A method and apparatus for monitoring a current transformer

Abstract

The application discloses a kind of monitoring method and equipment of current transformer, applied to electric power equipment online monitoring technical field, comprising: obtaining the physical constraint graph neural network of training completion;Physical constraint graph neural network is configured with the component coupling correlation graph of current transformer;Obtain the wideband impedance spectrum sequence of the current transformer to be measured, carry out frequency domain coding, obtain global frequency domain feature and map as the initial node feature of each node;According to the coupling weight of component coupling correlation graph and each edge, the initial node feature of each node is carried out multi-layer graph attention propagation;The node feature of each node after completing multi-layer graph attention propagation is aggregated, and the global deep feature obtained is used to determine the fault mode classification probability distribution, and the current operating state of the current transformer to be measured is determined according to the fault mode classification probability distribution. The application of the scheme can effectively realize the monitoring of current transformer, realize fault identification, and improve the accuracy and physical interpretability.

Description

Technical Field

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

[0002] Current transformers are crucial devices in power systems, used for current transformation, metering, and relay protection. Accurate monitoring of their operating status directly impacts the accuracy of power grid metering and the reliability of power supply. Abnormal conditions such as winding deformation, insulation degradation, and inter-turn short circuits not only alter the current transformer's transmission characteristics but can also increase metering errors, thereby affecting the accuracy of energy metering and the reliability of protection actions. Therefore, uninterrupted and continuous monitoring of the operating status of current transformers is necessary.

[0003] Current solutions often involve acquiring the broadband impedance spectrum of the current transformer and then determining the equipment status based on preset thresholds or relevant classification models. However, the port impedance response of a current transformer is the result of the combined action of multiple internal components. The same fault may exhibit different impedance spectrum changes under different load conditions, and different faults may also produce similar local impedance spectrum characteristics. Therefore, such solutions have limited accuracy.

[0004] Other solutions have introduced deep learning technology to predict the error of current transformers, but they still have problems such as poor generalization of pure data-driven models and insufficient integration of physical mechanisms and data-driven approaches, which makes the models prone to producing spurious correlation prediction results that violate physical laws.

[0005] In summary, how to effectively monitor current transformers and ensure the reliability of the monitoring is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide a monitoring method and device for current transformers, so as to effectively monitor current transformers and ensure the reliability of monitoring.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, the present invention provides a method for monitoring a current transformer, comprising:

[0009] Obtain a trained physical constraint graph neural network; wherein the physical constraint graph neural network is configured with a component coupling association graph of the current transformer, the component coupling association graph uses the components of the current transformer as nodes, the physical coupling relationship between different components as edges, and the coupling weight of each edge is used to reflect the degree of coupling between the corresponding components.

[0010] Obtain the broadband impedance spectrum sequence of the current transformer under test;

[0011] The broadband impedance spectrum sequence is frequency-domain encoded by the physical constraint graph neural network to obtain global frequency domain features and map them to the initial node features of each node.

[0012] Based on the component coupling association graph and the coupling weight of each edge, multi-layer graph attention propagation is performed on the initial node features of each node; wherein, in each layer of graph attention propagation, the attention weight between connected nodes is determined based on the current node features of connected nodes and the coupling weight of the corresponding edge, and the node features of each node are updated based on the attention weight.

[0013] The node features of each node after the multi-layer graph attention propagation is completed are aggregated to obtain global depth features;

[0014] Based on the global depth features, a fault mode classification probability distribution for the current transformer under test is obtained, and the current operating state of the current transformer under test is determined according to the fault mode classification probability distribution.

[0015] In one implementation, obtaining the broadband impedance spectrum sequence of the current transformer under test includes:

[0016] When the current transformer under test is in a energized operating state, a wideband scanning signal with a frequency varying within a preset frequency range is injected into the current transformer under test through an excitation source.

[0017] The impedance amplitude and phase angle of the current transformer under test are obtained at multiple sampling frequencies, and the broadband impedance spectrum sequence is formed by each sampling frequency and the corresponding impedance amplitude and phase angle.

[0018] In one implementation, the broadband impedance spectrum sequence is frequency-domain encoded using the physical constraint graph neural network to obtain global frequency domain features and map them to the initial node features of each node, including:

[0019] According to the order of the sampling frequencies, the corresponding impedance amplitudes and phase angles are arranged to obtain the impedance amplitude sequence and the phase angle sequence;

[0020] The impedance magnitude sequence and the phase angle sequence are jointly encoded by the frequency domain encoder in the physical constraint graph neural network to obtain global frequency domain features.

[0021] The global frequency domain features are linearly projected through the node projection layer corresponding to each node to obtain the initial node features of each node.

[0022] In one embodiment, the components of the current transformer include a primary winding, a secondary winding, a core main column, a core side yoke, a primary insulation layer, a secondary insulation layer, and secondary circuit terminals.

[0023] In the component coupling association diagram, the coupling weight of each edge is the electromagnetic coupling weight, heat conduction weight and mechanical coupling weight between the corresponding components obtained through multiphysics simulation, and the coupling weight is obtained by weighted summation according to the corresponding physical field contribution coefficients.

[0024] The electromagnetic coupling weight reflects the mutual inductance coefficient between the corresponding components, the thermal conduction weight reflects the thermal resistance between the corresponding components, and the mechanical coupling weight reflects the structural stiffness coefficient between the corresponding components.

[0025] In one implementation, during attention propagation at each graph layer, the attention weights between connected nodes are determined based on the current node characteristics of the connected nodes and the coupling weights of the corresponding edges, including:

[0026] For a target node, the nodes connected to the target node are considered as the neighbor nodes of the target node; wherein, the target node can be any node;

[0027] A linear transformation is performed on the current node features of the target node and each of the neighboring nodes, respectively;

[0028] For any of the neighboring nodes, the node features of the target node after linear transformation, the node features of the neighboring nodes after linear transformation, and the coupling weight of the edge between the target node and the neighboring nodes are concatenated into a vector, and the concatenation result is nonlinearly mapped to obtain the unnormalized attention coefficient between the target node and the neighboring nodes.

[0029] The unnormalized attention coefficients between the target node and each of the neighboring nodes are normalized to obtain the attention weights between the target node and each of the neighboring nodes.

[0030] In one implementation, during attention propagation at each graph layer, updating the node features of each node based on the attention weights includes:

[0031] Based on the attention weights between the target node and each of the neighboring nodes, the node features of each of the neighboring nodes after linear transformation are weighted and aggregated to obtain the neighborhood feature aggregation result.

[0032] The self-connection weight matrix of this layer is used to perform a self-connection transformation on the current node features of the target node to obtain the self-connection transformation result.

[0033] The self-connection transformation result is fused with the neighborhood feature aggregation result and activated by an activation function to obtain the updated node features of the target node.

[0034] In one embodiment, the fault mode classification probability distribution includes probabilities corresponding to winding deformation, insulation degradation, inter-turn short circuit, abnormal secondary circuit contact resistance, overload, and normal state, respectively.

[0035] The current operating state of the current transformer under test is determined based on the fault mode classification probability distribution, including:

[0036] The maximum probability in the failure mode classification probability distribution is used as the classification confidence level;

[0037] When the classification confidence level is lower than the confidence threshold, the current operating state of the current transformer under test is determined to be an unknown abnormal state, and the broadband impedance spectrum sequence is stored in the buffer to be reviewed.

[0038] When the classification confidence level is not lower than the confidence level threshold, the state corresponding to the highest probability is determined as the current operating state of the current transformer under test.

[0039] In one implementation, it further includes:

[0040] The global depth features are processed to obtain the metering error prediction value of the current transformer under test; wherein the metering error prediction value includes the ratio difference prediction value and the angle difference prediction value.

[0041] In one implementation, the training phase of the physical constraint graph neural network includes a first training phase and a second training phase.

[0042] The first loss function used in the first training phase is a weighted sum of fault classification loss and physical constraint loss; the second loss function used in the second training phase is a weighted sum of fault classification loss, measurement error regression loss, contrast loss, and physical constraint loss.

[0043] In one implementation, the physical constraint loss includes physical propagation consistency loss and energy dissipation consistency loss;

[0044] The methods for determining the physical propagation consistency loss include:

[0045] Construct a normalized coupling weight matrix based on the coupling weights of each edge;

[0046] In the attention propagation of each layer, the input node features of each node in this layer are weighted and aggregated according to the normalized coupling weight matrix to obtain the physical propagation features of each node.

[0047] Based on the difference between the output node characteristics of each node at this layer and the corresponding physical propagation characteristics, the propagation consistency loss corresponding to this layer is determined, and the propagation consistency losses corresponding to each layer are aggregated to obtain the physical propagation consistency loss.

[0048] The methods for determining the consistency loss of energy dissipation include:

[0049] The impedance real part characterization quantity is determined based on the real part of the impedance in the broadband impedance spectrum sequence of the training samples, and the node characteristic characterization quantity is determined based on the node characteristics of the nodes that generate active power loss.

[0050] The energy dissipation consistency loss is determined based on the difference between the real part of the impedance and the node characteristic.

[0051] In one implementation, the method for determining the contrast loss includes:

[0052] Anchor samples, positive samples, and multiple negative samples are selected from the dataset used for the second training phase; wherein the positive samples and the anchor samples have the same fault mode label and come from different load conditions or different current transformers, and each negative sample and the anchor sample have different fault mode labels.

[0053] The anchor point samples, the positive samples, and each of the negative samples are respectively input into the physical constraint graph neural network to obtain the corresponding global depth features;

[0054] Determine the first similarity between the global depth feature corresponding to the anchor point sample and the global depth feature corresponding to the positive sample, and the second similarity between the global depth feature corresponding to the anchor point sample and the global depth feature corresponding to each of the negative samples;

[0055] The contrast loss is determined based on the first similarity and each of the second similarities, so as to increase the feature similarity between samples with the same fault mode label and decrease the feature similarity between samples with different fault mode labels.

[0056] In one embodiment, the determination of the measurement error regression loss includes:

[0057] The training samples are input into the physical constraint graph neural network to obtain the ratio difference prediction value and angle difference prediction value corresponding to the training samples;

[0058] The ratio difference loss is determined based on the difference between the predicted ratio difference value and the corresponding true ratio difference value, and the angle difference loss is determined based on the difference between the predicted angle difference value and the corresponding true angle difference value.

[0059] The measurement error regression loss is obtained by weighted summation of the ratio difference loss and the angle difference loss.

[0060] In a second aspect, the present invention provides a monitoring device for a current transformer, comprising:

[0061] Memory, used to store computer programs;

[0062] A processor for executing the computer program to implement the steps of the current transformer monitoring method as described above.

[0063] By applying the technical solution provided in the embodiments of the present invention, a graph neural network based on physical constraints is set up, which embeds physical coupling weights into the graph attention mechanism to guide message passing, greatly improving the physical interpretability of the model and making it less likely to produce spurious correlation prediction results that violate physical laws, thus ensuring the monitoring reliability of the current transformer.

[0064] Specifically, the physical constraint graph neural network of this application is configured with a component coupling association graph of the current transformer. The component coupling association graph uses the components of the current transformer as nodes and the physical coupling relationships between different components as edges. The coupling weight of each edge reflects the degree of coupling between the corresponding components. Therefore, in this application, the compositional relationships of the internal components of the current transformer, the propagation paths of fault effects, and the coupling strength corresponding to the paths can be introduced into the monitoring process as stable physical priors, rather than establishing data associations without physical meaning based solely on port measurement data. During monitoring, the broadband impedance spectrum sequence of the current transformer under test can be obtained. Then, the broadband impedance spectrum sequence is frequency-domain encoded using the physical constraint graph neural network to obtain global frequency domain features, which are then mapped to the initial node features of each node, thus converting the broadband impedance spectrum sequence into a feature representation. Afterwards, multi-layer graph attention propagation is performed on the initial node features of each node based on the component coupling association graph and the coupling weights of each edge. In this application's scheme, during attention propagation at each layer of the graph, the attention weights between connected nodes are determined based on the current node features of the connected nodes and the coupling weights of the corresponding edges. The node features of each node are then updated based on these attention weights. This setup ensures that feature propagation is jointly regulated by the component state presented by the current sample and the coupling between components. This allows fault information to be transmitted layer by layer along paths with actual physical coupling relationships, thus highlighting fault-related propagation paths and suppressing weakly correlated interference paths. Subsequently, the node features of each node after multi-layer graph attention propagation are aggregated to obtain global depth features. Based on these global depth features, the fault mode classification probability distribution for the current transformer under test can be obtained, and the current operating state of the current transformer under test can be determined according to the fault mode classification probability distribution.

[0065] In summary, this solution can effectively monitor current transformers, identify faults, and improve the accuracy and physical interpretability of current transformer status monitoring. It is less likely to produce spurious correlation prediction results that violate physical laws, thus ensuring the reliability of current transformer monitoring. Attached Figure Description

[0066] 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.

[0067] Figure 1 The flowchart illustrates the implementation of a current transformer monitoring method according to a specific embodiment of the present invention.

[0068] Figure 2 This is a schematic diagram of the structure of a monitoring device for a current transformer provided in a specific embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0070] Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an implementation method for monitoring a current transformer according to a specific embodiment of the present invention. The monitoring method for the current transformer may include the following steps:

[0071] Step S101: Obtain the trained physical constraint graph neural network. The physical constraint graph neural network is configured with a component coupling relationship graph of the current transformer. The component coupling relationship graph uses the components of the current transformer as nodes and the physical coupling relationships between different components as edges. The coupling weight of each edge is used to reflect the degree of coupling between the corresponding components.

[0072] This application addresses the fact that traditional deep learning models typically treat impedance spectra as one-dimensional sequences or two-dimensional images. While this approach can extract frequency domain statistical features, it neglects the physical connections and fault propagation paths between functional components within the current transformer (CT). In reality, CT winding deformation often begins with mechanical stress concentration caused by magnetostriction of the core, subsequently propagating to the winding support through structural coupling, ultimately manifesting as a phase shift in a specific frequency band on the impedance spectrum. In other words, the propagation of this fault within the device exhibits significant topological and directional characteristics, and graph structures provide a natural mathematical abstraction for describing such spatial relationships. Therefore, this application abstracts the key functional components of the current transformer as graph nodes, thereby constructing a heterogeneous attribute graph, referred to as the component coupling relationship graph of the current transformer.

[0073] The component coupling diagram can be represented as follows: ,in, For a set of nodes, Let be the set of edges. The coupling weight matrix is ​​used. Each node corresponds one-to-one with an internal component of the current transformer. When at least one physical coupling relationship exists between two components—electromagnetic coupling, thermal conduction coupling, or mechanical coupling—an edge is set between the corresponding nodes. The coupling weight of the edge is used to quantify the overall coupling degree between the corresponding components. In this application, the topology and coupling weights of the component coupling graph can be configured before model training and remain unchanged during inference, while the node characteristics of each node dynamically change with each input broadband impedance spectrum sequence. Furthermore, it should be noted that different models of current transformers may have different physical structures. Therefore, in some embodiments, for a specific model of current transformer, a corresponding component coupling graph can be constructed, and a physical constraint graph neural network configured with that component coupling graph can be trained to further improve accuracy. Of course, in some embodiments, considering that the structure of current transformers is generally consistent, the component coupling graph can also be constructed according to a general structure. As long as the current transformer does not have a special structure, reliable state monitoring can be achieved based on this application.

[0074] The physical constraint graph neural network of this application is configured with a component coupling association graph of the current transformer, enabling fault information to be transmitted layer by layer along paths with actual physical coupling relationships during feature propagation. Of course, in other embodiments described below, loss constraints can also be applied based on physical coupling relationships to further ensure that the monitoring results of this application conform to physical laws.

[0075] In one specific embodiment of the present invention, the components of the current transformer include a primary winding, a secondary winding, a core main column, a core side yoke, a primary insulation layer, a secondary insulation layer, and secondary circuit terminals.

[0076] In the component coupling relationship diagram, the coupling weight of each side is the electromagnetic coupling weight, heat conduction weight and mechanical coupling weight between the corresponding components obtained through multiphysics simulation, and the coupling weight is obtained by weighted summation according to the corresponding physical field contribution coefficient.

[0077] Electromagnetic coupling weight reflects the mutual inductance coefficient between corresponding components, thermal conduction weight reflects the thermal resistance between corresponding components, and mechanical coupling weight reflects the structural stiffness coefficient between corresponding components.

[0078] The node set is represented as In this implementation, the component coupling diagram specifically includes seven nodes, namely N. v =7, refer to Table 1, which corresponds to the primary winding, secondary winding, iron core main column, iron core side yoke, primary insulation layer, secondary insulation layer and secondary circuit terminal in sequence.

[0079] Table 1: Schematic diagram of node sets in a specific scenario

[0080]

[0081] As shown in Table 1, the primary and secondary windings are the main carriers of electromagnetic, thermal, and mechanical coupling. The iron core main column is the main magnetic circuit component, primarily exhibiting magnetic, thermal, and mechanical effects. The iron core side yoke is an auxiliary magnetic circuit component, primarily exhibiting magnetic and mechanical effects. The primary and secondary insulation layers primarily exhibit dielectric, thermal, and mechanical effects. The secondary circuit terminals primarily exhibit electrical and thermal effects and are sensitive to changes in contact resistance. The node division in this implementation can balance the component resolution required for state recognition and the scale of graph computation.

[0082] In practical applications, to obtain the physical coupling relationships between components, a digital twin model of the current transformer's electromagnetic-thermal-mechanical multiphysics field can be pre-established. For different fault modes, broadband impedance spectrum scanning can be performed under different operating conditions to obtain a physically labeled simulation dataset D. sim The simulation process is not a simple calculation of a single physical field, but fully considers the bidirectional coupling between various physical phenomena: the Joule heat and core loss obtained from the electromagnetic field calculation are transferred to the temperature field as heat sources; the temperature rise distribution obtained from the temperature field calculation is fed back to the mechanical field through the coefficient of thermal expansion, generating thermal stress; the displacement and deformation in the mechanical field, in turn, change the geometric boundary conditions of the electromagnetic field calculation, thereby affecting the magnetic field distribution and impedance response.

[0083] Specifically, electromagnetic field calculations can obtain the winding current density J(f) and core magnetic flux density B(f), and the heat source of the thermal field is formed based on the Joule loss of the winding and the loss of the core. The thermal field calculations obtain the temperature rise ΔT(r) at each location. The thermal expansion caused by the temperature rise further acts on the mechanical field, and the mechanical field calculations obtain the winding electromagnetic force F. Lorentz and the magnetostrictive stress σ of the iron core mag This allows for the acquisition of component displacement or deformation. Component displacement and deformation can also provide feedback to change the geometry or boundary conditions used in the electromagnetic field calculation, thereby creating a two-way coupling between the electromagnetic field, thermal field, and mechanical field.

[0084] In this implementation, based on the multiphysics simulation results described above, not only can a large number of labeled training samples be obtained, but also the coupling strength coefficients between various components can be quantitatively extracted. These coefficients are not set by human experience, but can be directly derived from simulation calculations, thus providing physical priors for the initialization of edge weights in the subsequent graph neural network, avoiding the blindness of the traditional graph neural network where edge weights rely entirely on random initialization or data-driven approaches.

[0085] Specifically, for any two components i and j, the electromagnetic coupling weight w can be obtained through multiphysics simulation. ij (em) Heat conduction weight w ij (th) and mechanical coupling weight w ij (me) Among them, the mutual inductance coefficient M between components can be used as a reference. ij Determine the electromagnetic coupling weights; for example, the specific calculation method could be: w ij (em) =M ij / M max M max This represents the maximum mutual inductance coefficient of each pair of components with electromagnetic coupling relationships in the component coupling diagram. This can be determined based on the thermal resistance R between the components. th，ij Determine the heat conduction weight; for example, the specific calculation method can be: , This represents the maximum reciprocal of the thermal resistance of each component pair. It can be determined based on the structural stiffness coefficient k between the components. ij Determine the mechanical coupling weights, for example, w ij (me) =k ij / k max k max This represents the maximum structural stiffness coefficient for each component pair. It can be seen that, through the above normalization, physical quantities with different dimensions can be converted into a numerical range that can be weighted and combined.

[0086] In this application, the coupling weight of an edge refers to the comprehensive coupling weight of the edge, which can be obtained by weighted summation according to the corresponding physical field contribution coefficients. For example, in a specific implementation, it can be expressed as:

[0087] w ij =γ1·w ij (em) +γ2·w ij (th) +γ3·w ij (me) Here, γ1, γ2, and γ3 are the physical field contribution coefficients corresponding to the electromagnetic field, thermal field, and mechanical field, respectively. In practical applications, γ1, γ2, and γ3 are usually set to be no less than 0, and their sum is 1. The contribution coefficients of each physical field can be set based on simulation sensitivity analysis, expert experience, or training data. Furthermore, an adjacency matrix can be defined. and coupling weight matrix If there is a physical coupling relationship between node i and node j, then A in the adjacency matrix... ij =1, W in the coupling weight matrix ij =w ij If there is no physical coupling, then A ij =0 and W ij =0.

[0088] As described above, multiphysics digital twin models can be used to generate training samples. For example, in one specific implementation, the fault mode label space is defined as follows: Where K=6, corresponding to 5 fault modes (winding deformation, insulation degradation, inter-turn short circuit, abnormal secondary circuit contact resistance, and overload) and 1 normal state. For example, broadband impedance spectrum scanning can be performed under different load levels such as 20%, 50%, 80%, 100%, and 120% of the rated load to obtain a simulation dataset with physical labels. For example, in one case, there are no fewer than 20 samples in each category, covering 5 different load conditions, for a total of no fewer than 120 simulation samples.

[0089] Step S102: Obtain the broadband impedance spectrum sequence of the current transformer under test.

[0090] A broadband impedance spectrum sequence typically includes multiple sampling frequencies and the impedance amplitude and phase angle measured at each sampling frequency, used to reflect the port impedance response of the current transformer under test from low to high frequencies. It should be noted that the internal magnetic field, temperature field, and stress field data obtained through multiphysics simulation described above are used to pre-construct the component coupling correlation diagram. However, during the online monitoring of the current transformer, only the broadband impedance spectrum sequence of the current transformer under test needs to be obtained and input into the physical constraint graph neural network.

[0091] In one specific embodiment of the present invention, step S102 may specifically include:

[0092] When the current transformer under test is in a energized operating state, a wideband scanning signal with a frequency varying within a preset frequency range is injected into the current transformer under test through an excitation source.

[0093] The impedance amplitude and phase angle of the current transformer under test are obtained at multiple sampling frequencies, and the broadband impedance spectrum sequence is formed by each sampling frequency and the corresponding impedance amplitude and phase angle.

[0094] Specifically, when the current transformer under test (DUT) is energized, a broadband scanning signal can be injected into the measurement port of the DUT by controlling the excitation source. A non-invasive excitation source is typically used. The frequency of the broadband scanning signal can vary within a preset frequency range. When scanning to any sampling frequency, the voltage and current responses corresponding to that sampling frequency can be acquired by the measurement circuit, thereby calculating the impedance amplitude and phase angle. After completing measurements at multiple sampling frequencies, a broadband impedance spectrum sequence is formed by each sampling frequency and its corresponding impedance amplitude and phase angle, represented as follows: .

[0095] N f This represents the number of frequency sampling points, for example, 200. The preset frequency range is, for example, 10Hz to 10kHz. The sampling frequencies can be evenly arranged on the logarithmic frequency axis. n For the nth sampling frequency, and These represent the impedance amplitude and phase angle at the corresponding frequencies. This broadband impedance spectrum sequence X constitutes the most basic electrical measurement input for the model.

[0096] Step S103: Encode the broadband impedance spectrum sequence in the frequency domain using a physical constraint graph neural network to obtain global frequency domain features and map them to the initial node features of each node.

[0097] After obtaining the component coupling correlation graph, the next key step is to map the broadband impedance spectrum sequence into graph node features and extract a deep representation that fuses local fault responses and global coupling relationships on the graph structure through message passing. To this end, the physical constraint graph neural network in this application is configured with the component coupling correlation graph of the current transformer; that is, this application uses a physical constraint graph neural network (PC-GNN). This allows for the explicit embedding of multi-physics conservation law constraints on top of standard graph attention networks, ensuring that the feature learning process of the neural network aligns with the internal physical laws of the CT.

[0098] Frequency domain coding is used to extract the changes in impedance amplitude and phase angle along the frequency dimension. The resulting global frequency domain features reflect the frequency domain state of the electrical measurement input of the current transformer under test. Furthermore, this application considers that the impedance spectrum is essentially a measurement of the overall electrical port behavior of the CT, which is a global feature. However, the input of the graph neural network requires component-level node features. In other words, the global frequency domain features are the result of the combined action of various internal components of the current transformer. Therefore, the scheme of this application projects the global frequency domain features according to the projection parameters corresponding to each node, so that different nodes can obtain initial node features with different emphases. That is, differentiated initial features are obtained through node-specific projection.

[0099] In one specific embodiment of the present invention, step S103 may specifically include:

[0100] According to the order of each sampling frequency, the corresponding impedance amplitude and phase angle are arranged to obtain the impedance amplitude sequence and phase angle sequence;

[0101] By using a frequency domain encoder in a physical constraint graph neural network, the impedance magnitude sequence and phase angle sequence are jointly encoded to obtain global frequency domain features.

[0102] The global frequency domain features are linearly projected through the node projection layer corresponding to each node to obtain the initial node features of each node.

[0103] For example, the impedance values ​​corresponding to each sampling frequency can be arranged in ascending or descending order of sampling frequency to obtain an impedance value sequence. Similarly, the phase angles can be arranged in the same sampling frequency order to obtain a phase angle sequence.

[0104] A physically constrained graph neural network can include a frequency domain encoder as a front-end module and a graph attention layer that performs message passing between nodes. The frequency domain encoder can be a one-dimensional convolutional neural network or other encoding structures capable of extracting features along the frequency dimension. Taking a one-dimensional convolutional neural network (1D-CNN) as an example, the impedance amplitude sequence and phase angle sequence are concatenated and input into the frequency domain encoder. For example, they can be jointly encoded using multiple layers of one-dimensional convolution, nonlinear activation, and pooling to obtain the global frequency domain feature z, represented as:

[0105] .

[0106] D enc For the encoding dimension, the i-th node in the component coupling graph can be set as the node projection layer corresponding to that node, through... Obtain the initial node features of the i-th node. .in, and ...

[0107] As can be seen, this projection mechanism allows all nodes to share the same global frequency domain understanding while retaining their unique physical response priors, laying the foundation for differentiated message passing on the graph later. In other words, each node receives the same global frequency domain feature z, but can form initial node features that respectively emphasize the responses of windings, cores, insulation, or terminals.

[0108] Step S104: Based on the component coupling association graph and the coupling weights of each edge, perform multi-layer graph attention propagation on the initial node features of each node.

[0109] In each layer of graph attention propagation, attention weights between connected nodes are determined based on the current node features of the connected nodes and the coupling weights of the corresponding edges, and the node features of each node are updated based on these attention weights. The attention weights obtained in this application do not replace the preset physical coupling weights, but rather, based on the propagation path and coupling strength defined by the physical coupling weights, further reflect the contribution of different component features in the current sample to the state update of the target component. After multiple layers of propagation, the abnormal information carried by a component can propagate layer by layer through its physically coupled components.

[0110] In one specific embodiment of the present invention, during attention propagation at each graph layer, the attention weights between connected nodes are determined based on the current node characteristics of the connected nodes and the coupling weights of the corresponding edges, including:

[0111] For a target node, the nodes connected to the target node are considered as its neighbor nodes; where the target node can be any node.

[0112] Perform linear transformations on the current node features of the target node and each neighboring node respectively;

[0113] For any neighboring node, the node features of the linearly transformed target node, the node features of the linearly transformed neighboring nodes, and the coupling weights of the edges between the target node and the neighboring nodes are concatenated into vectors, and the concatenation result is nonlinearly mapped to obtain the unnormalized attention coefficients between the target node and the neighboring nodes.

[0114] Normalize the unnormalized attention coefficients between the target node and each neighboring node to obtain the attention weights between the target node and each neighboring node.

[0115] Let's take node i as the target node as an example. Let N be the set of neighboring nodes that have an edge with node i.i In the graph attention propagation at each layer (taking the l-th layer as an example), the current node features of the target node entering that layer are denoted as... That is, the feature vector of node i when it enters the attention operation of the l-th layer graph, after attention scoring, message aggregation and activation in the l-th layer, outputs the node features of the (l+1)-th layer. The current node characteristic of any neighboring node j of node i is denoted as... Therefore, it can be understood that in the first layer of graph attention propagation, the current node feature is the initial node feature of the corresponding node, and in subsequent graph attention propagation layers, the current node feature is the updated node feature output from the previous layer.

[0116] Taking the l-th layer as an example, it can be learned through the linear transformation matrix of this layer. , respectively and Perform a linear transformation. Then, analyze the target node features after the linear transformation. Features of neighboring nodes after linear transformation and the coupling weight w of the edge between node i and node j ij Perform vector concatenation and apply the attention parameters of this layer. Mapping with nonlinear functions yields unnormalized attention coefficients. For example, in one specific implementation, the unnormalized attention coefficient can be expressed as: In this example, the symbol ∥ denotes vector concatenation, and LeakyReLU is the linear unit with leakage correction used in this example. It is understandable that each graph attention propagation layer can set its own linear transformation matrix and attention parameters.

[0117] For target node i, the unnormalized attention coefficients between it and each neighboring node can be normalized using Softmax to obtain the attention weights between target node i and neighboring node j. , represented as At this point, the sum of the attention weights of all neighboring nodes for the same target node is 1. This is different from the conventional method of calculating unnormalized attention coefficients. In this implementation, the coupling weight w ij As a physical prior, it participates in the calculation of the unnormalized attention coefficient, forming a unique physical perception attention formula. That is, by explicitly embedding the attention mechanism through physical coupling weights, it can guide the message to propagate along a strongly physical coupling path, thereby improving the monitoring reliability of the proposed solution.

[0118] In one specific embodiment of the present invention, updating the node features of each node based on the attention weights during attention propagation at each layer of the graph can specifically include:

[0119] Based on the attention weights between the target node and each neighboring node, the node features of each neighboring node after linear transformation are weighted and aggregated to obtain the neighborhood feature aggregation result.

[0120] The self-connection weight matrix of this layer is used to perform a self-connection transformation on the current node features of the target node, and the self-connection transformation result is obtained.

[0121] The self-connection transformation result is fused with the neighborhood feature aggregation result and activated by an activation function to obtain the updated node features of the target node.

[0122] After obtaining the attention weights between the target node and each neighboring node, the features of the linearly transformed neighboring nodes can be weighted and summed according to each attention weight to obtain the aggregated neighborhood features of the target node. The neighborhood feature aggregation results reflect the combined impact of the current state of each connected component on the target component through the corresponding propagation path.

[0123] It should be noted that the physical coupling weight w ij It doesn't replace the attention coefficient, but rather serves as an additional prior input in the attention scoring. This allows the model to make reasonable inferences based on physical priors when data is insufficient, and to fine-tune the coupling strength through the attention mechanism when data is abundant, achieving a flexible fusion of physical knowledge and data-driven approaches. Specifically, to preserve the target node's own state information, a self-connection weight matrix is ​​used in this layer. Current node features of the target node Perform a self-connection transformation to obtain The self-connection transformation result is added to or concatenated with the neighborhood feature aggregation result and then fused. Activation is then performed using an activation function such as ELU to obtain the updated node features. For example, in a specific case, the node feature update process for node i can be represented as:

[0124] By performing the above processing on each node, one layer of graph attention propagation is completed. By using the node features output from this layer as the input to the next layer and repeating the above processing, multi-layer graph attention propagation can be achieved.

[0125] Step S105: Aggregate the node features of each node after completing the multi-layer graph attention propagation to obtain global depth features.

[0126] When aggregating node features, methods such as average aggregation and weighted average aggregation can be used. For example, if average aggregation is used, after L layers of propagation, the node features of each node in the last layer can be averaged to obtain the global depth features, represented as: .

[0127] The global depth features include both frequency domain information of the broadband impedance spectrum and component-level correlation information formed after the frequency domain information is propagated on the component coupling correlation diagram. In the specific implementation of this application, not only can fault classification be achieved based on this, but also measurement error can be predicted.

[0128] Step S106: Based on global deep features, obtain the fault mode classification probability distribution for the current transformer under test, and determine the current operating status of the current transformer under test according to the fault mode classification probability distribution.

[0129] In practical applications, global deep features can be input into the classification layer. After linear transformation and Softmax processing, the probability corresponding to each preset running state can be obtained. Then, the current running state can be output according to the probability magnitude and the preset confidence judgment rules.

[0130] In one specific embodiment of the present invention, the fault mode classification probability distribution includes probabilities corresponding to winding deformation, insulation degradation, inter-turn short circuit, abnormal secondary circuit contact resistance, overload, and normal state, respectively.

[0131] The current operating status of the current transformer under test is determined based on the probability distribution of fault mode classification, including:

[0132] The maximum probability in the failure mode classification probability distribution is used as the classification confidence level;

[0133] When the classification confidence level is lower than the confidence level threshold, the current operating state of the current transformer under test is determined to be an unknown abnormal state, and the broadband impedance spectrum sequence is stored in the buffer to be reviewed.

[0134] When the classification confidence level is not lower than the confidence level threshold, the state corresponding to the highest probability is determined as the current operating state of the current transformer under test.

[0135] Specifically, by inputting the global deep features H into the classification layer, a classification score can be obtained. For example, Softmax can be used to obtain the failure mode classification probability distribution ŷ, which is represented as:

[0136] This probability distribution includes the probabilities of winding deformation, insulation degradation, inter-turn short circuit, abnormal secondary circuit contact resistance, overload, and normal conditions. It should be noted that during training, the fault mode label y of the samples and the predicted probability can be used. Determine the fault classification loss in the form of cross-entropy. Among them One-hot encoding of the true label. During online inference, the maximum probability C = max( ) in the failure mode classification probability distribution can be used. ) is used as the classification confidence score, and argmax( Determine the state corresponding to the highest probability. If the classification confidence level C is not lower than the confidence threshold, then the state corresponding to the highest probability is determined as the current operating state of the current transformer under test. The confidence threshold can be set according to actual needs, for example, 0.85.

[0137] Furthermore, in this implementation, if the classification confidence level C is lower than the confidence threshold, it indicates that the matching degree between the current broadband impedance spectrum and the known operating state during the training phase is insufficient. Therefore, the current operating state of the current transformer under test is determined as an unknown abnormal state, and the current broadband impedance spectrum sequence and its associated information such as acquisition conditions and acquisition time can be stored in the audit buffer. The audit samples in the audit buffer may not be directly written into the confirmed state feature library to prevent dirty data caused by abnormal noise or acquisition failure from polluting the feature library. Of course, after the true state is determined through manual inspection or supplementary experiments, it can be used as a new measured training sample for incremental updates of the feature library.

[0138] In one specific embodiment of the present invention, it may further include:

[0139] The global depth features are processed to obtain the predicted value of the metering error of the current transformer under test; the predicted value of the metering error includes the predicted value of the ratio difference and the predicted value of the angle difference.

[0140] Metering error is a core evaluation indicator in the field of power metering. In practical applications, it is usually required to build an error state feature database, which not only needs to record what kind of fault it is, but also needs to answer how large the error is. However, in traditional solutions, fault identification and metering error assessment are separated, making it impossible to simultaneously achieve fault location and error quantification, and making it difficult to support the full life cycle status management of equipment. Therefore, the embodiment of this application also performs metering error prediction at the same time.

[0141] The physical constraint graph neural network in this embodiment includes a fault classification branch and a measurement error regression branch. It is particularly important to emphasize that the fault classification branch and the measurement error regression branch in this application share the global depth feature H output by the frequency domain encoder, node projection layer, each graph attention propagation layer and node feature aggregation layer. This multi-task sharing mechanism forces the depth feature H to simultaneously meet the discriminative requirements of intra-class compactness and inter-class separation, as well as the regression requirement of continuous mapping to measurement error. Therefore, it avoids the feature redundancy problem common in single classification tasks, so that each dimension can carry as much effective information related to the electrical state of the CT as possible.

[0142] In practical applications, the measurement error regression branch can be implemented by sequentially connecting a first linear layer, a nonlinear activation layer, and a second linear layer. For example, in a specific case, a multilayer perceptron (MLP) is constructed based on global deep features H to achieve end-to-end mapping of measurement errors.

[0143] .

[0144] in, σ represents the predicted measurement error, including the predicted ratio difference and the predicted angle difference, and σ is a nonlinear activation function. and These are the linear transformation weight matrices for the first layer and the second layer of the regression network, respectively. For the hidden layer dimension.

[0145] The ratio difference reflects the amplitude error of the secondary output current of the CT, calculated according to the rated transformation ratio and applied to the primary side, relative to the actual primary current. The phase difference reflects the phase error of the calculated secondary current of the CT relative to the primary current. During online inference, the predicted values ​​of the ratio difference and phase difference can be compared with the corresponding metering error alarm thresholds. If either predicted value exceeds the corresponding alarm threshold, an abnormal metering error message can be output.

[0146] In one specific embodiment of the present invention, the determination method for measurement error regression loss includes:

[0147] The training samples are input into the physical constraint graph neural network to obtain the ratio difference prediction value and angle difference prediction value corresponding to the training samples;

[0148] The ratio difference loss is determined based on the difference between the predicted ratio difference and the corresponding actual ratio difference, and the angle difference loss is determined based on the difference between the predicted angle difference and the corresponding actual angle difference.

[0149] By weighted summing of the difference loss and the angle difference loss, the regression loss of measurement error is obtained.

[0150] In this implementation, the predicted ratio difference is obtained through a regression branch of measurement error. Sum of angle difference predictions Based on the predicted and actual values ​​of the ratio difference. The difference between the predicted and actual angle difference values ​​determines the ratio loss. The difference between them determines the angle difference loss. For example, the ratio difference loss can be... Angle difference loss can be .

[0151] The measurement error regression loss is obtained by weighted summation of the difference loss and the angle difference loss. When performing scale normalization, both weights can be set to 1. It equals the sum of the squared prediction error of the ratio difference and the squared prediction error of the angle difference. For example, in one specific implementation, the mean square error is used for the regression loss of the measurement error: .

[0152] In one specific embodiment of the present invention, the training phase of the physical constraint graph neural network includes a first training phase and a second training phase.

[0153] The first loss function used in the first training phase is a weighted sum of fault classification loss and physical constraint loss; the second loss function used in the second training phase is a weighted sum of fault classification loss, measurement error regression loss, contrast loss, and physical constraint loss.

[0154] In this implementation, the physical constraint graph neural network is trained through a first training phase and a second training phase. For example, the training data may include a simulation dataset D generated from a multiphysics digital twin model. sim And the measured dataset D collected from field tests or actual operation. real Each training sample must include at least a broadband impedance spectrum sequence and a fault mode label. In addition, it is understood that the samples used for measurement error regression must also include the true values ​​of the ratio difference and the angle difference.

[0155] For example, in one specific implementation, the first training phase may use only the simulation dataset D. sim This allows the model to first learn basic features with explicit physical field correspondences. The first loss function used in the first training phase is a weighted sum of the fault classification loss and the physical constraint loss, for example, L... stage1 = , where λ phy These are the weights corresponding to the physical constraint loss. By limiting the propagation process of the model on the component coupling graph through physical constraint loss, we can reduce the representation of features that are significantly inconsistent with the internal working mechanism of the current transformer when the number of simulation samples is limited.

[0156] For example, in one specific implementation, the second training phase uses a simulation dataset D. sim and the measured dataset D real The hybrid dataset is used to jointly optimize the model obtained in the first training phase. The second loss function used in the second training phase is a weighted sum of fault classification loss, measurement error regression loss, contrastive loss, and physical constraint loss, for example, L... stage2 = . λ phy λ represents the weights corresponding to the physical constraint loss. cl To compare the weights corresponding to the loss, λreg The weights corresponding to the regression loss of measurement error can be set and adjusted according to actual needs. The gradients generated by each loss jointly update the parameters of the frequency domain encoder, node projection layer and graph attention propagation layer.

[0157] For example, the Adam optimizer can be used in the first training phase, with a learning rate of The batch size is 16, and the training run is 100 rounds. In the second training phase, the learning rate can be adjusted to... The batch size is 32, and the training run is 50 rounds. In practical applications, the second training phase can perform an early stopping operation based on the fault classification accuracy. For example, if the accuracy does not improve after 10 consecutive rounds, training can be stopped, and the model parameters can be restored to the optimal performance on the validation set.

[0158] In one specific embodiment of the present invention, the physical constraint loss includes physical propagation consistency loss and energy dissipation consistency loss;

[0159] Methods for determining the consistency loss in physical propagation include:

[0160] Construct a normalized coupling weight matrix based on the coupling weights of each side;

[0161] In the attention propagation of each layer, the input node features of each node in this layer are weighted and aggregated according to the normalized coupling weight matrix to obtain the physical propagation features of each node.

[0162] Based on the difference between the output node characteristics of each node in this layer and the corresponding physical propagation characteristics, the propagation consistency loss corresponding to this layer is determined, and the propagation consistency losses corresponding to each layer are aggregated to obtain the physical propagation consistency loss.

[0163] Methods for determining energy dissipation consistency loss include:

[0164] The impedance real part characterization quantity is determined based on the real part of the impedance in the broadband impedance spectrum sequence of the training samples, and the node characteristic characterization quantity is determined based on the node characteristics of the nodes that generate active power loss.

[0165] The energy dissipation consistency loss is determined based on the difference between the real part of the impedance and the nodal characteristic.

[0166] While graph attention networks can adaptively learn information transfer patterns between nodes, there is a risk of deviating from physical conservation laws if their optimization process is entirely driven by data labels. For example, the model might learn a numerically valid but physically unreasonable feature representation in order to reduce classification loss. In this application, to ensure that the latent space of the neural network is compatible with the energy propagation and diffusion laws within the CT, two types of physical loss terms can be introduced, embedding multiphysics mechanisms into the network optimization process.

[0167] Specifically, physical constraint loss Including physical propagation consistency loss Consistency loss with energy dissipation , and can be represented as , where λ prop and λ energy These are the weights for the corresponding losses. By... By incorporating the overall optimization objective, the feature learning process of PC-GNN no longer purely pursues the improvement of classification or regression accuracy, but seeks Pareto optimality between accuracy and physical consistency, thereby significantly enhancing the model's generalization ability and interpretability across different working conditions and equipment scenarios.

[0168] When determining the consistency loss of physical propagation, the degree matrix D can be constructed based on the coupling weight matrix W. This matrix characterizes the conservation and propagation properties of physical quantities on a graph structure, thus yielding the normalized coupling weight matrix. We can constrain the feature updates of each layer of nodes to approximately satisfy diffusion conservation, resulting in:

[0169] Its physical meaning is: node. The feature representation in the l-th layer should be consistent with the features of its neighboring nodes after being weighted and aggregated according to the physical coupling strength. This constraint forces the message passing process of the graph neural network to conform to the physical diffusion law, preventing the neural network from deviating too much from the physical conservation law and generating spurious correlation features.

[0170] When determining the energy dissipation uniformity loss, it is considered that the active power loss of the CT core and windings manifests as the cumulative effect of the real part of impedance (resistive component) in the frequency domain, and there is a quantifiable physical relationship between the two. Therefore, a sequence of real parts of impedance can be defined. Thus, energy dissipation loss is constructed, resulting in:

[0171] . The subset of nodes includes the core / winding components that directly generate active power losses (copper losses, iron losses) and whose state changes directly affect the real part of the impedance. Nodes that generate active power losses can include the nodes corresponding to the primary winding, secondary winding, core main column, and core side yoke. The L1 norm is used to aggregate node feature strengths. The normalized scaling factor is a learnable value that is optimized through backpropagation during training.

[0172] The deeper meaning of this loss term is that the depth characteristic strength of the core and winding nodes should reflect the overall energy dissipation level of the equipment. If the features extracted by the model are severely deviated from the real part of the measured impedance, it indicates that the model has failed to effectively capture the essence of electromagnetic-thermal coupling, and physical losses will drive the model to correct the latent space representation.

[0173] In one specific embodiment of the present invention, the method for determining the contrast loss includes:

[0174] Anchor samples, positive samples, and multiple negative samples are selected from the dataset used for the second training phase. Among them, the positive samples and anchor samples have the same fault mode label and come from different load conditions or different current transformers, while each negative sample and anchor sample has a different fault mode label.

[0175] The anchor point samples, positive samples, and each negative sample are respectively input into the physical constraint graph neural network to obtain the corresponding global depth features;

[0176] Determine the first similarity between the global depth features corresponding to the anchor point sample and the global depth features corresponding to the positive sample, and the second similarity between the global depth features corresponding to the anchor point sample and the global depth features corresponding to each negative sample;

[0177] The contrast loss is determined based on the first similarity and each second similarity, so as to increase the feature similarity between samples with the same failure mode label and decrease the feature similarity between samples with different failure mode labels.

[0178] The deep feature vectors extracted by PC-GNN have already incorporated frequency domain response, component topology, and physical conservation information. However, if conventional supervised classification and regression training is used directly, the model performance will heavily depend on the number and distribution balance of labeled samples. In CT fault diagnosis scenarios, there are a large number of normal state samples, while severe fault samples such as inter-turn short circuits and winding deformations are scarce. Traditional cross-entropy loss can easily lead to overfitting of the model to the majority class. Furthermore, supervised learning usually treats different fault categories as mutually exclusive discrete labels, making it difficult to explicitly model the compactness within the same fault category and the separation between different categories, which are precisely what a high-quality feature library is necessary for.

[0179] To overcome the aforementioned limitations, this application introduces a contrastive learning strategy, which explicitly optimizes the metric structure of the feature space through a self-supervised sample pair discrimination task. Simultaneously, considering the typical requirement for accurate quantification of the impact of faults on measurement errors, this application can embed a measurement error regression branch in parallel within the contrastive learning framework, achieving joint optimization of fault identification and error quantification.

[0180] For example, in one specific implementation, any anchor sample can be included in the hybrid dataset used in the second training phase. Select a positive sample and M negative samples Positive samples and anchor samples have the same failure mode labels. However, due to different load conditions or different current transformers, each negative sample and anchor point sample has different fault mode labels. Anchor point samples, positive samples, and each negative sample are all input into a physical constraint graph neural network to obtain the global depth features Ha of the anchor point and the global depth features of the positive samples. and global depth features of each negative sample The samples mentioned above do not need to be obtained in the same physical measurement; pairing between samples can be done in the training batch.

[0181] Such a sample pair construction strategy fully utilizes the multi-condition characteristics of simulation data: although the impedance spectrum details of the same fault differ under different loads, their physical essence is the same, and therefore they should be mapped to adjacent feature space regions. By forcing the model to identify such synonymous but different form sample relationships, the robustness of the model to load fluctuations is significantly improved.

[0182] In one specific implementation, normalized temperature-scale cross-entropy (NT-Xent) can be used as the contrastive loss. By maximizing the feature similarity of similar samples and minimizing the feature similarity of dissimilar samples, the discriminative power of the feature space is enhanced, as expressed as:

[0183] .

[0184] The cosine similarity function; Here, is a temperature coefficient used to adjust the level of attention given to difficult negative samples, and M is the number of negative samples. From an information theory perspective, This is equivalent to maximizing the lower bound of the mutual information between anchor features and positive samples. When the model training converges, similar fault samples in the feature space will form compact clusters, and clear boundaries will be maintained between different sample clusters. This provides a good geometric basis for subsequent feature library retrieval and incremental updates based on feature similarity.

[0185] By applying the technical solution provided in the embodiments of the present invention, a graph neural network based on physical constraints is set up, which embeds physical coupling weights into the graph attention mechanism to guide message passing, greatly improving the physical interpretability of the model and making it less likely to produce spurious correlation prediction results that violate physical laws, thus ensuring the monitoring reliability of the current transformer.

[0186] Specifically, the physical constraint graph neural network of this application is configured with a component coupling association graph of the current transformer. The component coupling association graph uses the components of the current transformer as nodes and the physical coupling relationships between different components as edges. The coupling weight of each edge reflects the degree of coupling between the corresponding components. Therefore, in this application, the compositional relationships of the internal components of the current transformer, the propagation paths of fault effects, and the coupling strength corresponding to the paths can be introduced into the monitoring process as stable physical priors, rather than establishing data associations without physical meaning based solely on port measurement data. During monitoring, the broadband impedance spectrum sequence of the current transformer under test can be obtained. Then, the broadband impedance spectrum sequence is frequency-domain encoded using the physical constraint graph neural network to obtain global frequency domain features, which are then mapped to the initial node features of each node, thus converting the broadband impedance spectrum sequence into a feature representation. Afterwards, multi-layer graph attention propagation is performed on the initial node features of each node based on the component coupling association graph and the coupling weights of each edge. In this application's scheme, during attention propagation at each layer of the graph, the attention weights between connected nodes are determined based on the current node features of the connected nodes and the coupling weights of the corresponding edges. The node features of each node are then updated based on these attention weights. This setup ensures that feature propagation is jointly regulated by the component state presented by the current sample and the coupling between components. This allows fault information to be transmitted layer by layer along paths with actual physical coupling relationships, thus highlighting fault-related propagation paths and suppressing weakly correlated interference paths. Subsequently, the node features of each node after multi-layer graph attention propagation are aggregated to obtain global depth features. Based on these global depth features, the fault mode classification probability distribution for the current transformer under test can be obtained, and the current operating state of the current transformer under test can be determined according to the fault mode classification probability distribution.

[0187] In summary, this solution can effectively monitor current transformers, identify faults, and improve the accuracy and physical interpretability of current transformer status monitoring. It is less likely to produce spurious correlation prediction results that violate physical laws, thus ensuring the reliability of current transformer monitoring.

[0188] To clarify the data interface of the model in engineering deployment and to standardize the storage format of the error state feature library, the complete input and output of the model are defined below.

[0189] Table 2: Model Input

[0190]

[0191] As described above, in a real-world online monitoring scenario, X can be automatically measured and uploaded by the front-end data acquisition module at fixed intervals (e.g., once per hour). and The system can be configured once during system initialization based on the CT model, and remains unchanged during operation.

[0192] Table 3: Intermediate Variables

[0193]

[0194] Table 4: Final Output of the Model

[0195]

[0196] In practical applications, each record in the error state feature library can be stored in the form of a 5-tuple: The quintuple consists of the fault mode label, deep feature vector, predicted measurement error, physical consistency check code (a concatenation of all node features used for physical validity verification during incremental updates), and the model's classification confidence for that record. This quintuple structure balances retrieval efficiency and physical interpretability. H supports fast nearest neighbor retrieval based on cosine similarity for identifying historically similar faults; the physical consistency check code... This preserves all node features. When a new sample is added to the database, its node feature distribution can be verified to ensure physical consistency with existing records, preventing dirty data caused by abnormal noise or acquisition failures from contaminating the feature database.

[0197] Corresponding to the embodiments of the current transformer monitoring method above, this embodiment of the invention also provides a current transformer monitoring device, which may include: a memory 201 for storing a computer program; and a processor 202 for executing the computer program to implement the steps of the current transformer monitoring method in any of the above embodiments.

[0198] Corresponding to the above method embodiments, this invention also provides a computer-readable storage medium and a computer program product, which can be referred to in conjunction with the above.

[0199] The computer program product includes a computer program / instructions that, when executed by a processor, implement the steps of the current transformer monitoring method as described in any of the above embodiments.

[0200] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the current transformer monitoring method as described in any of the above embodiments. The computer-readable storage medium referred to herein includes RAM (Random Access Memory), main memory, ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read Only Memory), registers, hard disks, removable disks, or any other form of storage medium known in the art.

[0201] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0202] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0203] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the technical solutions and core ideas of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the present invention.

Claims

1. A monitoring method for a current transformer, characterized in that, include: Obtain a trained physical constraint graph neural network; wherein the physical constraint graph neural network is configured with a component coupling association graph of the current transformer, the component coupling association graph uses the components of the current transformer as nodes, the physical coupling relationship between different components as edges, and the coupling weight of each edge is used to reflect the degree of coupling between the corresponding components. Obtain the broadband impedance spectrum sequence of the current transformer under test; The broadband impedance spectrum sequence is frequency-domain encoded by the physical constraint graph neural network to obtain global frequency domain features and map them to the initial node features of each node. Based on the component coupling association graph and the coupling weight of each edge, multi-layer graph attention propagation is performed on the initial node features of each node; wherein, in each layer of graph attention propagation, the attention weight between connected nodes is determined based on the current node features of connected nodes and the coupling weight of the corresponding edge, and the node features of each node are updated based on the attention weight. The node features of each node after the multi-layer graph attention propagation is completed are aggregated to obtain global depth features; Based on the global depth features, a fault mode classification probability distribution for the current transformer under test is obtained, and the current operating state of the current transformer under test is determined according to the fault mode classification probability distribution.

2. The monitoring method for a current transformer according to claim 1, characterized in that, Obtain the broadband impedance spectrum sequence of the current transformer under test, including: When the current transformer under test is in a energized operating state, a wideband scanning signal with a frequency varying within a preset frequency range is injected into the current transformer under test through an excitation source. The impedance amplitude and phase angle of the current transformer under test are obtained at multiple sampling frequencies, and the broadband impedance spectrum sequence is formed by each sampling frequency and the corresponding impedance amplitude and phase angle.

3. The monitoring method for a current transformer according to claim 2, characterized in that, The broadband impedance spectrum sequence is frequency-domain encoded using the physical constraint graph neural network to obtain global frequency domain features, which are then mapped to the initial node features of each node, including: According to the order of the sampling frequencies, the corresponding impedance amplitudes and phase angles are arranged to obtain the impedance amplitude sequence and the phase angle sequence; The impedance magnitude sequence and the phase angle sequence are jointly encoded by the frequency domain encoder in the physical constraint graph neural network to obtain global frequency domain features. The global frequency domain features are linearly projected through the node projection layer corresponding to each node to obtain the initial node features of each node.

4. The monitoring method for a current transformer according to claim 1, characterized in that, The components of the current transformer include a primary winding, a secondary winding, a core column, a core yoke, a primary insulation layer, a secondary insulation layer, and secondary circuit terminals. In the component coupling association diagram, the coupling weight of each edge is the electromagnetic coupling weight, heat conduction weight and mechanical coupling weight between the corresponding components obtained through multiphysics simulation, and the coupling weight is obtained by weighted summation according to the corresponding physical field contribution coefficients. The electromagnetic coupling weight reflects the mutual inductance coefficient between the corresponding components, the thermal conduction weight reflects the thermal resistance between the corresponding components, and the mechanical coupling weight reflects the structural stiffness coefficient between the corresponding components.

5. The monitoring method for a current transformer according to claim 1, characterized in that, In the attention propagation of each graph layer, the attention weights between connected nodes are determined based on the current node features of the connected nodes and the coupling weights of the corresponding edges, including: For a target node, the nodes connected to the target node are considered as the neighbor nodes of the target node; wherein, the target node can be any node; A linear transformation is performed on the current node features of the target node and each of the neighboring nodes, respectively; For any of the neighboring nodes, the node features of the target node after linear transformation, the node features of the neighboring nodes after linear transformation, and the coupling weight of the edge between the target node and the neighboring nodes are concatenated into a vector, and the concatenation result is nonlinearly mapped to obtain the unnormalized attention coefficient between the target node and the neighboring nodes. The unnormalized attention coefficients between the target node and each of the neighboring nodes are normalized to obtain the attention weights between the target node and each of the neighboring nodes.

6. The monitoring method for a current transformer according to claim 5, characterized in that, In each layer of graph attention propagation, the node features of each node are updated based on the attention weights, including: Based on the attention weights between the target node and each of the neighboring nodes, the node features of each of the neighboring nodes after linear transformation are weighted and aggregated to obtain the neighborhood feature aggregation result. The self-connection weight matrix of this layer is used to perform a self-connection transformation on the current node features of the target node to obtain the self-connection transformation result. The self-connection transformation result is fused with the neighborhood feature aggregation result and activated by an activation function to obtain the updated node features of the target node.

7. The monitoring method for a current transformer according to claim 1, characterized in that, The fault mode classification probability distribution includes probabilities corresponding to winding deformation, insulation degradation, inter-turn short circuit, abnormal secondary circuit contact resistance, overload, and normal state, respectively. The current operating state of the current transformer under test is determined based on the fault mode classification probability distribution, including: The maximum probability in the failure mode classification probability distribution is used as the classification confidence level; When the classification confidence level is lower than the confidence threshold, the current operating state of the current transformer under test is determined to be an unknown abnormal state, and the broadband impedance spectrum sequence is stored in the buffer to be reviewed. When the classification confidence level is not lower than the confidence level threshold, the state corresponding to the highest probability is determined as the current operating state of the current transformer under test.

8. The monitoring method for a current transformer according to any one of claims 1 to 7, characterized in that, Also includes: The global depth features are processed to obtain the metering error prediction value of the current transformer under test; wherein the metering error prediction value includes the ratio difference prediction value and the angle difference prediction value.

9. The monitoring method for a current transformer according to claim 8, characterized in that, The training phase of the physical constraint graph neural network includes a first training phase and a second training phase. The first loss function used in the first training phase is a weighted sum of fault classification loss and physical constraint loss; the second loss function used in the second training phase is a weighted sum of fault classification loss, measurement error regression loss, contrast loss, and physical constraint loss.

10. The monitoring method for a current transformer according to claim 9, characterized in that, The physical constraint loss includes physical propagation consistency loss and energy dissipation consistency loss; The methods for determining the physical propagation consistency loss include: Construct a normalized coupling weight matrix based on the coupling weights of each edge; In the attention propagation of each layer, the input node features of each node in this layer are weighted and aggregated according to the normalized coupling weight matrix to obtain the physical propagation features of each node. Based on the difference between the output node characteristics of each node at this layer and the corresponding physical propagation characteristics, the propagation consistency loss corresponding to this layer is determined, and the propagation consistency losses corresponding to each layer are aggregated to obtain the physical propagation consistency loss. The methods for determining the consistency loss of energy dissipation include: The impedance real part characterization quantity is determined based on the impedance real part in the broadband impedance spectrum sequence of the training sample, and the node characteristic characterization quantity is determined based on the node characteristics of the node that generates active power loss. The energy dissipation consistency loss is determined based on the difference between the real part of the impedance and the node characteristic.

11. The monitoring method for a current transformer according to claim 9, characterized in that, The methods for determining the contrast loss include: Anchor samples, positive samples, and multiple negative samples are selected from the dataset used for the second training phase; wherein the positive samples and the anchor samples have the same fault mode label and come from different load conditions or different current transformers, and each negative sample and the anchor sample have different fault mode labels. The anchor point samples, the positive samples, and each of the negative samples are respectively input into the physical constraint graph neural network to obtain the corresponding global depth features; Determine the first similarity between the global depth feature corresponding to the anchor point sample and the global depth feature corresponding to the positive sample, and the second similarity between the global depth feature corresponding to the anchor point sample and the global depth feature corresponding to each of the negative samples; The contrast loss is determined based on the first similarity and each of the second similarities, so as to increase the feature similarity between samples with the same fault mode label and decrease the feature similarity between samples with different fault mode labels.

12. The monitoring method for a current transformer according to claim 9, characterized in that, The determination methods for the regression loss of the measurement error include: The training samples are input into the physical constraint graph neural network to obtain the ratio difference prediction value and angle difference prediction value corresponding to the training samples; The ratio difference loss is determined based on the difference between the predicted ratio difference value and the corresponding true ratio difference value, and the angle difference loss is determined based on the difference between the predicted angle difference value and the corresponding true angle difference value. The measurement error regression loss is obtained by weighted summation of the ratio difference loss and the angle difference loss.

13. A monitoring device for a current transformer, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the monitoring method for a current transformer as described in any one of claims 1 to 12.

Images