gating mechanism.

[0031] A progressive feature extraction layer is constructed, which adopts a residual connection structure. The residual weights dynamically decay with the degree of aging to simulate the irreversible degradation process of insulation performance.

[0032] Based on the above technical solutions, preferably, the construction of the time-varying weight layer includes:

[0033] The aging time encoder adopts a position encoding structure to convert the cumulative running time into a high-dimensional time feature vector. The aging time encoder includes a time embedding layer, a periodic encoding layer, and a time feature enhancement layer.

[0034] The weight modulator adopts a gated linear unit structure and controls the time-varying degree of the weights through a gating mechanism. The gating function is constructed based on the insulation degradation dynamics equation and includes a degradation rate gate, a degradation degree gate, and a weight update gate.

[0035] The classification decision-maker adopts a multi-head structure, with each head corresponding to a different anomaly type. The heads interact with each other through a cross-attention mechanism and output the classification result through weighted voting.

[0036] Based on the above technical solutions, preferably, the acquisition of magnetic field data, electric field data, and thermal field data from the current transformer test involves feature extraction and weighted fusion of the magnetic field data, electric field data, and thermal field data to obtain a multi-physics fusion feature vector, including:

[0037] Acquire magnetic field data, electric field data and thermal field data for current transformer tests. The magnetic field data includes excitation current waveform and hysteresis loop data. The electric field data includes dielectric loss factor and insulation resistance data. The thermal field data includes temperature rise test data.

[0038] A hysteresis nonlinear sensing convolutional neural network, a dielectric loss frequency response adaptive network, and a temperature rise time delay compensation network were constructed respectively to extract features from the magnetic field data, electric field data, and thermal field data, thereby obtaining magnetic field features, electric field features, and thermal field features. Among them, the hysteresis nonlinear sensing convolutional neural network integrates a hysteresis memory unit to record the historical trajectory of the magnetization process, the weights of the dielectric loss frequency response adaptive network are adaptively adjusted based on the test frequency, and the temperature rise time delay compensation network performs time delay compensation based on the thermal time constant of the transformer.

[0039] An attention mechanism is used to weight and fuse the magnetic field features, electric field features, and thermal field features to obtain a unified multiphysics field fusion feature vector.

[0040] Based on the above technical solutions, preferably, the step of performing physical mechanism analysis on the filtered abnormal data, establishing a mapping relationship between abnormal data categories and physical faults, and obtaining an anomaly detection report includes:

[0041] A multiphysics coupled fault diagnosis tree model is constructed. Filtered abnormal data is input into the multiphysics coupled fault diagnosis tree model, and the corresponding mapping relationship between abnormality category and physical fault is output. An abnormality detection report is generated based on multiphysics characteristics. The abnormality detection report includes abnormality type, confidence level, physical mechanism analysis and targeted processing suggestions.

[0042] On the other hand, the present invention also provides an automatic filtering and analysis system for abnormal data from current transformer tests, the system comprising:

[0043] The feature extraction module is used to acquire magnetic field data, electric field data and thermal field data of the current transformer test, extract features from the magnetic field data, electric field data and thermal field data respectively and perform weighted fusion to obtain a multi-physics field fusion feature vector;

[0044] The electromagnetic induction anomaly analysis module is used to construct an electromagnetic induction chain propagation graph, using magnetic field, electric field, and thermal field as graph nodes, setting chain propagation edges and feedback coupling edges, calculating the weights of the edges in the electromagnetic induction chain propagation graph, constructing a magnetic flux conservation constraint graph convolutional neural network, and learning the propagation and evolution mode of transformer test anomalies in multiple physical fields through the magnetic flux conservation constraint graph convolutional neural network to obtain the coupling anomaly feature vector;

[0045] The insulation degradation filtering module is used to construct an insulation degradation progressive perception classifier and a multi-level anomaly identification network for oil-paper insulation. It classifies the coupled anomaly feature vectors and identifies normal data, single physical field anomaly data and multi-physical field coupled anomaly data. When the classification probability exceeds the dynamic adjustment threshold, the corresponding data is judged as anomaly data and filtered.

[0046] The anomaly data analysis module is used to perform physical mechanism analysis on the filtered anomaly data, establish a mapping relationship between anomaly data categories and physical faults, and obtain anomaly detection reports.

[0047] The automatic filtering and analysis method and system for abnormal test data of current transformers of the present invention have the following advantages over the prior art:

[0048] (1) By integrating multi-physics field fusion feature extraction, electromagnetic induction chain propagation graph modeling, insulation degradation progressive perception classification and physical mechanism source analysis, based on the magnetic flux conservation constraint graph convolutional neural network to dynamically learn the abnormal propagation evolution mode, adopting the insulation degradation progressive perception classifier architecture, introducing load history adaptive and dynamic threshold adjustment mechanism, the coupled abnormal modeling of magnetic field, electric field and thermal field was completed, ensuring that the abnormal data filtering results are highly consistent with the actual physical state of the transformer, realizing high-precision automatic identification, filtering analysis and physical mechanism source diagnosis of transformer test abnormal data;

[0049] (2) By constructing an electromagnetic induction chain propagation graph structure and a magnetic flux conservation constraint graph convolutional neural network, the abnormal coupling propagation mechanism between multiple physical fields of the transformer is accurately modeled. By setting chain propagation edges and feedback coupling edges, the one-way propagation process of the abnormality from the primary side to the secondary side and the two-way feedback relationship of electromagnetic induction are simulated. At the same time, the edge weights are calculated based on the electromagnetic induction law, the dielectric loss heating mechanism and the hysteresis loss heating mechanism, which ensures the accuracy of physical modeling, realizes the accurate modeling of the error propagation law between multiple test items, and improves the accuracy of coupling abnormality feature extraction and physical interpretability.

[0050] (3) By constructing a time-varying weighted neural network structure for insulation degradation progressive perception classifier, the progressive characteristics of insulation degradation are dynamically perceived and adaptively classified. The cascade structure of aging time encoder, weight modulator and classification decision-maker is designed, as well as the parallel processing mechanism of adaptive branch network for degradation stage. The network parameters are dynamically adjusted according to the equipment operation time and the degree of degradation, and the abnormal patterns of different stages of insulation in the early, middle and late stages are accurately identified. The abnormal classification accuracy is adaptively optimized with the aging state of the equipment, which improves the accuracy and reliability of filtering abnormal data of long-term operating transformers. Attached Figure Description

[0051] 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.

[0052] Figure 1 This is a flowchart of an automatic filtering and analysis method for abnormal test data of a current transformer according to the present invention. Detailed Implementation

[0053] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides an automatic filtering and analysis method for abnormal data from current transformer tests, comprising the following steps:

[0055] Magnetic field data, electric field data, and thermal field data of the current transformer test are acquired. Feature extraction is performed on the magnetic field data, electric field data, and thermal field data respectively, and weighted fusion is performed to obtain a multi-physics field fusion feature vector.

[0056] An electromagnetic induction chain propagation graph is constructed, with magnetic field, electric field, and thermal field as graph nodes. Chain propagation edges and feedback coupling edges are set, and the weights of the edges in the electromagnetic induction chain propagation graph are calculated. A magnetic flux conservation constraint graph convolutional neural network is constructed. The propagation and evolution modes of transformer test anomalies in multiple physical fields are learned through the magnetic flux conservation constraint graph convolutional neural network, and the coupling anomaly feature vector is obtained.

[0057] An insulation degradation progressive perception classifier and a multi-level anomaly identification network for oil-paper insulation are constructed to classify the coupled anomaly feature vectors, identify normal data, single physical field anomaly data and multi-physical field coupled anomaly data. When the classification probability exceeds the dynamic adjustment threshold, the corresponding data is judged as anomaly data and filtered.

[0058] Physical mechanism analysis is performed on the filtered abnormal data to establish a mapping relationship between abnormal data categories and physical faults, resulting in an anomaly detection report that includes anomaly type, confidence level, physical mechanism analysis, and handling suggestions.

[0059] Specifically, this embodiment integrates multi-physics field fusion feature extraction, electromagnetic induction chain propagation graph modeling, insulation degradation progressive perception classification, and physical mechanism tracing analysis. Based on the magnetic flux conservation constraint graph convolutional neural network, it dynamically learns the abnormal propagation evolution mode and adopts an insulation degradation progressive perception classifier architecture. It introduces load history adaptive and dynamic threshold adjustment mechanisms to complete the coupled abnormal modeling of magnetic field, electric field, and thermal field. This ensures that the abnormal data filtering results are highly consistent with the actual physical state of the transformer, and realizes high-precision automatic identification, filtering analysis, and physical mechanism tracing diagnosis of abnormal transformer test data.

[0060] The process involves acquiring magnetic field, electric field, and thermal field data from the current transformer test, extracting features from each data point, and then weighting and fusing them to obtain a multi-physics fusion feature vector, including:

[0061] Acquire magnetic field data, electric field data and thermal field data for current transformer tests. The magnetic field data includes excitation current waveform and hysteresis loop data. The electric field data includes dielectric loss factor and insulation resistance data. The thermal field data includes temperature rise test data.

[0062] A hysteresis nonlinear sensing convolutional neural network, a dielectric loss frequency response adaptive network, and a temperature rise time delay compensation network were constructed respectively to extract features from the magnetic field data, electric field data, and thermal field data, thereby obtaining magnetic field features, electric field features, and thermal field features. Among them, the hysteresis nonlinear sensing convolutional neural network integrates a hysteresis memory unit to record the historical trajectory of the magnetization process, the weights of the dielectric loss frequency response adaptive network are adaptively adjusted based on the test frequency, and the temperature rise time delay compensation network performs time delay compensation based on the thermal time constant of the transformer.

[0063] An attention mechanism is used to weight and fuse the magnetic field features, electric field features, and thermal field features to obtain a unified multiphysics field fusion feature vector.

[0064] In one specific embodiment, a hysteresis nonlinear sensing convolutional neural network is constructed for magnetic field data. This network integrates a hysteresis memory unit to store and track the historical change trajectory of magnetic flux density and magnetic field strength during magnetization, and considers the historical dependence of hysteresis effect when processing excitation current waveform and hysteresis loop data.

[0065] An adaptive frequency response network for dielectric loss is constructed based on electric field data. This network automatically adjusts its internal weight parameters according to the test frequency to adapt to the changing characteristics of dielectric loss factor at different frequencies and to process dielectric loss factor and insulation resistance data.

[0066] A temperature rise time delay compensation network is constructed for thermal field data. This network is based on a long short-term memory network architecture and incorporates a thermal time delay compensation mechanism. The thermal time constant is calculated based on the thermal capacity and heat dissipation coefficient of the current transformer to compensate for the time delay of the temperature rise response relative to the electromagnetic change.

[0067] In one specific embodiment, the hysteresis nonlinear sensing convolutional neural network includes:

[0068] The hysteresis memory unit is used to store historical state information of the magnetization process, recording the values ​​of magnetic flux density B and magnetic field strength H at different times and their changing trends.

[0069] The hysteresis-sensing convolutional layer considers both the current input data and the historical state information in the hysteresis memory unit when performing convolution operations, and extracts historically dependent magnetic field features.

[0070] The hysteresis loop analysis module is specifically designed to process BH hysteresis loop data and identify abnormal changes in key parameters such as loop shape, area, and coercivity.

[0071] In one specific embodiment, the dielectric loss frequency response adaptive network includes:

[0072] The frequency detection module monitors the current test frequency in real time and transmits the frequency information to the weighted modulation module;

[0073] The weight modulation module dynamically adjusts the weight parameters of each layer of the network according to the medium polarization mechanism and frequency information, so that the network can adapt to the differences in dielectric loss characteristics at different frequencies.

[0074] The dispersion characteristic modeling module establishes a model of the relationship between dielectric loss factor and frequency to guide the weighted modulation process;

[0075] Based on the current test conditions and the operating status of the transformer, the importance weights of the magnetic field characteristics, electric field characteristics, and thermal field characteristics are dynamically calculated.

[0076] The feature vectors of the three physical fields are merged by a weighted fusion method to generate a unified feature representation containing multi-physics coupling information. The weight allocation reflects the degree of contribution of each physical field to the current abnormal state.

[0077] In one specific embodiment, the attention mechanism for adaptive physical field weights includes:

[0078] The test condition evaluation module analyzes the current test voltage, test frequency, ambient temperature and other conditions of the current transformer to determine the relative importance of each physical field;

[0079] The weight calculation module calculates the attention weight coefficients of the magnetic field, electric field, and thermal field based on the evaluation results of the experimental conditions and the numerical distribution of the characteristic vectors of each physical field.

[0080] The weight normalization module ensures that the sum of the weight coefficients of the three physical fields equals 1, thus maintaining the numerical stability of the feature fusion process.

[0081] In one specific embodiment, the feature fusion process includes:

[0082] The weighted summation operation linearly combines the magnetic field eigenvectors, electric field eigenvectors, and thermal field eigenvectors according to the calculated weight coefficients.

[0083] Feature dimension alignment ensures that the feature vectors of the three physical fields have the same dimension, which facilitates fusion operations.

[0084] The fusion result verification checks whether the fused feature vector contains key information from each physical field, ensuring the fusion quality.

[0085] Specifically, this embodiment solves the problem that traditional single-physics field feature extraction cannot handle hysteresis effect, frequency response and thermal time delay characteristics by constructing a hysteresis nonlinear perception convolutional neural network, a dielectric loss frequency response adaptive network and a temperature rise time delay compensation network, and realizes in-depth mining and accurate characterization of multi-physics field data features;

[0086] By integrating hysteresis memory units to record magnetization history trajectories, adaptively adjusting network weights based on experimental frequency, and compensating for temperature rise lag according to thermal time constants, specialized processing is performed on the historical dependence of magnetic fields, the frequency correlation of electric fields, and the time delay of thermal fields, respectively. In addition, a physical field weight adaptive attention mechanism is adopted to dynamically weight and fuse the three features. This not only fully preserves the key information of each physical field, but also adaptively adjusts the importance weights of each physical field according to experimental conditions and equipment status, thereby improving the representation ability and anomaly sensitivity of the multi-physics fusion feature vector.

[0087] The construction of the electromagnetic induction chain propagation graph involves using magnetic, electric, and thermal fields as graph nodes, setting chain propagation edges and feedback coupling edges, calculating the weights of the edges in the electromagnetic induction chain propagation graph, and constructing a magnetic flux conservation constraint graph convolutional neural network. This magnetic flux conservation constraint graph convolutional neural network is used to learn the propagation and evolution patterns of transformer test anomalies among multiple physical fields, obtaining coupled anomaly feature vectors, including:

[0088] Construct an electromagnetic induction chain propagation graph structure, using magnetic field, electric field, and thermal field as graph nodes, and setting chain propagation edges and feedback coupling edges. The chain propagation edges are used to represent abnormal unidirectional propagation paths, and the feedback coupling edges are used to represent bidirectional feedback relationships of electromagnetic induction.

[0089] Calculate the weights of the edges in the electromagnetic induction chain propagation graph, including the magneto-electric coupling weight based on the law of electromagnetic induction, the electro-thermal coupling weight based on the dielectric loss heating mechanism, and the magneto-thermal coupling weight based on the hysteresis loss heating mechanism.

[0090] A magnetic flux conservation-constrained graph convolutional neural network is constructed. During the graph convolution process, the magnetic flux conservation law is used as a constraint condition. A variable ratio error propagation matrix is ​​established to simulate the propagation of error among multiple test items.

[0091] The magnetic flux conservation constraint graph convolutional neural network learns the propagation and evolution patterns of anomalies in multiple physics fields and outputs coupled anomaly feature vectors.

[0092] In one specific embodiment, the formula for calculating the magneto-electric coupling weight is:

[0093] ;

[0094] in, For magneto-electric coupling weights, Based on the coupling coefficient, The rate of change of magnetic flux, For periodic coupling modulation coefficients, Angular frequency, For time, The phase angle, This is the excitation current deviation compensation coefficient. This is the actual excitation current. This is the rated excitation current.

[0095] The formula for calculating the electro-thermal coupling weight is:

[0096] ;

[0097] in, For electro-thermal coupling weights, The coefficient of thermal loss is the coefficient of thermal activity. For dielectric loss factor, This is the voltage value. This is the temperature-related correction factor. This is the actual temperature. For reference temperature, This is a correction factor for insulation aging time. This refers to the insulation aging time.

[0098] The formula for calculating the magnetic-thermal coupling weight is:

[0099] ;

[0100] in, For magnetic-thermal coupling weights, The coefficient of heat generated by hysteresis loss. This refers to the hysteresis loss power. For the maximum magnetic flux density, The saturation magnetic flux density The magnetization characteristic index, This is the frequency deviation influence coefficient. For actual frequency, For the rated frequency, It is the hyperbolic tangent activation function.

[0101] Specifically, the magneto-electric coupling weights in this implementation are based on Faraday's law of electromagnetic induction, and introduce a triple nonlinear modulation mechanism through periodic modulation terms. To capture the dynamic characteristics of alternating electromagnetic fields, an exponential function is used. Excitation current deviation compensation is achieved. By extending the static electromagnetic induction model to a dynamic model with time-varying parameters, the limitations of traditional electromagnetic induction calculation are overcome. The excitation current deviation is used as an adaptive weight adjustment factor, realizing operating condition adaptability. The influence of excitation deviation is described by an exponential decay function, which conforms to the physical laws of the current transformer excitation characteristic curve.

[0102] This embodiment of the electro-thermal coupling weight is based on the dielectric loss heating theory, introducing a dual influence mechanism of temperature feedback and aging time. It considers the interaction between dielectric loss and temperature through a temperature-related correction term and uses an exponential saturation function to model the cumulative effect of insulation aging. By constructing an electro-thermal coupling model that includes a temperature feedback mechanism, it solves the problem of traditional models neglecting the influence of temperature on dielectric loss. It incorporates equipment aging time into the physical field coupling calculation, achieving adaptability throughout the equipment's entire lifecycle. This is achieved through the exponential saturation function. Accurately model the nonlinear cumulative characteristics of insulation degradation over time.

[0103] This embodiment of the magnetic-thermal coupling weight, based on hysteresis loss theory, introduces the nonlinear effect of magnetic saturation and a frequency deviation compensation mechanism. It uses a power function term to characterize the nonlinear amplification effect of magnetic saturation on loss and employs a hyperbolic tangent function to model the frequency deviation effect. By introducing the magnetic saturation ratio as a key variable, it overcomes the limitation of traditional hysteresis loss calculations in insufficiently describing the saturation effect, and uses a power function... It describes the nonlinear characteristics of magnetic saturation, accurately reflects the properties of the iron core material, and cleverly realizes the bounded influence of frequency deviation through the hyperbolic tangent function, which takes into account both directionality and avoids the calculation distortion caused by extreme frequency deviation.

[0104] The construction of the electromagnetic induction chain propagation diagram structure includes:

[0105] Magnetic field nodes, electric field nodes, and thermal field nodes are used as the basic nodes of the graph, and each node carries the characteristic information of the corresponding physical field.

[0106] Chain-like propagation edges are established between nodes to simulate the unidirectional propagation process of anomalies from the primary side to the secondary side, including chain paths for the propagation of magnetic field anomalies to the electric field and the propagation of electric field anomalies to the thermal field.

[0107] Feedback coupling edges are established between nodes to simulate the bidirectional feedback relationship of electromagnetic induction, including the feedback effect of secondary load changes on the primary side.

[0108] The chain propagation edge includes a primary side magnetic field anomaly propagation edge, a secondary side electric field anomaly propagation edge, and a hysteresis loss propagation edge, wherein:

[0109] The primary side magnetic field anomaly propagation edge is established by creating a directed edge from the magnetic field node to the electric field node to simulate the process of primary side excitation anomaly propagating to the secondary side through electromagnetic induction.

[0110] The abnormal propagation edge of the secondary electric field is established by creating a directed edge from the electric field node to the thermal field node to simulate the process of increased dielectric loss and subsequent temperature rise caused by abnormal secondary insulation.

[0111] Hysteresis loss propagation edge: Establish a directed edge from the magnetic field node to the thermal field node to simulate the process by which hysteresis loss directly causes the iron core to heat up.

[0112] The feedback coupling edge includes a load feedback edge and a thermal feedback edge:

[0113] Load feedback edge: Establish a reverse edge from electric field node to magnetic field node to simulate the feedback effect of secondary load change on primary excitation current.

[0114] Thermal feedback edges are established, creating reverse edges from thermal field nodes to magnetic field nodes and electric field nodes to simulate the feedback effect of temperature rise on magnetic permeability and insulation performance.

[0115] The construction of the flux conservation constraint graph convolutional neural network includes:

[0116] In each iteration of graph convolution, it is mandatory that the magnetic flux flowing into and out of each node remains in balance;

[0117] Establish a transformer ratio error propagation matrix to quantify the propagation coefficients of errors between different test items, including the influence of excitation characteristic errors on transformer ratio testing and the influence of insulation abnormalities on excitation characteristics.

[0118] In one specific embodiment, the formula for calculating the convolution of the flux conservation constraint graph is:

[0119] ;

[0120] ;

[0121] in, For graph convolutional networks The node feature matrix of the layer, For activation function, For the normalized adjacency matrix, For graph convolutional networks The node feature matrix of the layer For the first The weight matrix of the layer, Here is the magnetic flux conservation constraint weight matrix. For graph convolutional networks The magnetic flux correction term of the layer, This is the flux correction coefficient. For the target total magnetic flux, For the first The node at the th Magnetic flux of the layer The total number of nodes. For the first The node at the th The feature vector of the layer, For the first The node at the th The feature vector of the layer.

[0122] Specifically, this embodiment of magnetic flux conservation-constrained graph convolution embeds physical conservation constraints within the standard graph convolution framework. It dynamically adjusts node features through a magnetic flux correction term to ensure that the convolution process satisfies magnetic flux conservation. A feature weight allocation mechanism is designed to achieve magnetic flux correction based on node importance. Through a hybrid modeling paradigm of "physical conservation constraints + graph convolutional neural network," it achieves an organic integration of deep learning and physical laws, and designs a magnetic flux correction term. This solves the key problem of physical consistency in deep learning models by using proportional allocation terms. The allocation of physical quantities based on feature importance ensures the rationality of the correction process.

[0123] In one specific embodiment, the formula for calculating the ratio error propagation matrix is:

[0124] ;

[0125] ;

[0126] ;

[0127] in, The variable ratio error propagation matrix, For the node To the node The fundamental error propagation coefficient, Forward propagation time-varying function, For feedback propagation time-varying function, , , , Forward propagation parameters, and To provide feedback propagation parameters, At the initial time point, This represents the load change. This is the load change direction function.

[0128] Specifically, the variable ratio error propagation matrix in this embodiment describes the error propagation relationship between multiple test items by creating a matrix structure. Two types of time-varying propagation functions are designed to model the damped oscillation characteristics of forward propagation and the cumulative effect of feedback propagation, respectively. A load change direction factor is introduced to achieve directional error propagation. By constructing the error propagation matrix structure, the limitations of traditional single error analysis are overcome, achieving system-level error propagation modeling and designing damped oscillation functions. It accurately characterizes the dynamic characteristics of error forward propagation by introducing a sign function and a cumulative effect term. This enables the modeling of the directional impact of load changes on error propagation.

[0129] This embodiment constructs an electromagnetic induction chain propagation graph structure and a magnetic flux conservation constraint graph convolutional neural network to accurately model the anomaly coupling propagation mechanism between multiple physical fields of the current transformer. By setting chain propagation edges and feedback coupling edges, it simulates the unidirectional propagation process of anomalies from the primary side to the secondary side and the bidirectional feedback relationship of electromagnetic induction. At the same time, it calculates edge weights based on the law of electromagnetic induction, the dielectric loss heating mechanism, and the hysteresis loss heating mechanism, ensuring the accuracy of physical modeling. It realizes the accurate modeling of the propagation law of errors among multiple test items, and improves the accuracy of coupling anomaly feature extraction and physical interpretability.

[0130] The construction of a progressive insulation degradation perception classifier and a multi-level anomaly identification network for oil-paper insulation classifies the coupled anomaly feature vectors, identifying normal data, single-physical-field anomaly data, and multi-physical-field coupled anomaly data. When the classification probability exceeds a dynamically adjusted threshold, the corresponding data is judged as anomaly data and filtered, including:

[0131] An insulation degradation progressive sensing classifier is constructed, and a degradation sensing neuron is designed. The bias term of the degradation sensing neuron is dynamically adjusted based on the equipment operating time of the current transformer to adapt to the insulation progressive degradation process.

[0132] A multi-level anomaly identification network for oil-paper insulation was established, including an oil quality deterioration detection layer, a paper insulation deterioration detection layer, and a systemic deterioration assessment layer, to handle insulation anomalies at different levels respectively;

[0133] A load history adaptive threshold adjustment mechanism is constructed to dynamically adjust the anomaly judgment threshold based on the historical load of the current transformer, and to give higher weight to recent loads.

[0134] The coupling anomaly feature vectors are classified to identify normal data, single-physics-field anomaly data, and multi-physics-field coupling anomaly data. When the classification probability exceeds the dynamically adjusted threshold, the corresponding data is judged as anomaly data and filtered.

[0135] The insulation degradation progressive perception classifier adopts a time-varying weighted neural network structure, including:

[0136] A time-varying weight layer is constructed. The weight matrix of the time-varying weight layer varies exponentially according to the insulation aging time. The network structure of the time-varying weight layer includes three sub-modules connected in series: an aging time encoder, a weight modulator, and a classification decision unit.

[0137] An adaptive branch network for the degradation stage is constructed, which includes three parallel branches corresponding to the initial, middle and late degradation states of insulation, respectively. The active branch is dynamically selected through a degradation degree gating mechanism.

[0138] A progressive feature extraction layer is constructed, which adopts a residual connection structure. The residual weights dynamically decay with the degree of aging to simulate the irreversible degradation process of insulation performance.

[0139] The construction of the time-varying weight layer includes:

[0140] The aging time encoder adopts a position encoding structure to convert the cumulative running time into a high-dimensional time feature vector. The aging time encoder includes a time embedding layer, a periodic encoding layer, and a time feature enhancement layer.

[0141] The weight modulator adopts a gated linear unit structure and controls the time-varying degree of the weights through a gating mechanism. The gating function is constructed based on the insulation degradation dynamics equation and includes a degradation rate gate, a degradation degree gate, and a weight update gate.

[0142] The classification decision-maker adopts a multi-head structure, with each head corresponding to a different anomaly type. The heads interact with each other through a cross-attention mechanism and output the classification result through weighted voting.

[0143] Specifically, this embodiment constructs an insulation degradation progressive perception classifier with a time-varying weighted neural network structure to dynamically perceive and adaptively classify the progressive characteristics of insulation degradation. It designs a series structure of aging time encoder, weight modulator and classification decision-maker, as well as a parallel processing mechanism of adaptive branch network for degradation stage. The network parameters are dynamically adjusted according to the equipment operating time and the degree of degradation to accurately identify abnormal patterns in different stages of insulation in the early, middle and late stages. This achieves adaptive optimization of the abnormal classification accuracy with the aging state of the equipment, and improves the accuracy and reliability of filtering abnormal data of long-term operating transformers.

[0144] The process involves performing physical mechanism analysis on the filtered abnormal data, establishing a mapping relationship between abnormal data categories and physical faults, and obtaining an anomaly detection report, including:

[0145] A multiphysics coupled fault diagnosis tree model is constructed. Filtered abnormal data is input into the multiphysics coupled fault diagnosis tree model, and the corresponding mapping relationship between abnormal categories and physical faults is output. An abnormality detection report is generated based on multiphysics features. The abnormality detection report includes abnormality type, confidence level, physical mechanism analysis, and targeted processing suggestions.

[0146] Establish a traceable anomaly data filtering log system. For each filtered sample, record the process of feature extraction, anomaly propagation path analysis, classification judgment, and dynamic threshold adjustment to generate a complete and traceable filtering decision log to ensure the traceability of subsequent analysis and the system.

[0147] Specifically, this embodiment solves the problem of traditional anomaly detection lacking physical mechanism explanation and process transparency by constructing a multi-physics field coupled fault diagnosis tree model and a traceable anomaly data filtering log system, and realizes intelligent diagnosis and full-process traceable management of anomaly data through automatic filtering.

[0148] By accurately mapping anomaly categories to specific physical faults through a fault diagnosis tree, a complete detection report is generated, including anomaly type, confidence level, physical mechanism analysis, and handling suggestions, providing anomaly filtering results. By recording the complete process of feature extraction, anomaly propagation path analysis, classification judgment, and threshold adjustment, a traceable decision log system is established, ensuring the transparency and interpretability of the entire anomaly data filtering process, and improving the system's engineering practicality and the accuracy of fault diagnosis.

[0149] This invention also provides an automatic filtering and analysis system for abnormal data from current transformer tests, the system comprising:

[0150] The feature extraction module is used to acquire magnetic field data, electric field data and thermal field data of the current transformer test, extract features from the magnetic field data, electric field data and thermal field data respectively and perform weighted fusion to obtain a multi-physics field fusion feature vector;

[0151] The electromagnetic induction anomaly analysis module is used to construct an electromagnetic induction chain propagation graph, using magnetic field, electric field, and thermal field as graph nodes, setting chain propagation edges and feedback coupling edges, calculating the weights of the edges in the electromagnetic induction chain propagation graph, constructing a magnetic flux conservation constraint graph convolutional neural network, and learning the propagation and evolution mode of transformer test anomalies in multiple physical fields through the magnetic flux conservation constraint graph convolutional neural network to obtain the coupling anomaly feature vector;

[0152] The insulation degradation filtering module is used to construct an insulation degradation progressive perception classifier and a multi-level anomaly identification network for oil-paper insulation. It classifies the coupled anomaly feature vectors and identifies normal data, single physical field anomaly data and multi-physical field coupled anomaly data. When the classification probability exceeds the dynamic adjustment threshold, the corresponding data is judged as anomaly data and filtered.

[0153] The anomaly data analysis module is used to perform physical mechanism analysis on the filtered anomaly data, establish a mapping relationship between anomaly data categories and physical faults, and obtain anomaly detection reports.

[0154] Specifically, this embodiment of an automatic filtering and analysis system for abnormal data in instrument transformer tests integrates multi-physics field fusion processing of feature extraction module, chain propagation graph modeling of electromagnetic induction anomaly analysis module, progressive perception classification of insulation degradation filtering module, and physical mechanism tracing analysis of anomaly data analysis module. Based on magnetic flux conservation constraint graph convolutional neural network and progressive perception classifier of insulation degradation, a modular intelligent analysis architecture is constructed. It adopts multi-physics field coupled fault diagnosis tree and traceable filtering log system, and introduces dynamic threshold adjustment and physical field weight adaptive mechanism to realize systematic coupled modeling of magnetic field-electric field-thermal field and intelligent anomaly identification. This ensures high consistency between the anomaly data filtering results and the physical state of the instrument transformer and full-process traceability, realizing high-precision automatic filtering, intelligent classification and identification, and engineering application deployment of abnormal data in instrument transformer tests.

[0155] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic filtering and analysis method for abnormal data in current transformer tests, characterized in that, Includes the following steps: Magnetic field data, electric field data, and thermal field data of the current transformer test are acquired. Feature extraction is performed on the magnetic field data, electric field data, and thermal field data respectively, and weighted fusion is performed to obtain a multi-physics field fusion feature vector. An electromagnetic induction chain propagation graph is constructed, with magnetic field, electric field, and thermal field as graph nodes. Chain propagation edges and feedback coupling edges are set, and the weights of the edges in the electromagnetic induction chain propagation graph are calculated. A magnetic flux conservation constraint graph convolutional neural network is constructed. The propagation and evolution modes of transformer test anomalies in multiple physical fields are learned through the magnetic flux conservation constraint graph convolutional neural network, and the coupling anomaly feature vector is obtained. The construction of the electromagnetic induction chain propagation graph involves using magnetic, electric, and thermal fields as graph nodes, setting chain propagation edges and feedback coupling edges, calculating the weights of the edges in the electromagnetic induction chain propagation graph, and constructing a magnetic flux conservation constraint graph convolutional neural network. This magnetic flux conservation constraint graph convolutional neural network is used to learn the propagation and evolution patterns of transformer test anomalies among multiple physical fields, obtaining coupled anomaly feature vectors, including: Construct an electromagnetic induction chain propagation graph structure, using magnetic field, electric field, and thermal field as graph nodes, and setting chain propagation edges and feedback coupling edges. The chain propagation edges are used to represent abnormal unidirectional propagation paths, and the feedback coupling edges are used to represent bidirectional feedback relationships of electromagnetic induction. Calculate the weights of the edges in the electromagnetic induction chain propagation graph, including the magneto-electric coupling weight based on the law of electromagnetic induction, the electro-thermal coupling weight based on the dielectric loss heating mechanism, and the magneto-thermal coupling weight based on the hysteresis loss heating mechanism. A magnetic flux conservation-constrained graph convolutional neural network is constructed. During the graph convolution process, the magnetic flux conservation law is used as a constraint condition. A variable ratio error propagation matrix is ​​established to simulate the propagation of error among multiple test items. The magnetic flux conservation constraint graph convolutional neural network learns the propagation and evolution patterns of anomalies in multiple physics fields and outputs coupled anomaly feature vectors. An insulation degradation progressive perception classifier and a multi-level anomaly identification network for oil-paper insulation are constructed to classify the coupled anomaly feature vectors, identify normal data, single physical field anomaly data and multi-physical field coupled anomaly data. When the classification probability exceeds the dynamic adjustment threshold, the corresponding data is judged as anomaly data and filtered. Physical mechanism analysis is performed on the filtered abnormal data to establish a mapping relationship between abnormal data categories and physical faults, thereby obtaining an anomaly detection report.

2. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 1, characterized in that, The construction of the electromagnetic induction chain propagation diagram structure includes: Magnetic field nodes, electric field nodes, and thermal field nodes are used as the basic nodes of the graph, and each node carries the characteristic information of the corresponding physical field. Chain-like propagation edges are established between nodes to simulate the unidirectional propagation process of anomalies from the primary side to the secondary side, including chain paths for the propagation of magnetic field anomalies to the electric field and the propagation of electric field anomalies to the thermal field. Feedback coupling edges are established between nodes to simulate the bidirectional feedback relationship of electromagnetic induction, including the feedback effect of secondary load changes on the primary side.

3. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 2, characterized in that, The chain propagation edge includes a primary side magnetic field anomaly propagation edge, a secondary side electric field anomaly propagation edge, and a hysteresis loss propagation edge, wherein: The primary side magnetic field anomaly propagation edge is established by creating a directed edge from the magnetic field node to the electric field node to simulate the process of primary side excitation anomaly propagating to the secondary side through electromagnetic induction. The abnormal propagation edge of the secondary electric field is established by creating a directed edge from the electric field node to the thermal field node to simulate the process of increased dielectric loss and subsequent temperature rise caused by abnormal secondary insulation. Hysteresis loss propagation edge: Establish a directed edge from the magnetic field node to the thermal field node to simulate the process by which hysteresis loss directly causes the iron core to heat up.

4. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 1, characterized in that, The construction of a progressive insulation degradation perception classifier and a multi-level anomaly identification network for oil-paper insulation classifies the coupled anomaly feature vectors, identifying normal data, single-physical-field anomaly data, and multi-physical-field coupled anomaly data. When the classification probability exceeds a dynamically adjusted threshold, the corresponding data is judged as anomaly data and filtered, including: An insulation degradation progressive sensing classifier is constructed, and a degradation sensing neuron is designed. The bias term of the degradation sensing neuron is dynamically adjusted based on the equipment operating time of the current transformer to adapt to the insulation progressive degradation process. A multi-level anomaly identification network for oil-paper insulation was established, including an oil quality deterioration detection layer, a paper insulation deterioration detection layer, and a systemic deterioration assessment layer, to handle insulation anomalies at different levels respectively; A load history adaptive threshold adjustment mechanism is constructed to dynamically adjust the anomaly judgment threshold based on the historical load of the current transformer, and to give higher weight to recent loads. The coupling anomaly feature vectors are classified to identify normal data, single-physics-field anomaly data, and multi-physics-field coupling anomaly data. When the classification probability exceeds the dynamically adjusted threshold, the corresponding data is judged as anomaly data and filtered.

5. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 4, characterized in that, The insulation degradation progressive perception classifier adopts a time-varying weighted neural network structure, including: A time-varying weight layer is constructed. The weight matrix of the time-varying weight layer varies exponentially according to the insulation aging time. The network structure of the time-varying weight layer includes three sub-modules connected in series: an aging time encoder, a weight modulator, and a classification decision unit. An adaptive branch network for the degradation stage is constructed, which includes three parallel branches corresponding to the initial, middle and late degradation states of insulation, respectively. The active branch is dynamically selected through a degradation degree gating mechanism. A progressive feature extraction layer is constructed, which adopts a residual connection structure. The residual weights dynamically decay with the degree of aging to simulate the irreversible degradation process of insulation performance.

6. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 5, characterized in that, The construction of the time-varying weight layer includes: The aging time encoder adopts a position encoding structure to convert the cumulative running time into a high-dimensional time feature vector. The aging time encoder includes a time embedding layer, a periodic encoding layer, and a time feature enhancement layer. The weight modulator adopts a gated linear unit structure and controls the time-varying degree of the weights through a gating mechanism. The gating function is constructed based on the insulation degradation dynamics equation and includes a degradation rate gate, a degradation degree gate, and a weight update gate. The classification decision-maker adopts a multi-head structure, with each head corresponding to a different anomaly type. The heads interact with each other through a cross-attention mechanism and output the classification result through weighted voting.

7. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 1, characterized in that, The process involves acquiring magnetic field, electric field, and thermal field data from the current transformer test, extracting features from each data point, and then weighting and fusing them to obtain a multi-physics fusion feature vector, including: Acquire magnetic field data, electric field data and thermal field data for current transformer tests. The magnetic field data includes excitation current waveform and hysteresis loop data. The electric field data includes dielectric loss factor and insulation resistance data. The thermal field data includes temperature rise test data. A hysteresis nonlinear sensing convolutional neural network, a dielectric loss frequency response adaptive network, and a temperature rise time delay compensation network were constructed respectively to extract features from the magnetic field data, electric field data, and thermal field data, thereby obtaining magnetic field features, electric field features, and thermal field features. Among them, the hysteresis nonlinear sensing convolutional neural network integrates a hysteresis memory unit to record the historical trajectory of the magnetization process, the weights of the dielectric loss frequency response adaptive network are adaptively adjusted based on the test frequency, and the temperature rise time delay compensation network performs time delay compensation based on the thermal time constant of the transformer. An attention mechanism is used to weight and fuse the magnetic field features, electric field features, and thermal field features to obtain a unified multiphysics field fusion feature vector.

8. The automatic filtering and analysis method for abnormal test data of current transformers as described in claim 1, characterized in that, The process involves performing physical mechanism analysis on the filtered abnormal data, establishing a mapping relationship between abnormal data categories and physical faults, and obtaining an anomaly detection report, including: A multiphysics coupled fault diagnosis tree model is constructed. Filtered abnormal data is input into the multiphysics coupled fault diagnosis tree model, and the corresponding mapping relationship between abnormality category and physical fault is output. An abnormality detection report is generated based on multiphysics characteristics. The abnormality detection report includes abnormality type, confidence level, physical mechanism analysis and targeted processing suggestions.

9. An automatic filtering and analysis system for abnormal data from instrument transformer tests, used to execute the automatic filtering and analysis method for abnormal data from instrument transformer tests as described in any one of claims 1-8, characterized in that, The system includes: The feature extraction module is used to acquire magnetic field data, electric field data and thermal field data of the current transformer test, extract features from the magnetic field data, electric field data and thermal field data respectively and perform weighted fusion to obtain a multi-physics field fusion feature vector; The electromagnetic induction anomaly analysis module is used to construct an electromagnetic induction chain propagation graph, using magnetic field, electric field, and thermal field as graph nodes, setting chain propagation edges and feedback coupling edges, calculating the weights of the edges in the electromagnetic induction chain propagation graph, constructing a magnetic flux conservation constraint graph convolutional neural network, and learning the propagation and evolution mode of transformer test anomalies in multiple physical fields through the magnetic flux conservation constraint graph convolutional neural network to obtain the coupling anomaly feature vector; The insulation degradation filtering module is used to construct an insulation degradation progressive perception classifier and a multi-level anomaly identification network for oil-paper insulation. It classifies the coupled anomaly feature vectors and identifies normal data, single physical field anomaly data and multi-physical field coupled anomaly data. When the classification probability exceeds the dynamic adjustment threshold, the corresponding data is judged as anomaly data and filtered. The anomaly data analysis module is used to perform physical mechanism analysis on the filtered anomaly data, establish a mapping relationship between anomaly data categories and physical faults, and obtain anomaly detection reports.

Images