Power distribution station equipment monitoring and health evaluation method based on multi-modal information

Abstract

The invention discloses a power distribution station equipment monitoring and health evaluation method based on multi-modal information, and the method comprises the steps: collecting multi-dimensional parameters of power distribution station equipment, carrying out the dynamic clustering through an improved fuzzy C-means clustering algorithm, and carrying out the aggregation of related parameter subsets; and constructing a hypergraph neural network model, taking the parameter subset as a node, generating a hypergraph adjacency matrix containing a multi-parameter coupling relationship, extracting deep features through multilayer hyperedge convolution, generating a quantitative evaluation result through reasoning, and dividing health levels. According to the method, parameter adaptability is enhanced through a model formula, key steps are refined, parameter association is mined, the defects that traditional clustering precision is insufficient and a conventional neural network cannot model high-order association are overcome, and comprehensiveness and accuracy of health management of power distribution station equipment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power distribution station equipment monitoring, and particularly relates to a power distribution station equipment monitoring and health evaluation method based on multi-modal information. BACKGROUND

[0002] With the development of power distribution station equipment towards high capacity and high integration, the equipment presents the characteristics of multi-parameter strong coupling and nonlinear state change in the operation process. The traditional monitoring method based on single parameter or simple correlation analysis cannot fully capture the health status of the equipment. As a key node of the power system, the equipment of the power distribution station covers various types such as transformers, circuit breakers and switch cabinets. During operation, it involves multi-dimensional parameters such as voltage, current, temperature, vibration and insulation performance. These parameters have complex correlations in the time and space dimensions and present dynamic change characteristics affected by environmental factors and load fluctuations, which puts higher requirements on the accurate monitoring and evaluation of the health status of the equipment.

[0003] The existing technology has two significant shortcomings in the monitoring and health evaluation of power distribution station equipment. On the one hand, the traditional clustering algorithm has difficulty in effectively describing the fuzzy correlation characteristics between parameters when processing multi-parameter data, resulting in insufficient division accuracy of the similarity of equipment operating states, and unable to accurately aggregate parameter subsets with potential correlations, so that the subsequent state evaluation loses reliable data basis. On the other hand, the conventional neural network model can only handle simple correlations between pairs of parameters, and cannot model high-order coupling relationships between multiple parameters, ignoring the influence of the cooperative action of multiple parameters in the power distribution station equipment on the health status, resulting in evaluation results that do not reflect the overall characteristics of the equipment operation, and making it difficult to meet the needs of equipment health evaluation under complex working conditions. SUMMARY

[0004] In order to overcome the shortcomings and deficiencies of the existing technology, the present application provides a power distribution station equipment monitoring and health evaluation method based on multi-modal information.

[0005] The technical solution adopted by the present application is a power distribution station equipment monitoring and health evaluation method based on multi-modal information, comprising the following steps:

[0006] Step S1: Real-time acquisition of multi-dimensional parameter data including equipment operating voltage, load current, winding temperature, insulation resistance, mechanical vibration frequency and gas leakage concentration by deploying different types of sensing devices in different areas of the power distribution station, forming an original monitoring data set containing time and space correlation characteristics;

[0007] Step S2: Based on the improved fuzzy C-means clustering algorithm, the original monitoring data set is dynamically clustered, the optimal number of clusters is determined according to the nonlinear correlation characteristics of the equipment parameters, and the equipment parameter subsets with similar operating characteristics are divided;

[0008] Step S3: Construct the topology of the hypergraph neural network device monitoring and evaluation model, take the subset of device parameters as hypergraph nodes, and generate a hypergraph adjacency matrix by fusing multi-parameter coupling relationships through defining high-order association edges between nodes.

[0009] Step S4: Train the hypergraph neural network using the hypergraph adjacency matrix, extract deep hidden features of device parameters through multi-layer hyperedge convolution operations, and output a feature vector containing device health status related information.

[0010] Step S5: Based on the trained hypergraph neural network model, perform feature mapping and state inference on the subset of real-time input device parameters to generate a quantitative assessment result representing the health status of the device;

[0011] Step S6: Based on the quantitative assessment results, classify the equipment health status levels and generate a comprehensive assessment report on the equipment operating status that includes the influence of different parameters.

[0012] Furthermore, in step S2, the improved fuzzy C-means clustering algorithm performs dynamic clustering optimization by introducing device parameter weight coefficients, and defines the clustering objective function as:

[0013]

[0014] Among them, u ij This represents the membership degree of the i-th device parameter sample to the j-th cluster, where m is the fuzzy factor and x is the number of clusters. i Let v be the parameter sample vector of the i-th device. j Let λ be the cluster center vector of the j-th cluster, λ be the regularization parameter, n be the total number of samples, c be the number of clusters, p be the device parameter dimension, and ω be the number of clusters. k Let be the weighting coefficient for the k-th device parameter. Device parameters include voltage, current, and temperature, and satisfy [condition missing]. During the iteration process, the membership matrix is ​​updated. Clustering optimization is performed, and the weight coefficients ω are dynamically adjusted based on the time-varying characteristics of the device parameters. k This makes the clustering results more closely match the actual operating status of the power distribution station equipment.

[0015] Furthermore, in step S3, when constructing the hypergraph adjacency matrix, a spatiotemporal attenuation factor of the device parameters is introduced, and the hypergraph edge weight calculation formula is defined as follows:

[0016]

[0017] Among them, w e Let v be the weight of hyperedge e, and u be the device parameter nodes connected to hyperedge e. v and t uare the collection timestamps of parameters v and u, respectively, a is the time decay coefficient, β vu is the physical correlation coefficient of parameters v and u, γ(v, u) is the spatial distance weight of parameters v and u, and |e| is the number of nodes contained in the hyperedge e.

[0018] Further, in the step S4, the hyperedge convolution layer operation of the hypergraph neural network is defined as:

[0019]

[0020] where H (l) and H (l+1) are the hypergraph node feature matrices of the Ith layer and the (I+1)th layer, respectively, E is the set of all hyperedges in the hypergraph, w e is the weight of the hyperedge e, deg(v) is the degree of the node v, deg(e) is the degree of the hyperedge e, is the weight matrix corresponding to the hyperedge e of the Ith layer, b (l) is the bias vector of the Ith layer, and σ is the activation function.

[0021] Further, in the step S5, the quantitative evaluation function of the output layer of the hypergraph neural network is defined as:

[0022]

[0023] where S is the quantitative evaluation result of the device health status, θ k is the feature weight of the kth device parameter, f k (·) is the mapping function of the kth parameter feature, h k is the feature vector of the kth parameter output by the hypergraph neural network, μ e is the influence coefficient of the hyperedge e, g e (·) is the collaborative feature mapping function of the parameters associated with the hyperedge e, h e is the collaborative feature vector of the parameters associated with the hyperedge e.

[0024] Further, in the process of connecting the step S2 and the step S3, the mapping relationship function of the parameter subset to the hypergraph node is defined as:

[0025] N s =∑ i∈s η i x i +∑ i∈s ∑ j∈s，j≠i ξ ij (x i ⊙x j )

[0026] where N s is the hypergraph node feature corresponding to the parameter subset s, ηi is the node mapping weight of the i-th parameter in the subset s, ξ ij is the interaction mapping coefficient of the i-th and j-th parameters in the subset s, and is the element product operation.

[0027] Further, the step S3 comprises the following sub-steps:

[0028] Step S3.1: Based on the device parameter subset obtained in step S2, determine the hypergraph node set, each node corresponding to a comprehensive feature of a parameter subset, and preliminarily screen out the node combination with potential association by calculating the feature similarity between different parameter subsets;

[0029] Step S3.2: According to the physical connection relationship of the power distribution station equipment and the electrical characteristics of the parameters, define the generation rule of the hyperedge, and include the parameter subset nodes related to the same equipment component or having strong electrical coupling relationship into the same hyperedge to form an initial hyperedge set;

[0030] Step S3.3: Calculate the parameter fluctuation covariance of the nodes contained in each hyperedge, adjust the range of the hyperedge according to the covariance value, merge the hyperedges with consistent fluctuation trend, split the hyperedges containing nodes with too large fluctuation characteristics difference, and optimize the hypergraph structure;

[0031] Step S3.4: Based on the number of nodes contained in the hyperedge and the importance of the parameters, calculate the initial weight value of different hyperedges, establish a complete hypergraph adjacency matrix containing node, hyperedge and weight information, and provide topological structure data for the training of the hypergraph neural network.

[0032] Further, the step S4 comprises the following sub-steps:

[0033] Step S4.1: Input the hypergraph adjacency matrix obtained in step S3 and the device parameter subset features into the input layer of the hypergraph neural network, fuse the node features and hypergraph topological information through matrix transformation, and convert them into a feature matrix form suitable for hyperedge convolution operation;

[0034] Step S4.2: In the hidden layer of the hypergraph neural network, perform hyperedge convolution operation, weight and aggregate the node features through the hyperedge weight matrix, and introduce the physical constraint condition of the device parameters to adjust the convolution kernel parameters, so that the feature extraction process conforms to the operation law of the power distribution station equipment;

[0035] Step S4.3: After each layer of hyperedge convolution operation, perform nonlinear transformation on the output features, and enhance the model's ability to capture nonlinearly associated features through the activation function;

[0036] Step S4.4: Through the stacking of multiple layers of hyperedge convolution operation, gradually improve the abstraction level of the features, and convert the shallow physical features of the device parameters into deep health state related features.

[0037] Further, the step S5 comprises the following sub-steps:

[0038] Step S5.1: input the deep feature vector output by step S4 into the output layer of the hypergraph neural network, map the high-dimensional features to the feature space related to the device health state through feature dimension transformation, so that the dimension of the feature vector matches the dimension of the health assessment index;

[0039] Step S5.2: weighted fusion is performed on the mapped feature vector, and the weight value is determined according to the fault sensitivity of the device parameter, and the feature corresponding to the parameter with high fault sensitivity is given a higher weight, so as to highlight the influence of the calibration parameter on the health assessment result;

[0040] Step S5.3: the associated features between parameters in the hypergraph are introduced, the cooperative features are included in the evaluation calculation process by calculating the feature cooperation degree of the nodes connected by the hyperedge, and the limitation of single parameter feature evaluation is made up;

[0041] Step S5.4: the fused feature vector is converted into a quantitative evaluation result through a preset mapping rule, the result includes the health state information of different calibration components of the device, and forms the basic data directly used for state level division.

[0042] Beneficial effects: the present application proposes a power distribution station device monitoring and health evaluation method based on multi-modal information, which implements dynamic clustering on multi-dimensional parameter data of the power distribution station device through an improved fuzzy C-means clustering algorithm, improves the accuracy of operation state similarity division by accurately describing the fuzzy association characteristics between parameters, accurately aggregates the parameter subsets with potential association, and lays a data foundation for subsequent evaluation work, solving the problem of insufficient division accuracy when processing multiple parameters by traditional clustering algorithms. At the same time, relying on the hypergraph neural network device monitoring and evaluation model, a hypergraph topology structure containing multi-parameter coupling relationship is constructed, the high-order association between parameters is captured through hyperedge convolution operation, the individual feature and cooperative feature of the parameters are fused to generate a quantitative evaluation result, breaking the limitation of conventional neural networks that can only process simple association of paired parameters, and fully reflecting the overall characteristics of the device operation. By constructing the model formula to strengthen the adaptability of the parameters and the model, the key steps are refined to enhance the depth and breadth of parameter association mining, so that the method can accurately capture the characteristics of strong coupling of multiple parameters and nonlinearity of state change of the device, meet the needs of device health evaluation under complex working conditions, and improve the comprehensiveness and accuracy of power distribution station device health management. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is the overall step flowchart of the present application;

[0044] Figure 2 is the system architecture diagram of the present application. Detailed Implementation

[0045] It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. The following describes the application in further detail with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown, a method for monitoring and health assessment of substation equipment based on multimodal information includes the following steps:

[0047] Step S1: By deploying different types of sensing devices in different areas of the substation, real-time collection of multi-dimensional parameter data including equipment operating voltage, load current, winding temperature, insulation resistance, mechanical vibration frequency, and gas leakage concentration is carried out to form an original monitoring dataset containing spatiotemporal correlation characteristics.

[0048] Specifically, step S1 is the data foundation of the entire monitoring and health assessment method. Its core lies in acquiring multi-dimensional parameter data of the substation equipment during operation through various types of sensing devices. These parameters cover multiple aspects such as equipment electrical performance, thermal characteristics, mechanical condition, and safety indicators. Among them, voltage parameters reflect the potential difference state of the equipment in the power system, and its value range is usually related to the rated voltage of the equipment, directly affecting the insulation performance and power transmission efficiency of the equipment; load current reflects the size of the electrical load carried by the equipment, and its changes are closely related to the equipment's heat generation and losses; winding temperature is a key indicator for measuring the internal heat generation of the equipment, and excessively high temperatures will accelerate the aging of insulation materials; insulation resistance reflects the quality of the equipment's insulation performance and is an important parameter for ensuring the safe operation of the equipment; mechanical vibration frequency can reflect the operating status of the internal mechanical structure of the equipment, and abnormal vibration may indicate component wear or loosening; gas leakage concentration is for equipment containing oil or gas insulation, used to judge the sealing performance of the equipment and whether there are internal faults.

[0049] The implementation of this step is to reasonably deploy sensing devices in different areas of the power distribution station, such as installing voltage sensors, current sensors and winding temperature sensors in the transformer area, arranging mechanical vibration sensors near the circuit breaker, and setting gas leakage concentration sensors around gas insulated equipment. The sampling frequency of the sensing device is determined according to the variation characteristics of the parameters. For parameters that change quickly, such as voltage and current, the sampling frequency is set to a high value, usually not less than 200 times per second; for parameters that change slowly, such as temperature and insulation resistance, the sampling frequency can be appropriately reduced, generally 1 time per hour to 1 time per 10 minutes. The data collected by the sensing device is sent to the data processing center in real time through a dedicated data transmission line to form an original monitoring data set containing time stamps and location information. During data collection, the synchronization of each sensing device needs to be ensured to ensure the correlation of different parameters in the time dimension, and preliminary integrity checks are performed on the collected data to eliminate obviously missing or broken data segments to ensure that the original monitoring data set accurately reflects the operating status of the equipment at different times and different locations.

[0050] Step S2: based on the improved fuzzy C-means clustering algorithm, the original monitoring data set is dynamically clustered, the optimal cluster number is determined according to the nonlinear correlation characteristics of the equipment parameters, and the equipment parameter subsets with similar operating characteristics are divided;

[0051] Specifically, step S2 dynamically clusters the original monitoring data set based on the improved fuzzy C-means clustering algorithm, which mainly aims to effectively classify multi-dimensional parameter data and mine parameter subsets with similar operating characteristics, providing targeted data units for subsequent hypergraph model construction and health evaluation. The parameter data processed in this step has characteristics such as nonlinearity, high dimensionality and spatiotemporal correlation. The improved fuzzy C-means clustering algorithm can automatically optimize the cluster number according to the distribution characteristics of the parameter data by introducing a dynamic adjustment mechanism, avoiding the classification bias caused by the preset cluster number in traditional clustering algorithms. Through clustering, parameter data with similar variation trends and correlation characteristics can be classified into the same subset, so that subsequent model processing can focus on parameter groups with inherent connections, improving the relevance and accuracy of health evaluation.

[0052] In the implementation process, first, the original monitoring data set obtained in step S1 is converted into a data format suitable for clustering algorithm processing, and the numerical information, time stamp and position label of each parameter are retained. The improved fuzzy C-means clustering algorithm initially sets multiple clustering centers, and the initial values of these centers are determined according to the statistical characteristics of the parameter data, for example, initialized based on the mean and standard deviation of each parameter. The algorithm determines the possibility of data points belonging to different clusters by calculating the membership of each parameter data point to each cluster center, and iteratively updates the cluster centers according to the membership values. In the iteration process, according to the characteristics of the power distribution station equipment parameters, the algorithm dynamically adjusts the fuzzy coefficient of clustering. For parameters such as voltage and current that have strong correlation, the fuzzy coefficient is appropriately reduced to enhance the certainty of classification; for parameters such as vibration frequency and gas leakage concentration that are greatly affected by the environment, the fuzzy coefficient is appropriately increased to tolerate a certain fluctuation range. When the change of the cluster center is less than a preset threshold, the iteration is stopped, and at this time the obtained clusters are the parameter subsets of the equipment with similar operating characteristics. Each parameter subset contains complete data records of all parameters in the cluster, including parameter values at each time, corresponding equipment location information, etc., providing structured data input for subsequent hypergraph model construction.

[0053] Step S3: Constructing the topology structure of the hypergraph neural network equipment monitoring evaluation model, taking the equipment parameter subsets as hypergraph nodes, fusing the multi-parameter coupling relationship by defining high-order correlation edges between nodes, and generating a hypergraph adjacency matrix;

[0054] Specifically, the core of step S3 is to construct the topology structure of the hypergraph neural network equipment monitoring evaluation model. This step is the key link connecting the parameter clustering results and the neural network model, and its significance lies in converting the abstract parameter subsets into a graphical structure that can reflect the complex correlation between parameters. Compared with traditional graph models, hypergraph models can better handle high-order correlation relationships between multiple parameters, and are suitable for the characteristics of power distribution station equipment multi-parameter cooperation. By taking the equipment parameter subsets as hypergraph nodes and defining high-order correlation edges between nodes, the electrical coupling, physical connection, causal influence and other relationships between parameters can be integrated into the hypergraph structure, so that the hypergraph neural network can fully utilize these correlation information in the training and reasoning process, and improve the comprehensiveness and accuracy of health evaluation. The generation of the hypergraph adjacency matrix provides a mathematical basis for the calculation of the hypergraph neural network, and the elements in the matrix can quantitatively represent the correlation strength between nodes, providing a quantitative basis for subsequent feature extraction and state reasoning.

[0055] In implementation, first, the node set of the hypergraph is determined, each node corresponding to a device parameter subset obtained in step S2, and the characteristic information of the node includes the statistical characteristics of each parameter in the subset, such as mean, variance, and change rate. Subsequently, the hyperedge is defined according to the physical structure of the power distribution station device and the internal relationship between the parameters, and for the parameter subsets belonging to the same device component, such as the voltage, current, and winding temperature parameter subsets of the transformer, a hyperedge is constructed to connect them; for the parameter subsets that exist in the electrical loop association, such as the current parameter subset of the circuit breaker and the current parameter subset of the adjacent line, a hyperedge is also established to connect them. In defining the hyperedge, the type of the parameter, the connection relationship of the device, and the operation mechanism are comprehensively considered to ensure that the hyperedge can accurately reflect the substantial association between the parameters. In the construction process of the hypergraph adjacency matrix, the value of the matrix element is determined according to the parameter association strength of the nodes connected by the hyperedge, and the association strength is obtained by calculating the correlation coefficient and the common change frequency between the parameter subsets. For example, for the voltage and current parameter subsets of the transformer, since there is a close electrical association between them, the value of the corresponding adjacency matrix element is set to a high level; for parameter subsets with weak association, such as temperature parameter subsets in different regions, the value of the adjacency matrix element is set to a low level. The generated hypergraph adjacency matrix contains the association information between all nodes, and the dimension of the matrix is consistent with the number of parameter subsets, providing complete topological structure data for the next step of processing of the hypergraph neural network.

[0056] Step S4: training the hypergraph neural network using the hypergraph adjacency matrix, extracting deep implicit features of the device parameters through multi-layer hyperedge convolution operation, and outputting a feature vector containing device health state association information;

[0057] Specifically, step S4 trains the hypergraph neural network using the hypergraph adjacency matrix, and the main significance is to extract deep implicit features of the device parameters through multi-layer hyperedge convolution operation, and convert the original parameter data into an abstract feature vector that can represent the device health state. This step is the core process of the hypergraph neural network learning the parameter association pattern and the health state mapping relationship, and through training, the network can understand the influence of the complex interaction between parameters on the device health. The extraction of deep implicit features breaks away from the appearance restriction of the original parameters and can capture the potential health information hidden in the parameter changes, such as the precursor features of abnormal parameter fluctuations and the trend features of multi-parameter collaborative degradation, which are crucial for accurate evaluation of the device health state. Parameter adjustment during training can make the hypergraph neural network adapt to the specific operating environment and parameter characteristics of the power distribution station device, and improve the recognition ability of the model for the device health state.

[0058] In implementation, first, the hypergraph adjacency matrix generated in step S3 and the feature data of the parameter subset are input into the input layer of the hypergraph neural network. The initial weight parameters of the network are randomly initialized according to the normal distribution, and the bias parameters are set to zero. The training process adopts a batch processing manner, and the data set is divided into multiple batches, each batch containing a certain number of hypergraph samples. In each round of training, the hypergraph neural network processes the node features through the hyperedge convolution layer. The hyperedge convolution operation fuses and calculates the features of each node with the features of other nodes in the connected hyperedge according to the correlation strength in the hypergraph adjacency matrix, realizing the transmission and aggregation of the features. After multiple layers of hyperedge convolution operation, the output layer of the network generates a preliminary feature vector. The feature vector is compared with the preset device health state label, and the loss value between the two is calculated, which is calculated using the cross-entropy loss function. According to the loss value, the weight parameters and bias parameters of each layer of the network are adjusted through the back propagation algorithm. The adjustment process adopts the gradient descent optimization method, and the learning rate is dynamically adjusted according to the change of the loss value. When the loss value decreases rapidly, the learning rate is appropriately increased to accelerate convergence; when the loss value fluctuates slightly, the learning rate is reduced to improve accuracy. The training process continues until the loss value stabilizes at a preset lower level. At this time, the hypergraph neural network can stably output a feature vector containing device health state correlation information, and the training is completed.

[0059] Step S5: Based on the trained hypergraph neural network model, the feature mapping and state reasoning of the real-time input device parameter subset are performed to generate a quantitative evaluation result representing the health degree of the device;

[0060] Specifically, step S5 performs feature mapping and state reasoning on the real-time input device parameter subset based on the trained hypergraph neural network model, which means that the trained model is applied to the actual device monitoring scene to realize dynamic evaluation of the device health state. This step can utilize the parameter correlation pattern and health state mapping relationship learned by the model to quickly convert the real-time collected parameter data into a quantitative health evaluation result, reflecting the current running state of the device in a timely manner. Compared with the traditional periodic detection method, this real-time reasoning can timely discover early abnormalities of the device, providing timely decision basis for maintenance and repair of the device, reducing power outage accidents and economic losses caused by device failure. The generation of quantitative evaluation results makes the description of device health state more accurate, which can distinguish different degrees of health degradation and provide support for hierarchical maintenance.

[0061] In the implementation process, first, the real-time collected device parameters are divided into corresponding parameter subsets according to the clustering method of step S2, ensuring that the input parameter subset structure is consistent with the structure during training. These parameter subsets are input into the trained hypergraph neural network model. The model first extracts features from the input parameter subsets and maps the real-time parameter data to the feature space of the model through the same hyperedge convolution operation as in the training process. During the feature mapping process, the model uses the trained weight parameters and bias parameters to calculate and integrate the association information of the real-time parameters, generating a feature vector related to the device health status. Subsequently, the model converts the feature vector into a quantitative evaluation result through the state reasoning mechanism of the output layer. The numerical range of the evaluation result is set according to the health status level of the device, for example, setting a baseline value. When the evaluation result is higher than the baseline value, it indicates that the device health status is good; when the evaluation result is lower than the baseline value, it indicates that the device has different degrees of health problems. During the reasoning process, it is necessary to ensure that the input speed of real-time data matches the processing speed of the model. Typically, the reasoning time of the model is controlled within milliseconds to meet the real-time monitoring requirements. At the same time, the intermediate results generated during the reasoning process are recorded, including the numerical value of the feature vector and the reasoning path, providing data support for subsequent result analysis and model optimization.

[0062] Step S6: According to the quantitative evaluation result, divide the device health status level and form a comprehensive evaluation report of the device running state containing the influence of different parameter associations.

[0063] Specifically, step S6 divides the device health status level according to the quantitative evaluation result and forms a comprehensive evaluation report, which has the significance of converting abstract quantitative results into decision-making information with practical application value, providing clear basis for the operation management and maintenance strategy formulation of the power distribution station equipment. The division of health status level can intuitively reflect the health status of the device, enabling the management personnel to quickly understand the operation risk of the device and prioritize the processing of devices with poor health status. The comprehensive evaluation report integrates the association influence information of various device parameters, detailing the key parameters and associated factors affecting the device health status, helping technical personnel to deeply analyze the reasons for the deterioration of device health and develop targeted maintenance measures. This step realizes a complete closed loop from data collection to decision support, improving the scientificity and effectiveness of the health management of the power distribution station equipment.

[0064] In practice, the quantitative assessment results generated in step S5 are first categorized according to a preset health status level classification standard. The classification standard is typically determined based on historical equipment failure data and operational experience. For example, the assessment results may be divided into five levels: healthy, basically healthy, slightly deteriorated, moderately deteriorated, and severely deteriorated, with each level corresponding to a specific numerical range for the assessment results. Equipment at the slightly deteriorated level or above is highlighted in the report, and the main parameters leading to this level and their changing trends are listed in detail. When generating the comprehensive assessment report, the report content includes basic equipment information (such as equipment model, installation location, and years of operation), the number and distribution of equipment at each health status level, the quantitative assessment results for each piece of equipment, analysis of key parameters affecting health status, and explanations of the correlation between parameters. The report is generated using an automated template generation method, filling the processed information into the report template according to a preset format to ensure the report's standardization and consistency. Simultaneously, the report includes data visualization charts, such as parameter change trend graphs and health status distribution pie charts, to enhance readability. The generated report is sent to the substation's management and maintenance personnel via the internal network and stored in the database, forming a historical record of the equipment's health status and providing data support for the equipment's full lifecycle management. After the report is generated, its completeness and accuracy must be checked to ensure that all key information is included and the data is correct.

[0065] Preferably, in step S2, the improved fuzzy C-means clustering algorithm performs dynamic clustering optimization by introducing device parameter weight coefficients, and the clustering objective function is defined as:

[0066]

[0067] Among them, u ij This represents the membership degree of the i-th device parameter sample to the j-th cluster, where m is the fuzzy factor and x is the number of clusters. i Let v be the parameter sample vector of the i-th device. j Let λ be the cluster center vector of the j-th cluster, λ be the regularization parameter, n be the total number of samples, c be the number of clusters, p be the device parameter dimension, and ω be the number of clusters. k Let be the weighting coefficient for the k-th device parameter. Device parameters include voltage, current, and temperature, and satisfy [condition missing]. During the iteration process, the membership matrix is ​​updated. Clustering optimization is performed, and the weight coefficients ω are dynamically adjusted based on the time-varying characteristics of the device parameters. k This makes the clustering results more closely match the actual operating status of the power distribution station equipment.

[0068] Specifically, an improved fuzzy C-means clustering algorithm is used to introduce weight coefficients for equipment parameters to achieve dynamic clustering optimization. This is significant because it enhances the adaptability of clustering to the multi-parameter characteristics of substation equipment. In practice, weights are first assigned based on the degree of influence of each parameter on equipment health. Specifically, the weight for operating voltage is set to 0.22, load current to 0.21, winding temperature to 0.19, insulation resistance to 0.15, mechanical vibration frequency to 0.12, and gas leakage concentration to 0.11. During clustering, the algorithm uses these weights as a basis to iteratively update the membership matrix to adjust the cluster affiliation of data points. The initial cluster centers are set according to the mean values ​​of each parameter; for example, the voltage cluster center is set to 10kV, and the current cluster center is set to 200A. After each iteration, the clustering objective function value is calculated. The iteration stops when the change in the function value is less than 0.001 after three consecutive iterations. The resulting parameter subset covers similar feature parameters. For example, device data with voltage fluctuation range of 9.5-10.5kV and current fluctuation range of 180-220A are clustered into the same subset to ensure that the clustering results can accurately reflect the intrinsic relationship between parameters and provide high-quality input for subsequent hypergraph modeling.

[0069] Preferably, in step S3, a spatiotemporal attenuation factor of device parameters is introduced when constructing the hypergraph adjacency matrix, and the hypergraph edge weight calculation formula is defined as follows:

[0070]

[0071] Among them, w e Let v be the weight of hyperedge e, and u be the device parameter nodes connected to hyperedge e. v and t u These are the acquisition timestamps for parameters v and u, respectively, where α is the time decay coefficient and β is the time attenuation coefficient. vu Let γ(v, u) be the physical correlation coefficient between parameters v and u (such as the phase correlation coefficient between voltage and current), γ(v, u) be the spatial distance weight between parameters v and u (calculated based on the equipment installation location), and |e| be the number of nodes contained in the hyperedge e. The weight values ​​of all hyperedges in the hypergraph are calculated using this formula, and then the weighted hypergraph adjacency matrix is ​​constructed.

[0072] Specifically, a hypergraph adjacency matrix is ​​constructed by introducing a spatiotemporal decay factor to more accurately quantify the correlation strength between parameters. In implementation, the time decay coefficient α is set according to the parameter change rate: α for voltage and current is set to 0.7, α for temperature and insulation resistance is set to 0.4, and α for vibration frequency and gas concentration is set to 0.5. The physical correlation coefficient βvu is determined based on the degree of electrical coupling: βvu for transformer voltage and current is set to 0.85, and βvu for circuit breaker insulation resistance and vibration frequency is set to 0.5. vuis set to 0.65. The spatial distance weight γ(v, u) is calculated according to the distance between devices, the same cabinet parameter is set to 0.9, the adjacent cabinet parameter is set to 0.6, and the parameter of more than 3 interval cabinets is set to 0.3. When calculating the super-edge weight, for all node pairs in the same super-edge, the space-time correlation value is calculated according to the above parameters and the average value is taken. For example, for a super-edge containing transformer voltage, current and temperature, the weight is obtained by fusing the space-time parameters of the three. The finally constructed supergraph adjacency matrix has the same dimension as the number of parameter subsets, and the matrix element value is between 0 and 1, which accurately reflects the correlation between nodes.

[0073] Preferably, in the step S4, the super-edge convolution layer operation of the supergraph neural network is defined as:

[0074]

[0075] wherein H (l) and H (l+1) are the supergraph node feature matrices of the Ith layer and the I+1th layer respectively, E is the set of all super-edges in the supergraph, w e is the weight of the super-edge e, deg(v) is the degree of the node v (the number of super-edges connected to the node v), deg(e) is the degree of the super-edge e (the number of nodes contained in the super-edge e), is the weight matrix corresponding to the super-edge e of the Ith layer, b (l) is the bias vector of the Ith layer, and σ is the activation function; the deep features of the device parameters are extracted by multiple layers of the super-edge convolution operation.

[0076] Specifically, the super-edge convolution layer operation of the supergraph neural network is defined to extract the deep features of the device parameters, and the core is to strengthen the mining of parameter correlation features through multiple layers of convolution. In implementation, the initial dimension of the supergraph node feature matrix is set to 64x128 (64 parameter subsets, each containing 128-dimensional features), and the super-edge weight w e is taken from the adjacency matrix constructed in step 3. The node degree deg(v) is the number of super-edges connected to the node, and the average is set to 5-8. The super-edge degree deg(e) is the number of nodes contained in the super-edge, and the average is set to 4-6. The weight matrix We (l) of each layer of convolution has a dimension of 128x256 (first layer) and 256x512 (second layer), and the bias vector b(l) has the same dimension as the output feature. The activation function adopts LeakyReLU, and the slope is set to 0.02. In the convolution process, the node features are aggregated and updated at each layer. For example, the 64x128 feature matrix is converted to 64x256 in the first layer, and to 64x512 in the second layer. Through 3-5 layers of convolution operation, the feature dimension is improved to 64x1024, and the high-order correlation features of the parameters are fully extracted, providing deep feature support for health assessment.

[0077] Preferably, in the step S5, the quantification evaluation function of the hypergraph neural network output layer is defined as:

[0078]

[0079] wherein S is the device health state quantification evaluation result, θ k is the feature weight of the kth device parameter, f k (·) is the mapping function of the kth parameter feature, h k is the feature vector of the kth parameter output by the hypergraph neural network, μ e is the influence coefficient of the hyperedge e, g e (·) is the collaborative feature mapping function of the parameters associated with the hyperedge e, h e is the collaborative feature vector of the parameters associated with the hyperedge e; the function fuses the individual features and collaborative features of the parameters to generate the final evaluation result.

[0080] Specifically, the individual and collaborative features of the parameters are fused through the output layer quantification evaluation function to realize the accurate quantification of the device health state. In implementation, the parameter feature weight θ k is assigned according to importance, the winding temperature θ k is set to 0.23, the insulation resistance θ k is set to 0.21, the voltage θ k is set to 0.18, the current θ k is set to 0.17, the vibration frequency θ k is set to 0.11, and the gas concentration θk is set to 0.1. The hyperedge influence coefficient μe is set according to the association strength, the voltage-current hyperedge μ e is set to 0.25, the temperature-vibration hyperedge μ e is set to 0.18, the insulation-gas hyperedge μ e is set to 0.12. The mapping function converts the feature vector into a value of 0-100, the individual feature value is obtained by weighting the feature vector and θ k , and the collaborative feature value is obtained by weighting the hyperedge feature and μ e . The final evaluation result is the sum of the two, ranging from 0 to 100, wherein 90-100 corresponds to health, 70-89 corresponds to basic health, 50-69 corresponds to mild degradation, 30-49 corresponds to moderate degradation, 0-29 corresponds to severe degradation, and comprehensively reflects the device health status.

[0081] Preferably, in the process of connecting the step S2 and the step S3, the mapping relationship function of the parameter subset to the hypergraph node is defined as: N s =∑ i∈s η i x i +∑ i∈s ∑ j∈s，j≠i ξij (x i ⊙x j ), wherein N s is the hypergraph node feature corresponding to the parameter subset s, η i is the node mapping weight of the i-th parameter in the subset s, ξ ij is the interaction mapping coefficient of the i-th and j-th parameters in the subset s, and is the element product operation; the parameter subset obtained by clustering is converted into a hypergraph node through the function, and the clustering result is connected with the features of the hypergraph model.

[0082] Specifically, the conversion of the parameter subset to the hypergraph node is realized through the mapping function, and the information is ensured to be transmitted without loss. In implementation, the node mapping weight η i is set according to the proportion of the parameter in the subset, the voltage η i in the subset is set to 0.3, the current η i is set to 0.28, the temperature η i is set to 0.22, and the other parameters η i are set to 0.2 in total. The interaction mapping coefficient ξ ij is set according to the coupling strength of the parameters, the voltage and the current ξ ij are set to 0.25, the temperature and the vibration ξ ij are set to 0.2, and the insulation and the gas ξ ij are set to 0.15. In the mapping process, the product sum of the parameter value and η i is calculated first, and then the product sum of the element product of the parameters and ξ ij is calculated, and the node feature is obtained by adding the two. For example, if a subset contains a voltage of 10 kV, a current of 200 A, and a temperature of 60°C, the node feature is 10 x 0.3 + 200 x 0.28 + 60 x 0.22 + (10 x 200 x 0.25 + 200 x 60 x 0.2 + 10 x 60 x 0.15) x 0.001, which converts the parameter subset into a hypergraph node feature with uniform dimensions, laying a foundation for hypergraph modeling.

[0083] Preferably, the step S3 comprises the following sub-steps:

[0084] Step S3.1: based on the device parameter subset obtained in step S2, a hypergraph node set is determined, each node corresponding to a comprehensive feature of a parameter subset, and by calculating the feature similarity between different parameter subsets, a node combination with potential association is preliminarily screened out, providing a basis for the construction of hypergraph edges;

[0085] Step S3.2: according to the physical connection relationship of the distribution station equipment and the electrical characteristics of the parameters, the generation rule of the hyperedge is defined, the parameter subset nodes related to the same equipment component or having strong electrical coupling relationship are included in the same hyperedge, and an initial hyperedge set is formed;

[0086] Step S3.3: Calculate the parameter fluctuation covariance of the nodes contained in each super-edge, adjust the range of the super-edge according to the covariance value, merge the super-edges with highly consistent fluctuation trends, split the super-edges containing nodes with too large fluctuation characteristics difference, and optimize the supergraph structure;

[0087] Step S3.4: Calculate the initial weight values of different super-edges based on the number of nodes contained in the super-edge and the importance of the parameters, establish a complete supergraph adjacency matrix containing node, super-edge and weight information, and provide topological structure data for the training of the supergraph neural network.

[0088] Specifically, the supergraph topology is constructed in four steps. The first step is to determine the node set, each node corresponds to a parameter subset, and the cosine similarity between subsets is calculated. The subsets with a similarity greater than 0.7 are marked as potential associated node pairs. The second step is to define the super-edge according to the physical connection of the device. The voltage, current and temperature parameter subsets of the same transformer form a super-edge, and the insulation resistance and vibration parameter subsets of the same circuit breaker form a super-edge, forming 20-30 initial super-edges. The third step is to calculate the covariance of the parameters in the super-edge, and merge the super-edges with a covariance greater than 0.6, and split the super-edges with a covariance less than 0.3, and optimize the super-edge structure. The fourth step is to calculate the weight according to the number of nodes contained in the super-edge and the importance of the parameters. The weight of the super-edge containing more than 3 nodes increases by 0.2, and the weight of the super-edge containing voltage and current increases by 0.3. The final constructed supergraph adjacency matrix is n x n (n is the number of nodes), and the element value is between 0 and 1, accurately reflecting the node association strength.

[0089] Preferably, the step S4 comprises the following sub-steps:

[0090] Step S4.1: Input the supergraph adjacency matrix obtained in step S3 and the device parameter subset features into the input layer of the supergraph neural network, fuse the node features and supergraph topology information through matrix transformation, and convert them into a feature matrix form suitable for super-edge convolution operation, to prepare data for subsequent feature extraction;

[0091] Step S4.2: In the hidden layer of the supergraph neural network, perform super-edge convolution operation, weight and aggregate the node features through the super-edge weight matrix, and introduce the physical constraint condition of the device parameters to adjust the convolution kernel parameters, so that the feature extraction process is more in line with the operation law of the distribution station device;

[0092] Step S4.3: After each layer of super-edge convolution operation, perform nonlinear transformation on the output features, enhance the model's ability to capture nonlinearly associated features through the activation function, and improve the expression ability of deep features;

[0093] Step S4.4: Through the stacking of multiple layers of super-edge convolution operation, gradually improve the abstraction level of the features, and convert the shallow physical features of the device parameters into deep health state associated features, providing high-dimensional feature support for subsequent state evaluation.

[0094] Specifically, the supergraph neural network training is realized through four steps. In the first step, the supergraph adjacency matrix and the parameter features are input into the network, the feature matrix is standardized to the range of 0-1, and the adjacency matrix is normalized by row. In the second step, super-edge convolution is performed in the hidden layer, the convolution kernel size is set to 3x3, the step size is set to 1, the node features are aggregated according to the super-edge weight, and the physical constraint of voltage range 10±0.5kV and current range 200±20A is introduced to adjust the convolution parameters. In the third step, an activation function is applied after each convolution layer, the ReLU function slope is set to 0.01, and the output features are nonlinearly transformed to enhance the feature expression capability. In the fourth step, a 5-layer convolution stack is used, the feature dimension of each layer is doubled, and the input 64-dimensional feature is improved to 1024-dimensional feature. During training, the Adam optimizer is used, the learning rate is initially set to 0.001, it is attenuated by 0.1 every 100 rounds, and the iteration is stopped after 500 rounds to make the model loss value stable below 0.02, ensuring that the deep features of the parameters can be accurately extracted.

[0095] Preferably, the step S5 comprises the following sub-steps:

[0096] Step S5.1: input the deep feature vector output by step S4 into the output layer of the supergraph neural network, map the high-dimensional features to the feature space related to the device health state through feature dimension transformation, and match the dimension of the feature vector with the dimension of the health assessment index;

[0097] Step S5.2: weighted fusion is performed on the mapped feature vector, the weight value is determined according to the fault sensitivity of the device parameters, and the feature corresponding to the parameter with high fault sensitivity is given a higher weight to highlight the influence of the calibration parameter on the health assessment result;

[0098] Step S5.3: the correlation features between parameters in the supergraph are introduced, the collaborative features of the nodes connected by the super-edge are calculated, and the collaborative features are included in the evaluation calculation process to make up for the limitations of single parameter feature evaluation;

[0099] Step S5.4: the fused feature vector is converted into a quantitative evaluation result through a pre-set mapping rule, the result includes the health state information of different calibration components of the device, and forms basic data that can be directly used for state level division.

[0100] Specifically, the health state reasoning is realized through four steps. In the first step, 1024-dimensional deep features are mapped to a 20-dimensional health feature space. Principal component analysis is used to extract the top 20 features with variance contribution, retaining more than 95% of the information. In the second step, the 20-dimensional features are weighted and fused. The weights are sorted by feature importance, with the top 5 features weighted at 0.15, the 6th to 10th features weighted at 0.1, and the 11th to 20th features weighted at 0.05, highlighting the key feature influence. In the third step, the hyperedge collaborative feature degree is calculated. The associated node features are summed according to the hyperedge weight, and the collaborative feature proportion is set to 0.3. The individual features are weighted and fused. In the fourth step, the fused features are converted into quantitative results through pre-set rules. The rules are based on historical failure data, such as reducing 0.8 points for each 1-unit decrease in feature value. The final output is a quantitative value of 0-100, reflecting the device health status in 5 levels. The reasoning process takes less than 50 milliseconds, meeting the real-time monitoring requirements.

[0101] As shown in Figure 2 , a power distribution station equipment monitoring and health evaluation method based on multi-modal information is implemented through different units, including:

[0102] A multi-dimensional parameter real-time acquisition unit is connected to the sensing devices of different equipment in the power distribution station, used to obtain voltage, current, temperature, and other types of parameter data during equipment operation.

[0103] A parameter subset dynamic clustering processing unit is connected to the multi-dimensional parameter real-time acquisition unit, based on an improved fuzzy C-means clustering algorithm to cluster the collected parameter data.

[0104] A hypergraph topology structure construction unit is connected to the parameter subset dynamic clustering processing unit, taking the clustered parameter subsets as nodes to construct a hypergraph model containing multi-parameter coupling relationships.

[0105] A hypergraph neural network training and learning unit is connected to the hypergraph topology structure construction unit, using the hypergraph adjacency matrix and parameter feature data to train the hypergraph neural network.

[0106] A device health state reasoning and evaluation unit is connected to the hypergraph neural network training and learning unit, based on the trained model to process real-time parameters and generate health state quantitative results.

[0107] An evaluation result comprehensive analysis and output unit is connected to the device health state reasoning and evaluation unit, dividing health levels according to the quantitative results and forming an evaluation report. This unit maintains data interaction with the multi-dimensional parameter real-time acquisition unit and the device health state reasoning and evaluation unit.

[0108] A power distribution station equipment monitoring and health evaluation method based on multi-modal information, through an improved fuzzy C-means clustering algorithm, each parameter of the power distribution station equipment is implemented dynamic clustering, the fuzzy correlation characteristics between the parameters are accurately captured, the precision of the operation state similarity division is improved, and the parameter subset with potential correlation is accurately aggregated. This processing method solves the problem of insufficient division precision of the traditional clustering algorithm when facing multiple parameters, provides a more reliable data basis for subsequent equipment health evaluation work, and ensures that the evaluation process has high accuracy from the data source.

[0109] In terms of modeling multi-parameter correlation, the method constructs a hypergraph topology containing multi-parameter coupling relationship through a hypergraph neural network equipment monitoring evaluation model, breaking through the limitations of conventional neural networks. Through hyperedge convolution operation, high-order correlation between parameters can be captured, rather than simply handling the association of paired parameters. At the same time, the individual features and collaborative features of the parameters are fused to generate quantitative evaluation results, fully reflecting the overall characteristics of the device operation, overcoming the defects of conventional models that ignore the influence of multi-parameter collaboration on the health state, making the evaluation results more in line with the actual operation of the device.

[0110] In addition, the method strengthens the adaptability of the parameters and the model by constructing a model formula, and refines the key steps, further enhancing the depth and breadth of parameter correlation mining. This enables the method to accurately capture the characteristics of strong coupling of multiple parameters and non-linear state changes of the device, meeting the needs of device health evaluation under complex working conditions, significantly improving the comprehensiveness and accuracy of power distribution station equipment health management, and effectively making up for the deficiencies in the background technology.

[0111] In the description of the present application, it should be noted that unless otherwise specified and limited, the terms "arrangement", "installation", "connection", "connection", "fixing" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0112] Although embodiments of the present application have been shown and described, those skilled in the art can understand that various equivalent changes, modifications, replacements and variations of the embodiments can be made without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalent scope.

Claims

1. A power distribution station equipment monitoring and health assessment method based on multi-modal information, characterized in that, The method comprises the following steps: Step S1: Real-time collection of multi-dimensional parameter data including equipment operating voltage, load current, winding temperature, insulation resistance, mechanical vibration frequency, and gas leakage concentration through different types of sensing devices deployed in different areas of the power distribution station, to form an original monitoring data set containing time and space correlation characteristics; Step S2: Dynamic clustering processing of the original monitoring data set based on an improved fuzzy C-means clustering algorithm, determination of an optimal cluster number according to the nonlinear correlation characteristics of the equipment parameters, and division of equipment parameter subsets with similar operating characteristics; Step S3: Construction of the topology structure of the hypergraph neural network equipment monitoring evaluation model, taking the equipment parameter subsets as hypergraph nodes, fusing the multi-parameter coupling relationship by defining high-order correlation edges between nodes, and generating a hypergraph adjacency matrix; Step S4: Training of the hypergraph neural network using the hypergraph adjacency matrix, extraction of deep implicit features of the equipment parameters through multi-layer hyperedge convolution operations, and output of a feature vector containing equipment health state correlation information; Step S5: Feature mapping and state reasoning of the real-time input equipment parameter subsets based on the trained hypergraph neural network model, to generate a quantitative evaluation result representing the equipment health degree; Step S6: Division of equipment health state levels according to the quantitative evaluation result, and formation of an equipment operating state comprehensive evaluation report containing different parameter correlation influences.

2. The method of claim 1, wherein, In the step S2, the improved fuzzy C-means clustering algorithm performs dynamic clustering optimization by introducing equipment parameter weight coefficients, and defines a clustering objective function as: where u ij represents the membership degree of the ith device parameter sample belonging to the jth cluster, m is the fuzzy factor, x i is the ith device parameter sample vector, v j is the jth cluster center vector, λ is the regularization parameter, n is the total number of samples, c is the number of clusters, p is the device parameter dimension, ω k is the weight coefficient of the kth device parameter, including voltage, current, temperature, and satisfies In the iteration process, the membership degree matrix is updated to optimize clustering, and the weight coefficient ω k is dynamically adjusted according to the time-varying characteristics of the device parameters, so that the clustering results are more consistent with the actual operating state of the power distribution station equipment.

3. The method of claim 1, wherein, In the step S3, a time and space decay factor of the equipment parameters is introduced when constructing the hypergraph adjacency matrix, and a hypergraph edge weight calculation formula is defined as: where w e is the weight of hyper-edge e, v and u are the device parameter nodes connected by hyper-edge e, t v and t u are the collection timestamps of parameters v and u, respectively, a is the time decay coefficient, b vu is the physical correlation coefficient of parameters v and u, g(v, u) is the spatial distance weight of parameters v and u, and |e| is the number of nodes contained in hyper-edge e.

4. The method of claim 1, wherein, In the step S4, the hyperedge convolution layer operation of the hypergraph neural network is defined as: where H (l) and H (l+1) are the hypergraph node feature matrices of the I-th and I+1-th layers, respectively, E is the set of all hyperedges in the hypergraph, w e is the weight of the hyperedge e, deg(v) is the degree of the node v, deg(e) is the degree of the hyperedge e, is the weight matrix corresponding to the I-th layer hyperedge e, b (l) is the bias vector of the I-th layer, and σ is the activation function.

5. The method of claim 1, wherein, In the step S5, a quantitative evaluation function of the output layer of the hypergraph neural network is defined as: where S is the equipment health state quantitative evaluation result, θ k is the characteristic weight of the kth equipment parameter, f k (·) is the mapping function of the kth parameter feature, h k is the feature vector of the kth parameter output by the hypergraph neural network, μ e is the influence coefficient of the hyperedge e, g e (·) is the collaborative feature mapping function of the parameter associated with the hyperedge e, h e is the collaborative feature vector of the parameter associated with the hyperedge e.

6. The method of claim 1, wherein, In the connection process of the step S2 and the step S3, a mapping relationship function of the parameter subsets to the hypergraph nodes is defined as: N s =∑ i∈s η i x i +∑ i∈s ∑ j∈s，j≠i ξ ij (x i ⊙x j ) where N s is the hypergraph node feature corresponding to the parameter subset s, η i is the node mapping weight of the i-th parameter in the subset s, ξ ij is the interaction mapping coefficient of the i-th and j-th parameters in the subset s, and is the element product operation.

7. The method of claim 1, wherein, The step S3 comprises the following sub-steps: Step S3.1: Determination of a hypergraph node set based on the equipment parameter subsets obtained in the step S2, each node corresponding to a comprehensive feature of a parameter subset, preliminary screening of node combinations with potential correlations by calculating the feature similarity between different parameter subsets; Step S3.2: Definition of a hyperedge generation rule according to the physical connection relationship of the power distribution station equipment and the electrical characteristics of the parameters, inclusion of parameter subset nodes related to the same equipment component or having strong electrical coupling relationship into the same hyperedge, and formation of an initial hyperedge set; Step S3.3: Calculation of the parameter fluctuation covariance of the nodes included in each hyperedge, adjustment of the range of the hyperedge according to the covariance value, merging of hyperedges with consistent fluctuation trends, splitting of hyperedges containing nodes with too large fluctuation characteristics difference, and optimization of the hypergraph structure; Step S3.4: Calculation of the initial weight values of different hyperedges based on the number of nodes included in the hyperedge and the importance of the parameters, establishment of a complete hypergraph adjacency matrix containing node, hyperedge, and weight information, and provision of topology structure data for the training of the hypergraph neural network.

8. The method of claim 1, wherein, The step S4 comprises the following sub-steps: Step S4.1: input the supergraph adjacency matrix obtained in step S3 and the device parameter subset feature into the input layer of the supergraph neural network, fuse the node features and supergraph topology information through matrix transformation, and convert them into a feature matrix form suitable for superedge convolution operation; Step S4.2: in the hidden layer of the supergraph neural network, perform superedge convolution operation, weight and aggregate the node features through the superedge weight matrix, and introduce the physical constraint condition of the device parameter to adjust the convolution kernel parameter, so that the feature extraction process is consistent with the operation law of the power distribution station device; Step S4.3: after each layer of superedge convolution operation, perform nonlinear transformation on the output features, and enhance the model's ability to capture nonlinearly associated parameters through the activation function; Step S4.4: through the stacking of multiple layers of superedge convolution operation, gradually improve the abstract level of the features, and convert the shallow physical features of the device parameters into deep health state related features.

9. The method of claim 1, wherein, The step S5 includes the following sub-steps: Step S5.1: input the deep feature vector output in step S4 into the output layer of the supergraph neural network, map the high-dimensional features to the feature space related to the device health state through feature dimension transformation, so that the dimension of the feature vector matches the dimension of the health assessment index; Step S5.2: weight and fuse the mapped feature vectors, and the weight value is determined according to the fault sensitivity of the device parameter. The feature corresponding to the parameter with high fault sensitivity is given a higher weight, so as to highlight the influence of the calibration parameter on the health assessment result; Step S5.3: introduce the associated features between parameters in the supergraph, calculate the feature coordination degree of the nodes connected by the superedge, and include the coordination features in the evaluation calculation process to make up for the limitations of single parameter feature evaluation; Step S5.4: convert the fused feature vector into a quantitative evaluation result through a preset mapping rule. The result includes the health state information of different calibration components of the device, and forms the basic data directly used for state level division.

Images