A method for locating faults of an ultra-high voltage converter based on QGNNs and PH and related equipment

Abstract

The application discloses a kind of based on QGNNs and PH's extra-high voltage converter transformer fault location method and related equipment, it is related to power equipment fault diagnosis technical field. Including the acquisition of converter transformer multi-source heterogeneous monitoring data, and according to its physical structure graph model is constructed, sensor measuring point or monitored component is mapped as node;The sustained coherence analysis of the multidimensional time series data of each node is carried out, and topological invariant feature is extracted, and original node feature is fused to form enhanced node feature to construct enhanced graph model;Quantum-classical hybrid graph neural network is constructed and trained on the enhanced graph model, extracts multi-scale feature, outputs the fault type and corresponding fault location of converter transformer. It can be stable in strong noise and complex working condition Characterization of the structural characteristics of multi-source signal, make full use of the physical topological relationship of converter transformer internal components, realize the high-precision identification and component-level positioning of composite fault and early latent fault, improve the accuracy and reliability of fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of power equipment diagnosis, specifically to a fault location method and related equipment for ultra-high voltage converter transformers based on QGNNs and PH. Background Technology

[0002] Ultra-high voltage direct current (UHVDC) transmission is a core technology of my country's "West-to-East Power Transmission" strategy. The converter transformer, as one of the most critical and expensive large-scale pieces of equipment in a converter station, directly determines the safety and stability of the entire transmission system. Because UHVDC converter transformers are subjected to multiple stresses such as combined AC and DC electric fields, high temperatures, and mechanical vibrations over long periods, they are highly susceptible to internal faults such as insulation aging, winding deformation, and partial discharge. Without timely warning and location, these faults could lead to catastrophic accidents, causing enormous economic losses and social impact.

[0003] Currently, fault diagnosis for converter transformers mainly relies on the following methods:

[0004] Dissolved gas analysis (DGA): This is the most traditional and widely used method. It analyzes the concentration and ratio of fault-generated gases (such as H2, CH4, C2H2, etc.) in transformer oil and determines the fault type based on empirical criteria such as the IEC three-ratio method and the Rogers ratio method. However, this method heavily relies on expert experience, is insensitive to the identification of complex faults and early latent faults, and suffers from high rates of false positives and false negatives.

[0005] Vibration analysis: This method diagnoses the mechanical condition of the core and windings by analyzing vibration signals on the surface of the converter transformer housing. However, vibration signals attenuate significantly during propagation and are easily affected by complex electromagnetic noise and load fluctuations in the field, resulting in a low signal-to-noise ratio, making feature extraction difficult and limiting positioning accuracy.

[0006] Shallow machine learning-based methods: In recent years, some studies have attempted to apply shallow machine learning models such as support vector machines (SVM) and decision trees to fault diagnosis. While these methods reduce reliance on human subjective experience to some extent, their model capacity is limited, making it difficult to learn deep and complex fault modes from high-dimensional, nonlinear, and strongly coupled monitoring data. They also have insufficient generalization ability and poor diagnostic performance for novel or complex faults.

[0007] (Deep) Neural Network Methods: Deep learning models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs) can automatically learn deeper features. However, most existing methods treat the monitoring data from each sensor as independent time series or images, failing to effectively model the inherent physical connections and spatial topological relationships between the various components within the converter transformer. The generation and development of faults within the converter transformer is a dynamic process that propagates in physical space; ignoring this structural information will result in the loss of crucial diagnostic information.

[0008] Therefore, there is an urgent need in this field for an intelligent fault location method that can deeply integrate equipment spatial topology information, is robust to noise, and has strong deep pattern recognition capabilities. Summary of the Invention

[0009] The technical problem to be solved by this invention is that existing converter transformer fault diagnosis methods generally rely heavily on expert experience and manual features, are sensitive to field noise and operating condition fluctuations, and have difficulty in characterizing the high-dimensional nonlinear coupling relationship of multi-source heterogeneous time series data. The purpose is to provide a fault location method and related equipment for UHV converter transformers based on QGNNs and PH, which solves the problem of how to incorporate the physical topological relationship between internal components of the converter transformer into the modeling to improve the identification ability of composite faults and early latent faults and the fault component-level location accuracy.

[0010] This invention is achieved through the following technical solution:

[0011] A fault location method for UHV converter transformers based on QGNNs and PH includes:

[0012] Multi-source heterogeneous monitoring data of UHV converter transformers are collected, and the monitoring data is constructed into a graph model based on the physical structure of the converter transformers. The nodes of the graph model are used to characterize the sensor placement location or monitored components, the edges of the graph model are used to characterize the physical connection relationship or functional association relationship between the components, and the node features are used to characterize the multi-dimensional time-series monitoring data of the corresponding nodes.

[0013] By performing continuous homology analysis on multidimensional temporal data of at least some nodes in a graph model, topological invariant features characterizing the topological structure of the multidimensional temporal data are obtained.

[0014] The topology-invariant features are fused with the original node features of the corresponding nodes to obtain enhanced node features, and an enhanced graph model is constructed.

[0015] Based on the enhanced graph model, a quantum-classical hybrid graph neural network is constructed and trained to obtain a fault location model;

[0016] The real-time monitoring data of the UHV converter transformer to be diagnosed is processed into an enhanced graph model and input into the fault location model. The predicted results of the fault type and fault location are output as the fault diagnosis and location results of the UHV converter transformer.

[0017] Furthermore, the persistent cohomology analysis includes:

[0018] The multidimensional time-series data is preprocessed and constructed into a high-dimensional point cloud dataset based on phase space reconstruction.

[0019] Based on fault sensitivity weights and setting multi-scale parameters, continuous cohomology calculation is performed on the high-dimensional point cloud dataset to obtain the continuous interval set of each topologically invariant feature.

[0020] The set of persistent intervals is vectorized to obtain the topologically invariant features.

[0021] Furthermore, the preprocessing of the multidimensional time series data includes: performing outlier detection, noise filtering, and data standardization sequentially to obtain standard time series data;

[0022] The construction of a high-dimensional point cloud dataset includes: determining the time delay parameter τ and the embedding dimension m, and reconstructing the phase space of the standard time series data according to the time delay parameter τ and the embedding dimension m to obtain a phase space point cloud; and performing quality evaluation and optimization on the phase space point cloud to obtain a high-dimensional point cloud dataset for continuous cohomology computation.

[0023] Furthermore, based on fault sensitivity weights and setting multi-scale parameters, the method includes: assigning fault sensitivity weights to data points in a high-dimensional point cloud dataset, and determining the weighted distance between any pair of data points based on the fault sensitivity weights.

[0024] The continuous cohomology calculation includes: under the scale sequence corresponding to the multi-scale parameters, a simple complex filter flow is constructed stepwise based on the weighted distance; during the filter flow process, the birth scale and death scale of the topologically invariant features are calculated and recorded; and the birth scale and death scale are combined to form the continuous interval.

[0025] Furthermore, the weighted distance The determination steps include:

[0026] With data points With data points The original European distance Based on, and according to data points Fault sensitivity weight The weighted distance is obtained by weighting the original Euclidean distance. ;

[0027] Wherein, the fault sensitivity weight It is determined by a linear combination of at least the following three types of indicators:

[0028] Data points The energy index of the intrinsic mode function obtained by empirical mode decomposition of the corresponding signal;

[0029] Data points The deviation index of the current frequency of the corresponding signal from the fundamental frequency;

[0030] Data points Correlation coefficient index with historical normal state;

[0031] Weighting coefficients are assigned to the three types of indicators respectively to obtain the fault sensitivity weights. .

[0032] Furthermore, the feature vectorization of the set of sustained intervals includes:

[0033] The set of persistent intervals is filtered to retain persistent intervals whose persistence meets preset conditions;

[0034] The preserved persistent intervals are converted into corresponding persistent landscape functions, and at least one landscape layer is extracted;

[0035] The landscape layer is uniformly sampled within a preset scale range to obtain a discrete numerical sequence;

[0036] By concatenating discrete numerical sequences corresponding to different topological dimensions and different landscape layers, topologically invariant features with fixed dimensions are obtained.

[0037] Furthermore, the quantum-classical hybrid graph neural network comprises, in sequence: a feature encoding layer, a quantum graph convolutional layer, a quantum measurement layer, and a decoding and output layer;

[0038] The feature encoding layer is used to normalize the features of the enhanced nodes and map them to quantum states using amplitude encoding; the quantum graph convolution layer is used to perform quantum domain graph convolution on the quantum states based on the adjacency relationship of the enhanced graph model; the quantum measurement layer is used to perform quantum measurement on the convolved quantum states to obtain classical feature vectors; and the decoding and output layer is used to output the fault type probability and / or fault location probability based on the classical feature vectors.

[0039] Furthermore, the quantum graph convolutional layer achieves adjacency information interaction by applying a two-qubit entanglement gate to the corresponding qubits of adjacent nodes, and achieves node feature transformation by applying a single-qubit parameterized rotation gate to the corresponding qubits of nodes.

[0040] The quantum measurement layer includes: calculating the Pauli operator expectation value for the corresponding qubit in the output quantum state to obtain the node-level classical eigenvector; and summing or averaging the node-level classical eigenvectors to obtain the graph-level classical eigenvectors.

[0041] The decoding and output layer outputs the probability distribution of fault types based on the graph-level classical feature vectors, and outputs the probability distribution of the nodes corresponding to the fault locations based on the node-level classical feature vectors.

[0042] This invention also provides an ultra-high voltage converter transformer fault location system based on QGNNs and PH, used to implement the aforementioned ultra-high voltage converter transformer fault location method based on QGNNs and PH, comprising:

[0043] The graph modeling module is used to collect multi-source heterogeneous monitoring data of UHV converter transformers and construct a graph model of the monitoring data based on the physical structure.

[0044] The continuous homology feature extraction module is used to perform continuous homology analysis on the multidimensional time-series data of nodes in the graph model and output topologically invariant features.

[0045] The enhanced graph model building module is used to fuse topology-invariant features with original node features to obtain enhanced node features and build an enhanced graph model.

[0046] The quantum-classical hybrid graph neural network module is used to train and form a fault location model based on the augmented graph model, and to infer the augmented graph model to be diagnosed and output the predicted results of fault type and fault location.

[0047] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the fault location method for UHV converter transformers based on QGNNs and PH as described above.

[0048] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the UHV converter transformer fault location method based on QGNNs and PH as described above.

[0049] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0050] This invention extracts topologically invariant features from the time-series signals by reconstructing the phase space of multidimensional time-series data from each monitoring node and introducing continuous homology analysis. These features reflect the geometric and connectivity structure of the signal at multiple scales, rather than instantaneous amplitudes or local statistics, thus exhibiting inherent robustness to random noise, sampling disturbances, and load fluctuations. By introducing these topologically invariant features into the converter transformer fault diagnosis process, the risk of misdiagnosis and missed diagnosis caused by measurement errors and changes in operating conditions can be significantly reduced, fundamentally improving the stability and reliability of the features.

[0051] This invention models the electrical connections and mechanical structure of the converter transformer as a graph structure containing nodes and edges, allowing the physical relationships between different components to be explicitly entered into the model in topological form. Therefore, the propagation paths, spatial adjacency relationships, and functional coupling relationships of faults between different components can be directly learned and utilized within the model.

[0052] By fusing the topologically invariant features obtained through continuous cohomology with the original node features to form enhanced node features, and performing reasoning on the graph structure, the model of this invention can simultaneously obtain the global fault state and the local anomaly representation of each node. Thus, while outputting the fault type, it further outputs the probability distribution of the corresponding faulty component or measurement point, achieving precise positioning of specific locations such as windings, cores, and bushings.

[0053] Continuous cohomology characterizes the appearance and disappearance of signal structures at different scales, exhibiting higher sensitivity to minute structural changes. Graph neural networks, on the other hand, can propagate and aggregate local anomalies across the physical topology. Therefore, when multi-point coupling anomalies or latent faults such as early partial discharge or local overheating occur within the converter transformer, which have not yet manifested as significant amplitude changes, this invention can still detect them through changes in the topology and the relationships between nodes, thus achieving earlier and more reliable early warning than traditional methods.

[0054] This invention organically combines multi-source data acquisition, graph structure modeling, topological invariant feature extraction, feature fusion, and quantum graph neural network reasoning to form a complete closed loop. This eliminates the need for manual experience rules or manual feature selection in the fault location process, significantly improving the automation level and engineering feasibility of converter transformer fault diagnosis. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0056] Figure 1This is a flowchart of the UHV converter transformer fault location method based on QGNNs and PH in Example 1.

[0057] Figure 2 This is a schematic diagram of the graph model construction in Example 1.

[0058] Figure 3 This is a schematic diagram of the quantum-classical hybrid graph neural network in Example 1. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0060] Example 1

[0061] A fault location method for UHV converter transformers based on QGNNs and PH, such as Figure 1 As shown, it includes:

[0062] S100. Collect multi-source heterogeneous monitoring data of the UHV converter transformer, and construct the monitoring data into a graph model based on the physical structure of the converter transformer. The nodes of the graph model are used to characterize the sensor placement location or monitored component, the edges of the graph model are used to characterize the physical connection relationship or functional association relationship between the components, and the node features are used to characterize the multi-dimensional time-series monitoring data of the corresponding node.

[0063] S200. Perform continuous homology (PH) analysis on the multidimensional time-series data of at least some nodes in the graph model to obtain topologically invariant features characterizing the topological structure of the multidimensional time-series data.

[0064] S300. The topology-invariant features are fused with the original node features of the corresponding nodes to obtain enhanced node features, and an enhanced graph model is constructed.

[0065] S400. Construct and train a quantum-classical hybrid graph neural network based on the enhanced graph model to obtain a fault location model;

[0066] S500: Process the real-time monitoring data of the UHV converter transformer to be diagnosed into an enhanced graph model, input it into the fault location model, and output the predicted results of fault type and fault location as the fault diagnosis and location results of the UHV converter transformer.

[0067] QGNNs stands for Quantum Graph Neural Network, and PH stands for Continuous Coherence.

[0068] like Figure 2As shown, the left side is a schematic diagram of the physical structure and measurement point arrangement of an ultra-high voltage converter transformer, and the right side is a schematic diagram of a graphical model constructed based on the physical structure. The converter transformer includes winding components and core components, and various types of sensor measurement points are arranged at preset positions on the windings and core, such as high-voltage side measurement points, low-voltage side measurement points, core measurement points, and bushing measurement points. Each measurement point is used to collect multi-source monitoring signals related to fault conditions. The multi-source monitoring signals include at least one or more of the following: vibration signals, temperature signals, electrical quantity signals, and / or oil chromatography-related signals.

[0069] In this embodiment, the physical structure of the converter transformer is abstracted as a graphical model G. The nodes of the graphical model represent sensor measurement points or monitored components, the edges represent the physical connections or functional relationships between components, and the node features represent the multi-dimensional time-series monitoring data of the corresponding nodes. Specifically, for the component and measurement point arrangement shown in the left figure, the A-phase winding measurement point is abstracted as node W_A1, the core measurement point as node C_F1, the B-phase winding measurement point as node W_B1, the A-phase high-voltage side bushing measurement point as node B_HV, and the B-phase low-voltage side bushing measurement point as node B_LV. These nodes are only for illustration; in actual engineering, the number and type of nodes can be expanded according to the measurement point arrangement scheme.

[0070] Furthermore, an edge set is constructed based on the structural connection relationship of the converter transformer: for example, if there is direct structural coupling or fault propagation correlation between the A-phase winding and the core, an edge is established between node W_A1 and node C_F1; if there is structural coupling or fault propagation correlation between the core and the B-phase winding, an edge is established between node C_F1 and node W_B1; if there is an electrical connection or functional relationship between the winding measuring point and the corresponding bushing measuring point, an edge is established between node W_A1 and node B_HV, and between node W_B1 and node B_LV. Edges can be set as undirected edges to represent symmetrical coupling relationships, or as directed edges to represent the fault propagation direction; when using directed edges, the edge direction can be determined according to a preset propagation rule, for example, from the winding to the bushing or from the winding to the core.

[0071] In terms of node feature construction, multidimensional time-series data of each node v within a preset time window are collected to form the original node feature X_v. The preset time window can be a fixed-length sliding window, and the multidimensional time-series data can include different physical quantity channels of the same measurement point, or include multi-scale representations of similar signals.

[0072] In this embodiment, S200 includes the following steps:

[0073] S201. The multidimensional time series data is preprocessed, and a high-dimensional point cloud dataset is constructed from the multidimensional time series data based on phase space reconstruction.

[0074] S202. Based on the fault sensitivity weight and setting multi-scale parameters, perform continuous cohomology calculation on the high-dimensional point cloud dataset to obtain the continuous interval set of each topologically invariant feature.

[0075] S203. The set of continuous intervals is vectorized to obtain the topologically invariant features.

[0076] In this embodiment, S201 includes the following steps:

[0077] S201-1. Perform outlier detection, noise filtering, and data standardization in sequence to obtain standard time series data;

[0078] S201-2. Determine the time delay parameter τ and the embedding dimension m, and reconstruct the phase space of the standard time series data according to the time delay parameter τ and the embedding dimension m to obtain the phase space point cloud;

[0079] S201-3. Perform quality assessment and optimization on the phase space point cloud to obtain a high-dimensional point cloud dataset for continuous cohomology calculation.

[0080] In this embodiment, S202 includes the following steps:

[0081] S202-1. Assign fault sensitivity weights to data points in the high-dimensional point cloud dataset, and determine the weighted distance between any pair of data points based on the fault sensitivity weights.

[0082] S202-2. Under the scale sequence corresponding to the multi-scale parameters, a simple complex filter flow is constructed stepwise based on the weighted distance. During the filter flow process, the birth and death scales of the topologically invariant features are calculated and recorded, and the birth and death scales are combined to form the duration interval. Specifically:

[0083] When the weighted distance between data point pairs is less than the current scale, add an edge between them, and when there are edges connecting all data point pairs that are vertices, add a high-dimensional simplicial complex.

[0084] During the simple complex construction process, the birth and death scales of topological features in each dimension are calculated and tracked, and these are used as the persistence intervals of the corresponding topological features.

[0085] In this embodiment, the weighted distance The determination steps include:

[0086] With data points With data points The original European distance Based on, and according to data points Fault sensitivity weight The weighted distance is obtained by weighting the original Euclidean distance. ;

[0087] Wherein, the fault sensitivity weight It is determined by a linear combination of at least the following three types of indicators:

[0088] Data points The energy index of the intrinsic mode function obtained by empirical mode decomposition of the corresponding signal;

[0089] Data points The deviation index of the current frequency of the corresponding signal from the fundamental frequency;

[0090] Data points Correlation coefficient index with historical normal state;

[0091] Weighting coefficients are assigned to the three types of indicators respectively to obtain the fault sensitivity weights. .

[0092] Then the data points and data points The weighted distance between them is:

[0093]

[0094]

[0095] In the formula, Representing data points and data points The original Euclidean distance between them Indicates the weighted influence factor. Representing data points Fault sensitivity weights, Representing data points Energy of eigenmode functions obtained through empirical mode decomposition Representing data points The degree of deviation between the current frequency and the fundamental frequency. Representing data points Correlation coefficient with historical normal conditions , and They represent , and The weighting coefficients can be obtained by training and optimizing using historical fault data.

[0096] In the above process, the multi-scale parameters are:

[0097]

[0098]

[0099] In the formula, Indicates the range of scale parameters. This represents the mean distance between all data points. This represents the standard deviation of all data points with respect to distance. Indicates the number of scale parameters. Indicates the total number of scale parameters. This represents the sampling density adjustment parameter.

[0100] Specifically, when the weighted distance between data point pairs Less than or equal to the current scale When all vertex pairs are connected by edges, edges are added; furthermore, when all vertex pairs are connected by edges, higher-dimensional simplexes (such as triangles or tetrahedrons) are added. During this filtering process, the "birth" scale b (when the feature appears) and the "death" scale d (when the feature disappears) of each dimensional topological feature (0-dimensional: connected components, 1-dimensional: cycles, 2-dimensional: holes) are calculated and tracked. Each pair... This constitutes a continuous interval, the length of which is... (Persistence) characterizes the stability of the topological feature. Finally, the output is a set of multiple persistence intervals, each interval representing a topological feature that is stable at different scales.

[0101] In this embodiment, S203 includes the following steps:

[0102] S203-1. Filter the set of persistent intervals and retain persistent intervals whose persistence meets the preset conditions;

[0103] S203-2. Convert the retained persistent intervals into the corresponding persistent landscape functions and extract at least one landscape layer;

[0104] S203-3. Uniformly sample the landscape layer within a preset scale range to obtain a discrete numerical sequence;

[0105] S203-4. By splicing together discrete numerical sequences corresponding to different topological dimensions and different landscape layers, topological invariant features with fixed dimensions are obtained.

[0106] Specifically, the calculated persistent intervals are filtered, retaining those with significant persistence (e.g., length greater than 1.5 times the average length of all intervals) to filter out transient topological structures caused by noise. Then, each persistent interval... Transform into a landscape function:

[0107]

[0108] The landscape function reaches its maximum value at the center of the interval and decreases linearly to zero towards both ends. For all salient intervals, the corresponding landscape function is plotted at multiple scales. Take the values ​​from the top, sort them by function value, and take the first one. One (e.g., K=5) is used as the "landscape layer". ,in yes The maximum value of all landscape functions. This is the second largest value, and so on.

[0109] Finally, for each landscape layer Uniform sampling is performed within a preset scale range (from minimum birth value to maximum death value) to obtain discrete numerical sequences. These sequences (from homology groups of different dimensions, such as...) , (And different landscape layers) are pieced together to form a long feature vector. It represents the structural information of point clouds at different scales and topological dimensions, is robust to small deformations and noise in the data, and is a key interface connecting topological data analysis and quantum graph neural networks.

[0110] In S300 of this embodiment, the topologically invariant features with physical meaning extracted from each node are concatenated or weighted and fused with the original node features X to form enhanced node features. This results in enhanced graph structure data. This allows for the combination of local statistical features of the data with global topological features.

[0111] In S400 of this embodiment, as Figure 3 As shown, a quantum-classical hybrid graph neural network model specifically designed for fault diagnosis of UHV converter transformers is constructed and trained to learn the complex mapping relationship between fault modes and locations from enhanced graph structure data; for the enhanced graph model By deeply integrating quantum computing and classical computing, we can achieve accurate classification and location of faults.

[0112] The quantum-classical hybrid graph neural network comprises, in sequence, a feature encoding layer, a quantum graph convolutional layer, a quantum measurement layer, and a decoding and output layer;

[0113] The feature encoding layer is used to normalize the features of the enhanced nodes and map them to quantum states using amplitude encoding; the quantum graph convolution layer is used to perform quantum domain graph convolution on the quantum states based on the adjacency relationship of the enhanced graph model; the quantum measurement layer is used to perform quantum measurement on the convolved quantum states to obtain classical feature vectors; and the decoding and output layer is used to output the fault type probability and / or fault location probability based on the classical feature vectors.

[0114] The feature encoding layer enhances the node features in the graph model. Quantum states encoded as qubits serve as a bridge connecting classical data and quantum computing; specifically, they enhance node characteristics. Normalization is performed to make its norm 1 to meet the quantum state normalization requirements. Then, a classic fully connected neural network is used to compress or map the feature dimensions to... Dimension, where n is the number of qubits. The mapped eigenvectors. satisfy It can be directly used for amplitude encoding.

[0115] The normalized feature vector Each component is encoded as the ground state amplitude. In the superposition of qubits. For a single node Its corresponding quantum state Mapped to:

[0116]

[0117] in, It calculates the ground state.

[0118] The above process encodes the d-dimensional classical feature vector into a representation of the d-dimensional feature vector. A quantum bit spanning In Widhillbert space, by assigning each node in the graph Prepare its quantum state independently The quantum state of the entire graph G' can be represented as the tensor product of the quantum states of all nodes. , which serves as the input to the quantum graph convolutional layer.

[0119] Quantum graph convolutional layers aim to simulate the propagation and aggregation of information along a graph model through quantum circuits. Their design is inspired by classical graph convolutional networks, but is implemented in Hilbert space through quantum gate operations.

[0120] Specifically, to simulate the propagation of node information along edge E in graph convolution, a parameterized two-qubit controlled gate sequence is designed. For each edge in the graph... In the corresponding qubit and Apply a parameterized two-bit unitary gate between them It is represented as:

[0121]

[0122] in, These are trainable parameters. This operation makes the node... and The quantum states become entangled, allowing information to be exchanged. This process is equivalent to encoding the graph adjacency matrix A in a quantum circuit, where... This indicates that there is an interaction.

[0123] During information exchange, the characteristics of each node also need to undergo non-linear transformation. This is done by representing each node... qubit Apply a series of single-bit parameterized rotating gates To achieve this, it is represented as:

[0124]

[0125] In the formula, It is a trainable 3D parameter vector. This operation is analogous to the weight transformation of node features in classic graph convolution.

[0126] By combining the aforementioned "adjacency interaction" and "node feature transformation" operations, a basic "quantum graph convolutional layer" is constructed; based on this, in the entire quantum graph convolutional layer, the first quantum circuit... The graph convolution operation of a layer is represented as:

[0127]

[0128] In the formula, Indicating the first quantum circuit The output quantum state of the layer quantum graph convolution structure. Indicates the first Layer nodes Single-bit quantum gate operations (parameters are) ), Indicates the first Layer connection nodes and Two-qubit quantum gate operation (parameters are) ), It indicates that it is the first Layer nodes The single-qubit transformation parameters, Indicates the first Layer edge The parameters of the two-qubit entanglement gate, Represents nodes in an augmented graph model and nodes The connecting edge, This represents the set of edges in an augmented graph model. This represents the set of nodes in the augmented graph model.

[0129] By stacking multiple layers, the model can simulate multi-hop propagation of information in a graph within a quantum state, thereby capturing more complex global structural information and all parameters. It is learned through optimization algorithms during the training process.

[0130] In the quantum measurement layer, the final quantum state is obtained by processing the quantum graph convolution layer. To use it for classical post-processing, quantum measurements are required to collapse it into classical data as extracted multi-scale feature vectors. This process includes:

[0131] For each qubit in the output quantum state, calculate its Pauli operator expectation value, which is used as the classical eigenvector corresponding to that node;

[0132] The classical feature vectors of all nodes are summed or averaged to obtain the global graph-level feature vector, while retaining the classical feature vector of each node.

[0133] Global graph-level feature vectors and classical feature vectors are used as multi-scale feature vectors for nodes.

[0134] Specifically, Pauli operators are used. The expected value is used as the measurement result. For each node Calculate the Pauli operator expectation value for the corresponding qubit (or qubit group). ,in The expectation value can be approximated by preparing and measuring the output state multiple times on a quantum computer and then statistically averaging the results. These expectation values ​​constitute the classical eigenvectors of each node. .

[0135] Since fault location requires outputting the global state of the entire graph (fault type) and the local state of the nodes (fault location), this step employs a dual readout strategy. First, the feature vectors of all nodes are... Summation or averaging yields a global graph-level feature vector. At the same time, the classic feature vector of each node is preserved. Used for subsequent node-level (localization) tasks. Among them, the global graph-level feature vector... It will be used as the input for subsequent decoding and output layers.

[0136] The decoding and output layer consists of a multilayer sensing mechanism, whose input is a global graph-level feature vector obtained from quantum measurement. This multilayer perceptron contains several fully connected layers and introduces nonlinearity using activation functions such as ReLU. The final output layer has the same number of neurons as the number of fault categories, C, and uses the Softmax activation function to transform the output into a predicted probability distribution for each fault category. For node-level localization tasks, another approach using classic feature vectors can be similarly employed. For a multilayer perceptron with input, output a probability for each node that "this node belongs to the fault location".

[0137] In this embodiment, the quantum-classical hybrid graph neural network is trained using a hierarchical progressive strategy, and its training loss function is... for:

[0138]

[0139] In the formula, Cross-entropy loss represents the classification of fault types. This represents the mean square error of the fault location regression. , They represent and The weighting coefficients.

[0140] in, , One-hot encoded labels that indicate the fault type.

[0141] In the training process of a quantum-classical hybrid graph neural network, optimization is performed on a classical computer using gradient descent (e.g., the Adam optimizer). Since the parameters of the quantum circuits are continuously differentiable, their gradients can be calculated using quantum gradient estimation methods such as parameter shift rules. The entire training process is end-to-end, with the gradient propagating back from the classical output layer through quantum measurements and quantum circuits to the classical encoding layer, thus achieving joint optimization of all parameters. The model parameters to be trained include the classical encoding layer weights and the quantum circuit parameters. And the classic output layer weights.

[0142] In S500 of this embodiment, the real-time monitoring data of the UHV converter transformer to be diagnosed is processed into an enhanced graph model, and then input into the trained fault location model for forward inference. The fault category with the highest probability output by the model is the final fault diagnosis and location result.

[0143] In this embodiment, a typical ±800kV UHV converter transformer is used as the implementation object. The fault location process is as follows:

[0144] First, multiple sensors were deployed on the converter transformer to collect multi-source monitoring data, including: oil chromatography data: concentration values ​​of H2, CH4, C2H2, C2H4, and C2H6 were obtained from an online dissolved gas monitoring device in the oil. Vibration data: triaxial vibration acceleration sensors were installed at six key locations on the converter transformer tank wall (such as the winding corresponding area and the core corresponding area) to collect vibration signals. Temperature data: top oil temperature, winding hot spot temperature, etc., were collected. A total of 15 sensor measurement points were selected as nodes in the graph.

[0145] Then, construct the graphical model of the commutator. Among them, node set : Includes the above 15 sensor measurement points, for example Edge set Based on the internal electrical connection diagram and physical structure of the converter transformer, the relationships between components are defined. For example, winding nodes on the same core column are connected, and vibration measurement points that are spatially adjacent are connected. A total of 20 undirected edges are constructed. Node characteristics. Each node corresponds to a feature vector. For example, the feature of a vibration measuring point can be the time domain and frequency domain features (root mean square, peak value, frequency components, etc.) extracted from its vibration signal within 1 hour; the feature of an oil chromatography measuring point is the concentration of five gases.

[0146] Taking a node of a triaxial vibration sensor in ZZDFP-345000 / 1000 as an example, the continuous coherence analysis process is as follows:

[0147] Point cloud construction: Take three-channel data from 1024 consecutive sampling points of the sensor. Without performing complex feature extraction, the triaxial readings at each time step are directly treated as a point in three-dimensional space. This results in a 3D point cloud containing 1024 points. Continuous cohomology calculation: Using Python's giotto-tda library, Vietoris-Rips complex flow is applied to this point cloud. The "birth" and "death" times of topological features (0-dimensional connected components, 1-dimensional loops) are recorded, starting from a scale parameter ε=0, to obtain a set of persistent intervals. Feature vectorization: Using the "persistent landscape" method, the above persistent intervals are transformed into a 50-dimensional feature vector. This vector represents the topological feature of the vibration signal at that node. The above operation is performed on each of the 15 vibration and temperature time-series data nodes, generating a unique topological feature vector for each node.

[0148] The original feature vector (e.g., 10-dimensional) of each node is concatenated with its topological feature vector (50-dimensional) to form a new 60-dimensional feature vector, resulting in the augmented graph model. ,in This refers to the enhanced node features.

[0149] Construct a quantum-classical hybrid graph neural network using quantum machine learning libraries such as PennyLane or TensorFlow Quantum, specifically:

[0150] Classical coding layer: A fully connected neural network that maps 60-dimensional augmented node features to the quantum state amplitude of 4 qubits (requires 2^4=16-dimensional data, which can be processed by padding or truncation).

[0151] Quantum graph convolutional layer: Design a 4-qubit variable quantum circuit, which contains multiple layers. Each layer includes: applying a Ry rotation gate to each qubit, with the rotation angle being a trainable parameter; and applying a controlled Z gate to the qubits corresponding to associated nodes according to the adjacency matrix of the enhanced graph model, thereby realizing the interaction of information between nodes.

[0152] Quantum Measurement and Feature Extraction Layer: Measurements are performed on the four qubits in the Z direction, and the expected values ​​form a 4-dimensional classical eigenvector.

[0153] Output layer: A 3-layer fully connected neural network. The input is 4-dimensional measurement features, and the final output is a 6-dimensional vector, which represents the predicted probability of 6 states: "normal", "inter-turn short circuit in winding", "multi-point grounding of iron core", "bushing discharge", "overheating fault" and "partial discharge".

[0154] Training process: Collect 500 sets of valid data (including annotations of the above 6 states) from the past few years of operation of this type of converter transformer as the training set. Use the Adam optimizer with cross-entropy as the loss function to jointly train the classical parameters and quantum circuit parameters in the hybrid model until the model's accuracy on the validation set converges.

[0155] When diagnosing a converter transformer, data from its current 15 measuring points are collected to generate an enhanced graph model. This graph data is then input into a pre-trained model. The model outputs six probability values, and the category corresponding to the highest probability is taken as the diagnostic result. For example, if the output is [0.02, 0.01, 0.05, 0.90, 0.01, 0.01], the diagnostic result is "buffer discharge," and the fault location is achieved by combining this high-probability node with its physical location.

[0156] Example 2

[0157] A fault location system for UHV converter transformers based on QGNNs and PH is provided to implement the fault location method for UHV converter transformers based on QGNNs and PH in Example 1, comprising:

[0158] The graph modeling module is used to collect multi-source heterogeneous monitoring data of UHV converter transformers and construct a graph model of the monitoring data based on the physical structure.

[0159] The continuous homology feature extraction module is used to perform continuous homology analysis on the multidimensional time-series data of nodes in the graph model and output topologically invariant features.

[0160] The enhanced graph model building module is used to fuse topology-invariant features with original node features to obtain enhanced node features and build an enhanced graph model.

[0161] The quantum-classical hybrid graph neural network module is used to train and form a fault location model based on the augmented graph model, and to infer the augmented graph model to be diagnosed and output the predicted results of fault type and fault location.

[0162] Example 3

[0163] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the fault location method for UHV converter transformers based on QGNNs and PH as described in Embodiment 1.

[0164] Example 4

[0165] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the UHV converter transformer fault location method based on QGNNs and PH as described in Example 1.

[0166] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0167] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0168] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0169] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0170] Those skilled in the art will understand that all or part of the steps in the above facts and methods can be implemented by a program instructing related hardware. The program or the program described therein can be stored in a computer-readable storage medium. When the program is executed, it includes the following steps: at this time, the corresponding method steps are introduced. The storage medium can be ROM / RAM, magnetic disk, optical disk, etc.

[0171] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault location method for ultra-high voltage converter transformers based on QGNNs and PH, characterized in that, include: Multi-source heterogeneous monitoring data of UHV converter transformers are collected, and the monitoring data is constructed into a graph model based on the physical structure of the converter transformers. The nodes of the graph model are used to characterize the sensor placement location or monitored components, the edges of the graph model are used to characterize the physical connection relationship or functional association relationship between the components, and the node features are used to characterize the multi-dimensional time-series monitoring data of the corresponding nodes. By performing continuous homology analysis on multidimensional temporal data of at least some nodes in a graph model, topological invariant features characterizing the topological structure of the multidimensional temporal data are obtained. The topology-invariant features are fused with the original node features of the corresponding nodes to obtain enhanced node features, and an enhanced graph model is constructed. Based on the enhanced graph model, a quantum-classical hybrid graph neural network is constructed and trained to obtain a fault location model; The real-time monitoring data of the UHV converter transformer to be diagnosed is processed into an enhanced graph model and input into the fault location model. The predicted results of the fault type and fault location are output as the fault diagnosis and location results of the UHV converter transformer.

2. The method for fault location of UHV converter transformers based on QGNNs and PH as described in claim 1, characterized in that, The persistent cohomology analysis includes: The multidimensional time-series data is preprocessed and constructed into a high-dimensional point cloud dataset based on phase space reconstruction. Based on fault sensitivity weights and setting multi-scale parameters, continuous cohomology calculation is performed on the high-dimensional point cloud dataset to obtain the continuous interval set of each topologically invariant feature. The set of persistent intervals is vectorized to obtain the topologically invariant features.

3. The UHV converter transformer fault location method based on QGNNs and PH according to claim 2, characterized in that, Preprocessing of the multidimensional time series data includes: performing outlier detection, noise filtering, and data standardization sequentially to obtain standard time series data; The construction of a high-dimensional point cloud dataset includes: determining the time delay parameter τ and the embedding dimension m, and reconstructing the phase space of the standard time series data according to the time delay parameter τ and the embedding dimension m to obtain a phase space point cloud; and performing quality evaluation and optimization on the phase space point cloud to obtain a high-dimensional point cloud dataset for continuous cohomology computation.

4. The UHV converter transformer fault location method based on QGNNs and PH according to claim 2, characterized in that, The fault sensitivity weight and multi-scale parameter setting include: assigning fault sensitivity weights to data points in a high-dimensional point cloud dataset, and determining the weighted distance between any pair of data points based on the fault sensitivity weights; The continuous cohomology calculation includes: under the scale sequence corresponding to the multi-scale parameters, a simple complex filter flow is constructed stepwise based on the weighted distance; during the filter flow process, the birth scale and death scale of the topologically invariant features are calculated and recorded; and the birth scale and death scale are combined to form the continuous interval.

5. The UHV converter transformer fault location method based on QGNNs and PH according to claim 2, characterized in that, The feature vectorization of the set of sustained intervals includes: The set of persistent intervals is filtered to retain persistent intervals whose persistence meets preset conditions; The preserved persistent intervals are converted into corresponding persistent landscape functions, and at least one landscape layer is extracted; The landscape layer is uniformly sampled within a preset scale range to obtain a discrete numerical sequence; By concatenating discrete numerical sequences corresponding to different topological dimensions and different landscape layers, topologically invariant features with fixed dimensions are obtained.

6. The UHV converter transformer fault location method based on QGNNs and PH according to claim 2, characterized in that, The quantum-classical hybrid graph neural network comprises, in sequence, a feature encoding layer, a quantum graph convolutional layer, a quantum measurement layer, and a decoding and output layer; The feature encoding layer is used to normalize the features of the enhanced nodes and map them to quantum states using amplitude encoding; the quantum graph convolution layer is used to perform quantum domain graph convolution on the quantum states based on the adjacency relationship of the enhanced graph model; the quantum measurement layer is used to perform quantum measurement on the convolved quantum states to obtain classical feature vectors; and the decoding and output layer is used to output the fault type probability and / or fault location probability based on the classical feature vectors.

7. The UHV converter transformer fault location method based on QGNNs and PH according to claim 6, characterized in that, The quantum graph convolutional layer achieves adjacency information interaction by applying a two-qubit entanglement gate to the corresponding qubits of adjacent nodes, and achieves node feature transformation by applying a single-qubit parameterized rotation gate to the corresponding qubits of nodes. The quantum measurement layer includes: calculating the Pauli operator expectation value for the corresponding qubit in the output quantum state to obtain the node-level classical eigenvector; and summing or averaging the node-level classical eigenvectors to obtain the graph-level classical eigenvectors. The decoding and output layer outputs the probability distribution of fault types based on the graph-level classical feature vectors, and outputs the probability distribution of the nodes corresponding to the fault locations based on the node-level classical feature vectors.

8. A fault location system for ultra-high voltage converter transformers based on QGNNs and PH, used to implement the fault location method for ultra-high voltage converter transformers based on QGNNs and PH as described in any one of claims 1-7, characterized in that, include: The graph modeling module is used to collect multi-source heterogeneous monitoring data of UHV converter transformers and construct a graph model of the monitoring data based on the physical structure. The continuous homology feature extraction module is used to perform continuous homology analysis on the multidimensional time-series data of nodes in the graph model and output topologically invariant features. The enhanced graph model building module is used to fuse topology-invariant features with original node features to obtain enhanced node features and build an enhanced graph model. The quantum-classical hybrid graph neural network module is used to train and form a fault location model based on the augmented graph model, and to infer the augmented graph model to be diagnosed and output the predicted results of fault type and fault location.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the UHV converter transformer fault location method based on QGNNs and PH as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the UHV converter transformer fault location method based on QGNNs and PH as described in any one of claims 1 to 7.

Images