Operation and maintenance debugging monitoring system suitable for electric field

Abstract

The application discloses a system suitable for operation and maintenance debugging monitoring of an electric field, relates to the technical field of operation and maintenance debugging monitoring, and comprises a collection unit, which collects operation data of electric field equipment at multidimensional monitoring points, wherein the operation data comprises electrical parameters, mechanical vibration parameters, environmental parameters and chemical gas component data, and a multimodal monitoring data set is constructed; and a feature analysis module, which performs space-time feature extraction on the multimodal monitoring data set, generates a multidimensional feature tensor containing time domain features, frequency domain features and spatial distribution features, and obtains the space-time coupling relationship between features through tensor decomposition technology. The application provides a full-process intelligent solution from monitoring, early warning to decision-making for electric field operation and maintenance, greatly reduces the equipment failure risk, shortens the power-off time, and improves the safety and economy of power grid operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of operation and maintenance debugging monitoring, and particularly relates to an operation and maintenance debugging monitoring system suitable for an electric field. BACKGROUND

[0002] Under the background of the intelligent development of the power system, the operating conditions of electric field equipment are becoming increasingly complex, and the uncertainty of the operating environment is significantly increasing. The existing electric field operation and maintenance debugging monitoring technology has weak fusion and analysis capabilities for multi-source heterogeneous data, and cannot deeply mine the potential correlation between data. Therefore, it is difficult to accurately deconstruct the spatio-temporal coupling relationship of the device operating state, resulting in a large error in the device health state evaluation, and the fault risk prediction lacks timeliness and accuracy. At the same time, the traditional monitoring technology cannot effectively combine the device topology structure for systematic analysis, and it is difficult to make scientific and reasonable operation and maintenance decisions, which cannot meet the actual needs of the safe and stable operation of the electric field equipment. SUMMARY

[0003] The purpose of the present application is to provide an operation and maintenance debugging monitoring system suitable for an electric field, which solves the technical problem that the prior art cannot effectively combine the device topology structure for systematic analysis, and it is difficult to make scientific and reasonable operation and maintenance decisions, which cannot meet the actual needs of the safe and stable operation of the electric field equipment.

[0004] The operation and maintenance debugging monitoring system suitable for an electric field comprises:

[0005] A collection unit collects operating data of electric field equipment at multiple dimensions of monitoring points, the operating data comprising electrical parameters, mechanical vibration parameters, environmental parameters and chemical gas component data, and a multi-modal monitoring data set is constructed;

[0006] A feature analysis module extracts spatio-temporal features from the multi-modal monitoring data set, generates a multi-dimensional feature tensor containing time domain features, frequency domain features and spatial distribution features, and obtains the spatio-temporal coupling relationship between features through tensor decomposition technology;

[0007] A calculation module constructs a device health state space based on the spatio-temporal coupling relationship, calculates the density distribution of historical fault cases in the state space, and determines a fault warning threshold surface of key parameters;

[0008] A risk assessment module maps real-time monitoring data to the device health state space, measures the deviation degree of a real-time state point from the warning threshold surface through Mahalanobis distance, and generates a fault risk heat map;

[0009] A decision module combines the fault risk heat map and the device topology structure, uses a graph neural network to predict a fault propagation path, and generates an operation and maintenance decision matrix containing inspection priority and maintenance window period;

[0010] The verification unit verifies the validity of the operation and maintenance decision matrix through digital twinning technology, and dynamically adjusts the parameter weight of the spatio-temporal feature extraction algorithm according to the verification result.

[0011] As a further scheme of the present application: the spatio-temporal feature extraction on the multi-modal monitoring data set is specifically implemented in the following manner:

[0012] A hybrid neural network model composed of a two-dimensional convolutional neural network layer, a long short-term memory network layer and a full connection layer is built.

[0013] The two-dimensional convolutional neural network layer is provided with a plurality of convolutional layers and pooling layers with different convolution kernel sizes, for extracting data spatial features.

[0014] The long short-term memory network layer utilizes the gating mechanism to process the time sequence features of the data.

[0015] Finally, the spatio-temporal features are fused through the full connection layer to output a multi-dimensional feature tensor containing time domain features, frequency domain features and spatial distribution features.

[0016] As a further scheme of the present application: the device health state space is constructed based on the spatio-temporal coupling relationship, specifically:

[0017] A high-dimensional state space is constructed by using Riemannian manifold;

[0018] The multi-dimensional feature tensor is first subjected to whitening processing to eliminate the correlation between the features;

[0019] Then, the processed data is mapped to the Riemannian manifold space through exponential mapping;

[0020] In the Riemannian manifold space, the probability density distribution of historical failure cases is calculated by using kernel density estimation method, and then the failure warning threshold surface of the key parameters is determined.

[0021] As a further scheme of the present application: the graph neural network is used to predict the failure propagation path by combining the failure risk thermodynamic map and the device topology structure, specifically:

[0022] A graph neural network composed of a graph convolutional layer, a graph attention layer and an output layer is constructed;

[0023] The graph convolutional layer is used to aggregate the neighborhood information of the device nodes to realize the update of the node features;

[0024] The graph attention layer is used to calculate the attention coefficients between the nodes to enhance the recognition ability of the key nodes in the device topology structure;

[0025] Finally, the output layer predicts the propagation path of the failure in the device topology structure based on the updated node features, and outputs the path probability distribution.

[0026] As a further scheme of the present application: the deviation degree of the real-time state point from the early warning threshold surface is measured by the Mahalanobis distance, and a fault risk heat map is generated, specifically:

[0027] First, the real-time monitoring data is standardized to have zero mean and unit variance;

[0028] The local weighted regression method is used to locally fit the early warning threshold surface;

[0029] The Mahalanobis distance from the real-time state point to the fitted surface is calculated, and the Mahalanobis distance calculation result is smoothed by combining the Gaussian kernel density estimation;

[0030] Taking the geographical position of the equipment as the coordinate reference, a fault risk heat map is finally generated.

[0031] As a further scheme of the present application: the effectiveness of the operation and maintenance decision matrix is verified by digital twinning technology, specifically:

[0032] A digital twin containing equipment geometry, material properties, operation data and environmental impact is established;

[0033] A three-dimensional modeling software is used to construct a geometric model of the equipment, and a material attribute model is formed by setting parameters according to the material properties;

[0034] The operation data model is established by synchronizing the operation data of the equipment, and the environmental impact model is constructed by simulating environmental factors;

[0035] The operation and maintenance decision matrix is simulated in the digital twin, and at least the operation state and fault probability results are collected.

[0036] As a further scheme of the present application: the electrical parameters include at least three of partial discharge quantity, insulation resistance, dielectric loss factor, leakage current, dynamic capacitance and capacitance change rate, ultra-high frequency discharge signal and pulse current signal, the mechanical vibration parameters include acceleration amplitude, vibration frequency distribution and vibration phase characteristics, the environmental parameters include temperature, humidity, air pressure and wind speed, and the chemical gas component data includes SF6 decomposition product concentration and dissolved gas content in transformer oil.

[0037] As a further scheme of the present application: when the processed data is mapped to the Riemannian manifold space by exponential mapping, it further includes:

[0038] The geodesic distance between the current data point and the origin of the Riemannian manifold space is calculated;

[0039] The scale parameter of the exponential mapping is dynamically adjusted according to the geodesic distance, and when the distance is greater than a preset threshold, a contraction type exponential mapping is used to suppress the influence of outliers;

[0040] When the distance is less than the preset threshold, an expanding exponential mapping is used to enhance the discrimination of normal state data;

[0041] The geometric relationship between data points is kept unchanged through parallel movement operation on the manifold.

[0042] As a further scheme of the application: when the fault early warning threshold surface of the key parameter is determined, a topological data analysis method is used to extract the fault feature topological structure in the Riemannian manifold space, specifically:

[0043] The Vietoris-Rips complex of historical fault cases is constructed to capture the hole and connected branch features in the manifold space.

[0044] The persistent homology groups under different scales are calculated to identify the fault features with topological persistence.

[0045] The persistent homology features are associated and mapped with the physical fault modes of the equipment to determine the topological critical points of the key parameters.

[0046] The threshold surface with geometric invariance is generated in the Riemannian manifold space through the topological critical points.

[0047] As a further scheme of the application: when the effectiveness of the operation and maintenance decision matrix is verified by the digital twin technology, the method further comprises:

[0048] A digital twin model consistent with the parameters of the physical electric field equipment is constructed, and the digital twin model comprises an equipment electrical parameter model, a mechanical structure model and an environmental influence model.

[0049] The debugging strategy parameters in the operation and maintenance decision matrix are input into the digital twin model to simulate a predicted state data set.

[0050] The geodesic distance of the predicted state data set and the physical equipment measured data in the high-dimensional state space is calculated as a quantitative index of decision effectiveness.

[0051] When the geodesic distance exceeds a preset threshold, the density distribution of the historical fault cases in the high-dimensional state space is recalculated, and the fault early warning threshold surface of the key parameter is adjusted.

[0052] By comparing the topological feature changes in the predicted state data set with the topological features of the historical fault cases, a verification report containing fault mode matching degrees is generated, and the parameter weights of the spatiotemporal feature extraction algorithm are adjusted based on the verification report.

[0053] Compared with the prior art, the application has the following beneficial effects:

[0054] The application solves the core pain points of the traditional electric field monitoring system through the technical path of "multi-modal data fusion-temporal and spatial feature mining-risk dynamic assessment-intelligent decision optimization". The combination of multi-modal data acquisition and tensor decomposition technology significantly enhances the early fault warning capability; the fault propagation prediction model based on the graph neural network improves the inspection path planning efficiency and reduces unnecessary manual inspection; the digital twin verification mechanism ensures the reliability of the operation and maintenance decision, and improves the prediction accuracy during the maintenance window period. Thus, the application provides an intelligent solution for the whole process from monitoring, early warning to decision-making for the electric field operation and maintenance, greatly reduces the equipment failure risk, shortens the power outage time, and improves the safety and economy of the power grid operation. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 FIG. 1 is a schematic diagram of the system framework structure of the application. DETAILED DESCRIPTION

[0056] The technical solutions of the application will be described below in conjunction with the embodiments. Obviously, the described embodiments are only a part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0057] In the field of electric field operation and maintenance debugging monitoring, the traditional system faces problems such as difficulty in multi-dimensional data fusion, lag in fault early warning, lack of overall optimization in operation and maintenance decision, etc. On the one hand, the running state of electric field equipment (such as transformers, circuit breakers, etc.) is affected by multi-dimensional factors such as electrical parameters, mechanical vibration, environmental factors, and chemical gas components, but the traditional monitoring system can only analyze single type data in isolation, and it is difficult to capture the spatio-temporal coupling relationship between different parameters. For example, in the early stage of transformer partial discharge, the electrical parameter changes slightly, but the chemical gas component (such as hydrogen and acetylene concentration) may have abnormally increased, and the traditional system cannot associate these cross-dimensional information, resulting in failure of early fault warning. On the other hand, when the equipment fails, the traditional system cannot accurately predict the fault propagation path, often leading to blind inspection by the operation and maintenance personnel, missing the best maintenance window period, increasing the power outage time and maintenance cost.

[0058] Referring to Figure 1 The application provides an operation and maintenance debugging monitoring system suitable for electric field, which comprises:

[0059] The acquisition unit acquires the running data of the electric field equipment at the multi-dimensional monitoring points, and the running data includes electrical parameters, mechanical vibration parameters, environmental parameters, and chemical gas component data, and constructs a multi-modal monitoring data set;

[0060] a feature analysis module, which extracts spatio-temporal features from the multi-modal monitoring dataset to generate a multi-dimensional feature tensor containing time-domain features, frequency-domain features and spatial distribution features, and obtains spatio-temporal coupling relationships between the features by tensor decomposition technology;

[0061] The acquisition unit is in communication connection with the feature analysis module and is configured to transmit the multi-modal monitoring dataset to the feature analysis module.

[0062] The calculation module constructs a device health state space based on the spatio-temporal coupling relationships, calculates the density distribution of historical fault cases in the state space, and determines a fault warning threshold surface of the key parameters;

[0063] The risk assessment module maps real-time monitoring data to the device health state space, measures the deviation degree of a real-time state point from the warning threshold surface by Mahalanobis distance, and generates a fault risk heat map;

[0064] The decision module combines the fault risk heat map and the device topology structure, uses a graph neural network to predict a fault propagation path, and generates an operation and maintenance decision matrix containing an inspection priority and a maintenance window period.

[0065] The verification unit verifies the effectiveness of the operation and maintenance decision matrix by digital twinning technology and dynamically adjusts the parameter weights of the spatio-temporal feature extraction algorithm according to the verification result.

[0066] As an optional embodiment, the electrical parameters include at least three of partial discharge quantity, insulation resistance, dielectric loss factor, leakage current, dynamic capacitance and capacitance change rate, ultra-high frequency discharge signal and pulse current signal, the mechanical vibration parameters include acceleration amplitude, vibration frequency distribution and vibration phase characteristics, the environmental parameters include temperature, humidity, air pressure and wind speed, and the chemical gas component data includes SF6 decomposition product concentration and dissolved gas content in transformer oil.

[0067] Specifically, the acquisition unit is configured to construct the multi-modal monitoring dataset, which comprehensively covers electrical parameters (such as voltage and current) of the electric field device, mechanical vibration parameters (such as vibration frequency and amplitude), environmental parameters (such as temperature and humidity) and chemical gas component data (such as SF6 decomposition product concentration).

[0068] The feature analysis module extracts spatio-temporal features from the dataset to generate a multi-dimensional feature tensor, analyzes the spatial distribution characteristics of data from different monitoring points in combination with the spatial layout of the device, and uses tensor decomposition technology (such as CANDECOMP / PARAFAC decomposition) to reveal the spatio-temporal coupling relationships between fluctuations in electrical parameters and mechanical vibrations and changes in environmental temperature, for example, it is found that an increase in temperature will intensify mechanical vibrations at certain frequencies, which in turn affects the stability of electrical connections.

[0069] Among them, the device health state space takes historical failure cases as samples, and determines the failure warning threshold surface of the key parameters through density distribution analysis. For example, the transformer insulation breakdown cases in history are analyzed to form a high-risk area in the state space, and a warning is triggered when the real-time monitoring data is mapped near the area. The risk assessment module measures the deviation degree of the real-time state point from the warning threshold surface by using Mahalanobis distance, and generates a failure risk heat map. Compared with Euclidean distance, Mahalanobis distance considers the correlation between parameters, and can more accurately reflect the actual risk of the device. For example, when voltage fluctuations, abnormal increase of vibration frequency and synchronous temperature rise are monitored, the Mahalanobis distance will increase significantly, and the corresponding area in the heat map is displayed as a high-risk red area;

[0070] Among them, taking the substation device network as an example, the nodes represent the devices, and the edges represent the electrical connection relationship. By training the graph neural network model, the propagation rule of the failure in the network is learned. When a circuit breaker appears a high-risk warning, the model can predict that the failure may propagate along the bus to the adjacent transformer or other load devices, and generate an operation and maintenance decision matrix containing inspection priority (such as preferentially checking the elements directly connected to the failure device) and maintenance window period (such as predicting that the failure probability exceeds 80% in the next 24 hours). The verification unit simulates the execution of the operation and maintenance decision in the virtual environment by using digital twinning technology, and evaluates its effectiveness, thereby providing a full-process intelligent solution from monitoring, early warning to decision-making for the operation and maintenance of the electric field, which is beneficial to greatly reducing the device failure risk, shortening the power outage time, and improving the safety and economy of the power grid operation.

[0071] As an optional embodiment, spatio-temporal feature extraction is performed on the multi-modal monitoring data set, specifically in the following manner:

[0072] A hybrid neural network model composed of a two-dimensional convolutional neural network layer, a long short-term memory network layer and a fully connected layer is built;

[0073] Among them, the hybrid neural network model combines the advantages of convolutional neural network in spatial feature extraction and recurrent neural network in time series processing, and realizes deep fusion of features through the fully connected layer, providing efficient network structure support for multi-dimensional feature tensor generation.

[0074] The two-dimensional convolutional neural network layer is provided with a plurality of convolution layers and pooling layers with different convolution kernel sizes for extracting data spatial features;

[0075] Different sizes of convolution kernels (e.g., 3x3, 5x5) can capture different ranges of spatial correlation information, for example, small size convolution kernels are suitable for extracting signal details of local monitoring points (e.g., vibration peak value of a certain sensor), and large size convolution kernels can capture the spatial distribution law of multiple adjacent monitoring points (e.g., temperature gradient of different parts of a transformer). The pooling layer reduces the data dimension through downsampling operation, retains key features, and avoids overfitting. For example, after the convolution operation on the temperature monitoring data of each area of the substation, the spatial aggregation mode of the temperature can be extracted, and the regional distribution characteristics of the poor heat dissipation of the equipment are identified.

[0076] The gating mechanism of the long short-term memory network layer is used to process the time sequence characteristics of the data.

[0077] The gating mechanism includes an input gate, a forget gate, and an output gate, which helps to solve the gradient vanishing or explosion problem of traditional recurrent neural networks and is suitable for capturing dynamic change rules in long time series data. For example, for voltage fluctuation data of electrical equipment, the long short-term memory network can remember voltage abnormal fluctuations hours ago and establish a correlation with the current vibration signal changes to identify the cumulative impact of long-term voltage instability on the mechanical structure of the equipment. Through the gating mechanism, the network can automatically ignore irrelevant short-term noise and focus on long-term trend features that have a significant impact on the health status of the equipment.

[0078] Finally, the spatial and temporal features are fused through a fully connected layer to output a multi-dimensional feature tensor containing time domain features, frequency domain features, and spatial distribution features.

[0079] The fully connected layer non-linearly combines the spatial features extracted by the convolutional neural network and the temporal features extracted by the long short-term memory network to explore the deep association between them. For example, the vibration frequency domain features of the transformer (energy proportion of a certain specific frequency) are fused with the temperature change features over time (temperature fluctuation amplitude caused by diurnal temperature difference) to form comprehensive features that reflect the thermal-mechanical coupling effect of the equipment. The output multi-dimensional feature tensor provides structured feature data for subsequent tensor decomposition and spatio-temporal coupling relationship analysis, ensuring that the association between features is completely preserved.

[0080] Specifically, the hybrid neural network model breaks through the limitations of a single network structure, improving the extraction accuracy of spatial and temporal features. The combination of multi-size convolution kernels and the gating mechanism enables precise capture of multi-scale and long time series features, which helps to avoid feature loss or redundancy. The fusion processing of the fully connected layer establishes a deep association between spatial and temporal features, providing a high-quality feature basis for subsequent equipment health status evaluation and fault warning. This helps to improve the utilization rate of multi-modal data features, enhances the perception ability of the running state of electrical equipment through the cooperation of tensor decomposition technology, and provides a strong guarantee for the accuracy of operation and maintenance decisions.

[0081] As an optional embodiment, a device health state space is constructed based on a space-time coupling relationship, specifically:

[0082] A Riemannian manifold is used to construct a high-dimensional state space;

[0083] First, the multi-dimensional feature tensor is subjected to whitening processing to eliminate the correlation between the features;

[0084] Then, the processed data is mapped to a Riemannian manifold space through exponential mapping;

[0085] In the Riemannian manifold space, the probability density distribution of historical failure cases is calculated using the kernel density estimation method, and then the failure warning threshold surface of the key parameters is determined.

[0086] In the operation and debugging monitoring of the electric field, on the one hand, the multi-dimensional features of the device operation have complex correlations, such as the mutual influence of electrical parameters, temperature, vibration and other features, and the traditional Euclidean space cannot accurately describe this nonlinear relationship, which easily leads to ambiguous state division; on the other hand, the distribution of historical failure cases in the high-dimensional feature space is non-uniform, and the traditional threshold setting method (such as a single threshold or a hyperplane) cannot adapt to this distribution characteristic, often resulting in false alarms or missed alarms, affecting the accuracy of the judgment of the device health state. These problems make the device health state evaluation lack reliable spatial model support, restricting the effectiveness of failure warning;

[0087] Among them, Riemannian manifold is selected as the carrier of the device health state space. Compared with the traditional Euclidean space, Riemannian manifold can better depict the nonlinear correlation and geometric structure between high-dimensional features. The running features of electric field devices (such as voltage fluctuation, vibration frequency, gas concentration, etc.) present a complex curved distribution in high-dimensional space. Riemannian manifold can adapt to this distribution through its inherent geometric properties, which is beneficial to avoid state description deviation caused by unreasonable space assumptions. For example, the combination of health features of a transformer under different loads and environments can form a more realistic state region division in the Riemannian manifold, providing a more accurate spatial basis for subsequent failure probability calculation;

[0088] Among them, in the multi-dimensional feature tensor, the features of different monitoring parameters often have coupling, such as temperature rise may cause resistance change and vibration enhancement at the same time. This correlation will interfere with the construction of the state space. Whitening processing converts the features into uncorrelated form through linear transformation, normalizes the variance of each feature, for example, the correlation coefficient of the temperature feature and the vibration feature is reduced to below 0.1 after processing from the original 0.8. This process simplifies the relationship between the features, ensures that the data mapped to the Riemannian manifold space can more truly reflect the independent contribution of each feature, and improves the discrimination of the state space;

[0089] The exponential mapping is an important tool in Riemannian geometry, which can convert data points in a linear space to the curved structure of a manifold space, preserving the local geometric characteristics of the data. For example, after the whitening processing of the transformer feature data (including voltage, temperature, vibration, etc.), the point set corresponding to the actual health status of the equipment is formed on the Riemannian manifold after the exponential mapping. The points in the normal operating state are concentrated in a certain area of the manifold, while the points in the fault state are distributed in another area, and the point sets of different types of faults have obvious geometric separation.

[0090] The density estimation assigns a kernel function to each historical fault point, fitting the fault probability distribution of the entire space, and can accurately capture the clustering characteristics of fault cases on the manifold. For example, for transformer insulation aging faults, historical cases show a specific high-density distribution area on the Riemannian manifold, and the kernel density estimation can clearly outline the boundary of this area. Based on this distribution, the fault warning threshold surface of the key parameters is determined, which can adapt to the fault probability contour. When the real-time monitoring data mapped to the point on the manifold enters the inside of the surface (high probability area), the fault warning is triggered, effectively avoiding the problem that the traditional flat threshold cannot adapt to non-uniform distribution, thereby providing a precise spatial model for equipment health status evaluation. The feature analysis module and the risk assessment module cooperate to improve the reliability of the electric field equipment fault warning and provide a solid basis for operation and maintenance decision-making.

[0091] As an optional embodiment, when the processed data is mapped to the Riemannian manifold space through the exponential mapping, it further includes:

[0092] calculating the geodesic distance between the current data point and the origin of the Riemannian manifold space;

[0093] dynamically adjusting the scale parameter of the exponential mapping according to the geodesic distance, and when the distance is greater than a preset threshold, using a contraction type exponential mapping to suppress the influence of outliers;

[0094] When the distance is less than the preset threshold, an expansion type exponential mapping is used to enhance the discrimination of normal state data;

[0095] The geometric relationship between the data points is maintained by parallel movement operation on the manifold.

[0096] Specifically, in the electric field operation and maintenance debugging monitoring system, when the data is mapped to the Riemannian manifold space, on the one hand, there may be outliers in the data (such as extreme values generated by abnormal fluctuations of sensors), and if these points are directly mapped to the manifold space, the distribution structure of normal state data will be distorted, resulting in the deviation of the fault warning threshold surface; on the other hand, the subtle data differences in the normal state (such as the difference between the characteristics of slight aging and normal operation of the equipment) may be compressed in the manifold space, making it difficult to form an effective distinction and affecting the accuracy of the health state evaluation. In addition, the geometric relationship between data points in the traditional mapping process is easily affected by the change of scale, resulting in distortion of the relative position of the originally related state points on the manifold, which cannot accurately reflect the relevance of the actual operation of the equipment.

[0097] wherein the geodesic distance is the shortest path length between two points on the Riemannian manifold, which is different from the straight line distance in Euclidean space, and it can more accurately describe the actual distance between data points in the curved structure of the manifold. For example, in the health state space of the electric field equipment, the origin can be set as the characteristic point of the ideal operating state of the equipment, and the geodesic distance between the real-time monitoring data point of a transformer and the origin is 5.2 (unit: manifold space measurement unit), which directly reflects the deviation degree of the current state of the equipment from the ideal state, providing a quantitative basis for subsequent mapping scale adjustment;

[0098] wherein when the distance is greater than a preset threshold (such as a threshold of 8.0), the data point is determined to be an outlier, and a shrinking exponential mapping is used. The shrinking mapping compresses the outlier point towards the origin of the manifold by reducing the scale parameter, reducing its interference with the overall distribution. For example, an extreme vibration data point generated by a sensor fault has a geodesic distance of 12.0, which is reduced to 9.5 after shrinking mapping, avoiding distortion of the clustering structure of normal state data. When the distance is less than the preset threshold, an expanding exponential mapping is used, which expands the distance between normal state data points by amplifying the scale parameter, enhancing the distinction of subtle differences. For example, the geodesic distance between temperature characteristic points under different loads during normal operation of the equipment is only 1.0, which is increased to 1.5 after expansion, so that the "light load normal" and "heavy load slight overheating" states are clearly separated on the manifold;

[0099] Among them, parallel movement is a transformation method in Riemann geometry that keeps the direction of vectors unchanged. In the data mapping process, it ensures that the relative positions and angle relationships between data points are consistent with the original space. For example, in the topology of electric field equipment, the state points of three adjacent devices A, B, and C form a specific triangular relationship in the original space. After parallel movement, this relationship is completely preserved on the Riemann manifold, which helps to avoid geometric distortion caused by mapping scale adjustment. This operation ensures that the correlation between device states (such as the causal relationship between A device failure and B device abnormality) is not destroyed in the manifold space, providing a reliable geometric basis for subsequent kernel density estimation and fault probability distribution calculation. This provides key support for the accurate construction of Riemann manifold space, enabling the device health state space to more accurately reflect the operating characteristics of the electric field equipment, laying the foundation for the accuracy of subsequent fault warning threshold surface determination and risk assessment, and significantly improving the adaptability of the electric field operation and maintenance monitoring system to complex data distribution.

[0100] As an optional embodiment, when determining the fault warning threshold surface of the key parameter, a topology data analysis method is used to extract the fault feature topology structure in the Riemann manifold space, specifically:

[0101] Constructing a Vietoris-Rips complex of historical fault cases to capture the hole and connected branch features in the manifold space;

[0102] Calculating persistent homology groups at different scales to identify fault features with topological persistence;

[0103] Associating the persistent homology features with the physical fault modes of the device to determine the topological critical points of the key parameter;

[0104] Generating a threshold surface with geometric invariance in the Riemann manifold space through the topological critical points.

[0105] Specifically, in the electric field operation and maintenance monitoring, when determining the fault warning threshold surface of the key parameter, on the one hand, relying only on statistical methods or empirical values to set the threshold value makes it difficult to capture the complex topology structure of fault features in high-dimensional space. For example, different fault modes of the device may form nested feature regions on the Riemann manifold, and traditional threshold surfaces cannot accurately outline their boundaries. On the other hand, fault features may exhibit dynamic topological changes with changes in device operating state and environment, and traditional methods lack the ability to identify such topologically persistent features, resulting in poor adaptability of the threshold surface and leading to warning delays or false alarms, affecting timely judgment and handling of device faults;

[0106] Among them, the Vietoris-Rips complex is an important tool in topological data analysis, which connects historical failure case data points into polyhedral structures by setting different distance parameters. For example, in the Riemannian manifold health state space of electrical field equipment, when the distance parameter is set to 2 for the insulation failure case data points of the transformer, a complex structure composed of multiple triangles can be formed, in which the holes may correspond to a certain characteristic blank area under a certain insulation failure mode, and the connected branches represent different types of insulation failure groups. In this way, abstract failure data can be converted into topological structures with geometric meaning, and the distribution pattern of failure characteristics in the manifold space can be intuitively presented;

[0107] Persistent homology groups are used to measure the existence time of topological features (such as points, edges, holes, etc.) at different distance parameter scales. The longer the topological feature exists, the stronger its topological persistence, and the more it represents the essential failure characteristics. For example, when analyzing circuit breaker failure cases, as the distance parameter gradually increases from 1 to 10, some hole features persist at multiple scales, while some temporary edge features quickly disappear. Those persistent hole features are the key failure characteristics with topological persistence. This process can filter out non-essential features caused by noise or accidental factors, and focus on the core topological structure that is important for failure warning;

[0108] Among them, the persistent homology feature is an abstract topological description, which needs to be combined with the actual device physical failure mode to have practical significance. For example, by correlating the identified hole feature with strong persistence with the specific physical failure modes such as excessive dissolved gas concentration in transformer oil and excessive local discharge intensity, it is found that this hole feature corresponds to the typical characteristics of transformer insulation aging failure. Further, the topological critical points of key parameters (such as gas concentration and discharge intensity) are determined under this correlation. These critical points are key numerical points for the transition of failure characteristics from one topological state to another, such as when a certain gas concentration reaches the critical point, the topological structure of the failure characteristics changes significantly;

[0109] Among them, the topological critical point has a stable location attribute in the manifold space, and the threshold surface generated based on these points will not change dramatically due to slight deformation of the manifold or small amplitude fluctuations of the data, i.e. it has geometric invariance. For example, after determining the multiple topological critical points of transformer insulation failure, the threshold surface formed by connecting these points can stably enclose the characteristic area of insulation failure in the Riemannian manifold space. No matter what small range of state fluctuations occur in the normal operation of the device, this threshold surface can accurately define the failure area and ensure the stability and reliability of the failure warning;

[0110] Specifically, by constructing the Vietoris-Rips complex and calculating the persistent homology group, the limitations of traditional methods in capturing complex fault features in high-dimensional space are overcome, the accuracy of fault feature recognition is improved, the association mapping of persistent homology features and physical fault modes enables the conversion of abstract topological features into critical points with practical physical significance, and the problem of disconnection between topological features and actual faults is solved. The threshold surface generated has geometric invariance, which helps to avoid the deviation of early warning caused by data fluctuations or manifold deformation, and improves the stability of fault early warning. This scheme provides a precise and stable threshold standard for electric field equipment fault early warning, and cooperates with the Riemannian manifold state space construction, risk assessment and other modules to significantly improve the prediction ability and reliability of the electric field operation and debugging monitoring system for equipment faults.

[0111] As an optional embodiment, the fault risk thermodynamic map is combined with the device topology structure, and a graph neural network is used to predict the fault propagation path, specifically:

[0112] A graph neural network including a graph convolution layer, a graph attention layer and an output layer is constructed;

[0113] The graph convolution layer is used to aggregate the neighborhood information of the device nodes to realize node feature updating;

[0114] The graph attention layer is used to calculate the attention coefficient between nodes to enhance the identification ability of key nodes in the device topology structure;

[0115] Finally, the output layer predicts the propagation path of the fault in the device topology structure based on the updated node features, and outputs the path probability distribution.

[0116] Specifically, in the electric field operation and debugging monitoring, on the one hand, the electric field equipment forms a mutually related system through a complex topology structure (such as bus connection and cable laying), and a single device fault may spread to other devices along the topology link, while the traditional prediction method (such as rule-based reasoning) is difficult to capture this nonlinear and dynamic propagation rule, on the other hand, the fault propagation influence of key nodes (such as hub substations and main transformers) in the device topology is much greater than that of ordinary nodes, but the traditional method has insufficient ability to distinguish the importance of nodes, resulting in a lack of priority in the predicted fault path, which increases the probability of invalid inspection. These problems make it difficult for operation and maintenance personnel to accurately locate the fault propagation direction, delay the best maintenance opportunity, and even expand the fault impact range;

[0117] The graph neural network takes the device as a node and the electrical connection relationship as an edge, and converts the device topology into graph data, which can naturally fit the association relationship between devices. Compared with the traditional neural network, the graph neural network does not need to rely on the grid structure of the data, and can directly process irregular device topology connections, for example, it can effectively model the star connection or ring connection structure of "transformer-bus-circuit breaker" in the substation, and provide an adaptive network architecture for fault propagation path prediction.

[0118] The convolution layer captures the information interaction within the local topology range by performing weighted summation on the features of each device node and its directly connected neighbor nodes. For example, when a disconnector has a fault risk, the graph convolution layer will fuse its features with the features of the adjacent bus and mutual inductor, update the feature vectors of these nodes, and reflect the preliminary impact of the fault transmission. This aggregation mechanism simulates the physical propagation process of the fault between devices, for example, an abnormal current will affect the voltage or temperature state of the adjacent device through electrical connection, and the graph convolution layer quantifies this influence as a change in node features, providing a basis for subsequent path prediction.

[0119] The attention coefficient reflects the influence weight of the neighbor nodes on the current node. Key nodes (such as main transformers) have more connections with other nodes due to their central position in the topology, and their attention coefficients are significantly higher than those of ordinary nodes. For example, in the substation topology, the main transformer is connected to multiple feeders and circuit breakers. When the graph attention layer calculates the attention coefficients of the main transformer and its neighbor nodes, the weight of the main transformer may reach 0.6 (the average weight of ordinary nodes is 0.2), making the network pay more attention to the state change of the main transformer when predicting fault propagation. This mechanism solves the problem of insufficient distinction of node importance in traditional methods, ensuring that the predicted fault path focuses on the key propagation link.

[0120] The output layer converts the node features into the probability of fault propagation to each neighbor node through the softmax function. For example, when a fault in a circuit breaker is detected, the output layer may predict that the probability of fault propagation to the bus is 70%, the probability of propagation to the adjacent load switch is 20%, and the probability of propagation to other devices is 10%. This probability distribution provides a quantitative basis for operation and maintenance decisions, such as prioritizing the inspection of the bus with the highest propagation probability and allocating corresponding inspection resources to different paths. The output fault propagation path clearly shows the link from the fault source to the potentially affected devices, such as "circuit breaker A → bus B → transformer C", which helps operation and maintenance personnel quickly lock the risk range, accurately predicts the diffusion path of the fault in the electric field device topology, focuses on high-risk links for operation and maintenance personnel, optimizes the allocation of inspection resources, and significantly shortens the fault location time and reduces the risk of fault expansion, providing strong support for the scientificity and timeliness of electric field operation and maintenance decisions.

[0121] As an optional embodiment, the deviation degree of the real-time state point from the early warning threshold surface is measured by Mahalanobis distance, and a fault risk heat map is generated, specifically:

[0122] Firstly, the real-time monitoring data is standardized to have zero mean and unit variance;

[0123] The local weighted regression method is used to locally fit the early warning threshold surface;

[0124] The Mahalanobis distance of the real-time state point to the fitted surface is calculated, and the Mahalanobis distance calculation result is smoothed by combining the Gaussian kernel density estimation;

[0125] Taking the geographical position of the equipment as the coordinate reference, a fault risk heat map is finally generated.

[0126] Specifically, in the operation and debugging monitoring of the electric field, firstly, the real-time monitoring data is directly used for distance calculation, which will lead to deviation evaluation distortion due to the difference in dimension and uneven distribution, and secondly, the early warning threshold surface is mostly globally fitted, which is difficult to adapt to the local nonlinear characteristics of the equipment health state space, resulting in insufficient risk assessment accuracy in some areas;

[0127] The real-time monitoring data comes from different sensors, and the numerical range of electrical parameters (such as voltage kV) and environmental parameters (such as humidity %) is significantly different, and direct calculation will lead to large numerical parameter dominating the distance result. Standardization processing eliminates the dimensional effect through the formula (wherein μ is the mean, and σ is the standard deviation) eliminates the dimensional effect, for example, the temperature data (mean 25℃, standard deviation 5℃) and vibration data (mean 0.1mm / s, standard deviation 0.02mm / s) in a certain area are uniformly converted into standardized data with zero mean and unit variance. This process ensures that each parameter has balanced weight in distance calculation, which helps to avoid deviation evaluation deviation caused by dimensional difference, and provides a reliable data basis for subsequent Mahalanobis distance calculation;

[0128] Among them, the threshold surface of global fitting is difficult to adapt to the complex nonlinear distribution in the Riemannian manifold space, for example, in the equipment health state space, the fault feature distribution of the transformer area and the circuit breaker area may present different curvatures. Local weighted regression assigns weights to each to be evaluated point (nearby point weight is high, far point weight is low), constructs a local polynomial model, for example, in the vicinity of the transformer state point, only 50 historical fault points around it are used to fit the surface, so that the surface is more fitted to the local characteristics.

[0129] wherein the Mahalanobis distance considers the correlation between parameters, such as the positive correlation between temperature and resistance, and can more truly reflect the degree of state deviation than the Euclidean distance. For example, the Mahalanobis distance of a certain real-time state point to the threshold surface is 2.5, indicating that the degree of deviation from the normal range is 2.5 times the average level. Since real-time data may contain noise, Gaussian kernel density estimation smoothes local fluctuations by weighted average of distance values (such as a bandwidth of 0.3), making the risk value transition more continuous. For example, the distance value originally fluctuating between 2.3-2.7 is smoothed to 2.5±0.1, avoiding risk misjudgment caused by noise;

[0130] wherein the risk value corresponding to the Mahalanobis distance is mapped to a color gradient (such as red representing high risk and blue representing low risk) and superimposed on the electrical field device topology map. For example, a certain transformer area in the substation is displayed in deep red (risk value 0.85), and its adjacent bus area is yellow (risk value 0.6), intuitively presenting the spatial distribution of risk. The operation and maintenance personnel can quickly identify the high-risk link of "transformer A-bus B" through the heat map and prioritize inspection, solving the poor readability problem of traditional numerical list and improving the risk positioning efficiency. Thus, a precise and readable risk distribution map is provided for electrical field operation and maintenance, which cooperates with the graph neural network fault propagation prediction to improve the identification efficiency of high-risk areas and the rationality of maintenance resource allocation, and is conducive to significantly enhancing the risk perception and decision support capability of the electrical field operation and maintenance monitoring system.

[0131] As an optional embodiment, the effectiveness of the operation and maintenance decision matrix is verified by digital twinning technology, specifically:

[0132] A digital twin is established, which includes device geometry, material properties, operation data, and environmental impact;

[0133] A device geometry model is constructed using three-dimensional modeling software, and a material property model is formed by setting parameters according to material characteristics;

[0134] An operation data model is established by synchronizing device operation data, and an environmental impact model is constructed by simulating environmental factors;

[0135] The operation and maintenance decision matrix is simulated in the digital twin, and at least the operation state and failure probability results are collected.

[0136] Specifically, in the electrical field operation and maintenance monitoring, on the one hand, directly executing unverified decisions on physical devices poses a safety risk; on the other hand, relying solely on historical data or simulation models for offline verification cannot reflect the current actual state of the device and the impact of complex environments. These problems make it difficult to ensure the effectiveness of operation and maintenance decisions, and may result in "over-maintenance" or "insufficient maintenance", increasing operation and maintenance costs and reducing device reliability;

[0137] Wherein, the digital twin is established to include device geometry, material properties, operation data and environmental impact, and the real characteristics of the physical device are fully reproduced. A three-dimensional modeling software (such as SolidWorks) is used to construct a device geometry model, which is accurate to millimeter level, for example, the number of turns of the transformer winding, the shape of the core, etc. The material property model is formed according to the material property setting parameters, for example, the resistivity of copper winding, the permeability of silicon steel sheet, etc., which directly affect the electrical and thermal performance of the device. The operation data model is established synchronously with the device operation data, and the voltage, current, temperature and other parameters are obtained through real-time interface, so that the digital twin can reflect the instantaneous state of the device. The environmental impact model is simulated by setting the temperature variation range, humidity fluctuation rule, pollution level, etc. in different seasons, and environmental factors have a significant impact on device aging and failure probability, such as high temperature environment which can accelerate the aging of insulation materials;

[0138] The operation and maintenance decision matrix includes inspection priority, maintenance window period, parameter adjustment scheme, etc. For example, when the decision matrix proposes "to perform cooling system maintenance on a certain transformer in the next 24 hours", the digital twin will simulate this operation and update the device operation state in real time. Through the operation data model, the temperature change before and after maintenance is monitored, such as whether the hot spot temperature decreases from 85℃ to 75℃ after cooling system maintenance; the reliability improvement of the device after maintenance is evaluated by the failure probability model, such as whether the insulation aging failure probability decreases from 15% to 8%. For complex decisions (such as coordinated operation involving multiple devices), the digital twin can simulate the chain reaction during the execution of the decision, for example, adjusting the on-off time of a certain circuit breaker may affect the voltage stability of adjacent devices, and the digital twin can quantify this impact and output the corresponding operating parameter change curve;

[0139] In summary, a virtual verification environment is provided for operation and maintenance decision-making, which enables operation and maintenance personnel to evaluate the effectiveness of the decision before implementation, optimize maintenance plans, reduce unnecessary maintenance work, and reduce operation and maintenance costs, which is conducive to significantly improving the decision reliability and economy of the power field operation and maintenance monitoring system.

[0140] As an optional embodiment, when the digital twin technology verifies the effectiveness of the operation and maintenance decision matrix, it further includes:

[0141] A digital twin model consistent with the parameters of the physical power field device is constructed, which includes a device electrical parameter model, a mechanical structure model and an environmental impact model;

[0142] The tuning strategy parameters in the operation and maintenance decision matrix are input into the digital twin model to simulate the generation of a predicted state data set;

[0143] The geodesic distance between the predicted state data set and the measured data of the physical device in the high-dimensional state space is calculated as a quantitative indicator of the effectiveness of the decision;

[0144] When the geodesic distance exceeds the preset threshold, the density distribution of historical fault cases in the high-dimensional state space is recalculated, and the fault warning threshold surface of the key parameters is adjusted;

[0145] By comparing the topological feature changes in the predicted state data set with the topological features of historical fault cases, a verification report containing fault mode matching degree is generated, and the parameter weight of the space-time feature extraction algorithm is adjusted based on the verification report.

[0146] Among them, the electrical parameter model is accurate to the volt-ampere characteristic, impedance frequency characteristic, etc. of each element, such as the contact resistance of the circuit breaker, the eddy current loss coefficient of the transformer winding, etc. The mechanical structure model uses finite element analysis method to simulate the structural response of the equipment under the action of vibration and stress, such as fatigue life prediction of isolating switch operating mechanism. The environmental impact model considers the influence of temperature, humidity, pollution level and other factors on the performance of the equipment, such as establishing a nonlinear relationship model between temperature and insulation aging rate. The model parameters are constantly calibrated by field measurement data to ensure that the error between the digital twin model and the physical equipment is less than 5%, reaching the highest precision standard in the industry;

[0147] Among them, when the decision matrix proposes "to increase the reactive power compensation capacity of a certain line by 20%", the digital twin model will output the parameter changes such as voltage distribution and current density of the equipment after the operation. The geodesic distance between the predicted state data set and the physical equipment measured data in the high-dimensional state space is calculated as a quantitative index of decision effectiveness. The geodesic distance takes into account the geometric characteristics of the state space, which can more accurately reflect the difference between the two data sets. For example, after adjusting the opening and closing time of a circuit breaker, the geodesic distance between the predicted contact wear rate of the digital twin model and the measured value of the physical equipment in the high-dimensional state space is 3.2, which intuitively indicates the deviation between the implementation effect of the decision and the expectation;

[0148] Among them, if it is found that the geodesic distance between the predicted state and the measured state of a certain transformer continues to increase, the system will automatically analyze the density distribution of historical insulation fault cases, and adjust the warning threshold surface of key parameters such as temperature and partial discharge to make the warning more in line with the actual operation state of the equipment. By comparing the topological feature changes in the predicted state data set with the topological features of historical fault cases, a verification report containing fault mode matching degree is generated. For example, the matching degree of the hole feature in the predicted state data set with the topological features of historical insulation aging faults is 85%, which indicates the risk of insulation aging. Based on the verification report, the parameter weight of the space-time feature extraction algorithm is adjusted to strengthen the recognition ability of such fault features, for example, the weight of temperature change rate feature is increased from 0.3 to 0.5 to enhance the sensitivity to abnormal temperature changes, which is conducive to improving the intelligent level and reliability of the electric field operation and debugging monitoring system.

[0149] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A monitoring and maintenance system suitable for electric fields, characterized in that, include: The data acquisition unit collects operational data of the electric field equipment at multi-dimensional monitoring points. The operational data includes electrical parameters, mechanical vibration parameters, environmental parameters, and chemical gas composition data, and constructs a multi-modal monitoring dataset. The feature analysis module extracts spatiotemporal features from the multimodal monitoring dataset, generating a multidimensional feature tensor containing temporal features, frequency domain features, and spatial distribution features. The spatiotemporal coupling relationship between features is obtained through tensor decomposition technology. The calculation module constructs a device health state space based on the spatiotemporal coupling relationship, calculates the density distribution of historical fault cases in the state space, and determines the fault early warning threshold surface of key parameters. The risk assessment module maps real-time monitoring data to the device health status space, measures the deviation between the real-time status point and the warning threshold surface using Mahalanobis distance, and generates a fault risk heat map. The decision-making module combines the fault risk heat map and equipment topology, uses graph neural networks to predict fault propagation paths, and generates an operation and maintenance decision matrix that includes inspection priorities and maintenance windows. The verification unit verifies the effectiveness of the operation and maintenance decision matrix using digital twin technology, and dynamically adjusts the parameter weights of the spatiotemporal feature extraction algorithm based on the verification results. The construction of the device health state space based on the spatiotemporal coupling relationship specifically includes: A high-dimensional state space is constructed using Riemannian manifolds; First, whiten the multidimensional feature tensor to eliminate the correlation between features; The processed data is then mapped to the Riemannian manifold space via an exponential mapping. Within the Riemannian manifold space, the kernel density estimation method is used to calculate the probability density distribution of historical failure cases, thereby determining the failure early warning threshold surface of key parameters. The method of combining the fault risk heatmap with the equipment topology and using a graph neural network to predict the fault propagation path is as follows: Construct a graph neural network that includes graph convolutional layers, graph attention layers, and an output layer; By aggregating the neighborhood information of device nodes using graph convolutional layers, node feature updates can be achieved. By calculating the attention coefficients between nodes through the graph attention layer, the ability to identify key nodes in the device topology is enhanced. Finally, the output layer predicts the propagation path of the fault in the device topology based on the updated node features and outputs the path probability distribution. The process of generating a fault risk heatmap by measuring the deviation between real-time state points and the warning threshold surface using Mahalanobis distance is as follows: First, the real-time monitoring data is standardized to have zero mean and unit variance. A local weighted regression method is used to locally fit the warning threshold surface; Calculate the Mahalanobis distance from the real-time state point to the fitted surface, and smooth the Mahalanobis distance calculation results by combining Gaussian kernel density estimation. Using the geographical location of the equipment as a coordinate reference, a heat map of failure risk is finally generated; The verification of the effectiveness of the operation and maintenance decision matrix using digital twin technology specifically involves: Establish a digital twin that includes equipment geometry, material properties, operational data, and environmental impact; Use 3D modeling software to construct the geometric model of the equipment, and set parameters according to the material properties to form a material property model; Synchronize equipment operation data to establish an operation data model, and simulate environmental factors to construct an environmental impact model; Simulate the execution of the operation and maintenance decision matrix in a digital twin, and collect results including at least the operating status and failure probability. When determining the fault early warning threshold surface for key parameters, a topological data analysis method is used to extract the fault feature topology in the Riemannian manifold space, specifically: Construct Vietoris-Rips complexes of historical failure cases to capture the features of holes and connected branches in the manifold space; Calculate persistent homology groups at different scales and identify fault features with topological persistence; By associating persistent cohomology features with equipment physical failure modes, the topological critical points of key parameters are determined. Generate geometrically invariant threshold surfaces in Riemannian manifold space by using topological critical points; When the digital twin technology verifies the effectiveness of the operation and maintenance decision matrix, it also includes: Construct a digital twin model that is consistent with the parameters of the physical electric field equipment. The digital twin model includes an electrical parameter model, a mechanical structure model, and an environmental impact model. Input the debugging strategy parameters in the operation and maintenance decision matrix into the digital twin model to simulate and generate a predicted state dataset; The geodesic distance between the predicted state dataset and the measured data of physical devices in the high-dimensional state space is calculated as a quantitative indicator of decision effectiveness. When the geodetic distance exceeds the preset threshold, the density distribution of historical fault cases in the high-dimensional state space is recalculated, and the fault warning threshold surface of key parameters is adjusted. By comparing the changes in topological features in the predicted state dataset with the topological features of historical failure cases, a verification report containing the failure mode matching degree is generated, and the parameter weights of the spatiotemporal feature extraction algorithm are adjusted based on the verification report.

2. The operation, maintenance, commissioning, and monitoring system for electric fields according to claim 1, characterized in that, The spatiotemporal feature extraction of the multimodal monitoring dataset is performed in the following manner: Construct a hybrid neural network model consisting of two-dimensional convolutional neural network layers, long short-term memory network layers, and fully connected layers; The two-dimensional convolutional neural network layer is configured with multiple convolutional layers and pooling layers with different kernel sizes to extract spatial features of the data. The time-series features of data are processed by utilizing the gating mechanism of the Long Short-Term Memory network layer; Finally, the spatiotemporal features are fused through a fully connected layer to output a multidimensional feature tensor containing temporal features, frequency domain features, and spatial distribution features.

3. The operation, maintenance, commissioning, and monitoring system for electric fields according to claim 1, characterized in that, The electrical parameters include at least three of the following: partial discharge quantity, insulation resistance, dielectric loss factor, leakage current, dynamic capacitance and capacitance change rate, ultra-high frequency discharge signal, and pulse current signal. The mechanical vibration parameters include acceleration amplitude, vibration frequency distribution, and vibration phase characteristics. The environmental parameters include temperature, humidity, air pressure, and wind speed. The chemical gas composition data includes... Concentration of decomposition products and content of dissolved gases in transformer oil.

4. The operation, maintenance, commissioning, and monitoring system for electric fields according to claim 1, characterized in that, The process of mapping the processed data to the Riemannian manifold space via exponential mapping further includes: Calculate the geodesic distance between the current data point and the origin of the Riemannian manifold space; The scale parameter of the exponential mapping is dynamically adjusted based on the geodesic distance. When the distance is greater than the preset threshold, a contraction-type exponential mapping is used to suppress the influence of outliers. When the distance is less than a preset threshold, an expanded exponential mapping is used to enhance the distinguishability of normal state data; The geometric relationships between data points are preserved by parallel movement operations on the manifold.

Images