A power equipment overheating fault early warning method based on principal component analysis

Abstract

The application discloses a power equipment overheating fault early warning method based on principal component analysis, and relates to the technical field of power equipment state monitoring, comprising: collecting multi-source monitoring data, performing preprocessing, and generating standardized time series data set; performing principal component analysis on the standardized data, and obtaining score sequence and characteristic vector; constructing a thermodynamic evolution interval, comparing evolution trajectories, and generating deviation judgment results; calculating disturbance propagation speed, dynamically modeling change relationship, and generating break judgment results; reconstructing residual time series data, inputting improved MTGNN, and identifying propagation path and source node; and comprehensively integrating all early warning indexes, and outputting overheating grade and overheating point position. The application realizes early warning of power equipment overheating fault and accurate positioning of potential overheating points by performing principal component analysis on multi-source operation data of power equipment and combining graph structure time series modeling to jointly analyze abnormal thermal evolution characteristics and propagation behavior.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring technology, and in particular to a method for early warning of overheating faults in power equipment based on principal component analysis. Background Technology

[0002] With the continuous expansion of power system scale and the increasing complexity of equipment operating conditions, key power equipment such as transformers, switchgear, and cable joints are prone to localized overheating during long-term operation due to load fluctuations, environmental changes, or structural aging. To ensure the safe and stable operation of the power system, existing technologies typically involve deploying temperature, current, and voltage sensors at critical locations on the equipment to continuously monitor its operating status and conduct overheating fault early warning analysis based on the collected operating data. This type of method has become an important technical means for power equipment condition monitoring in practical applications.

[0003] Existing methods for early warning of overheating in power equipment mostly rely on threshold judgments from single sensors or static models built based on historical statistical features. For example, they identify anomalies by setting upper temperature limits, rate of change thresholds, or using simple time series analysis methods. Some methods introduce dimensionality reduction techniques such as principal component analysis to compress features from multi-source monitoring data before anomaly detection. However, most of these techniques focus on numerical deviations at a single moment or over a short period, making it difficult to systematically model the coupling relationships between multi-sensor data under complex operating conditions.

[0004] In actual operating environments, overheating faults in power equipment often manifest as a gradual diffusion of heat within the equipment's internal structure. Their abnormal characteristics gradually emerge through the combined changes in multi-dimensional operating data before a significant temperature increase. Current technologies generally lack the ability to jointly analyze the principal component evolution process, thermal disturbance propagation behavior, and anomaly source location. This makes it difficult to accurately characterize the formation trend and propagation path of overheating anomalies under conditions of strong correlation among multiple sensors and rapid changes in operating conditions. This easily leads to delayed early warnings, high false alarm rates, and a lack of clear explanations for anomaly propagation in the early warning results, failing to meet the practical needs for refined early warning and location of overheating faults in power equipment.

[0005] Therefore, how to provide a method for early warning of overheating faults in power equipment based on principal component analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method for early warning of overheating faults in power equipment based on principal component analysis (PCA). This invention digitizes and analyzes multi-source time-series operating data of power equipment, extracts key features reflecting the evolution of the equipment's thermal state through PCA, and further combines graph-structured time-series modeling to jointly characterize the spatial propagation and temporal evolution of the reconstructed residuals from the principal components. This enables a comprehensive determination of the formation process, propagation behavior, and location of the overheating anomaly in the power equipment. By constructing mechanisms for determining deviations in principal component thermal evolution, analyzing thermal disturbance propagation, and identifying the propagation path of anomaly residuals, this invention provides early warning of potential overheating risks before the temperature rises. It boasts advantages such as high timeliness, low false alarm rate, clear anomaly location, and strong interpretability, making it suitable for refined intelligent early warning of overheating faults in power equipment under complex operating conditions.

[0007] A method for early warning of overheating faults in power equipment based on principal component analysis according to an embodiment of the present invention includes:

[0008] Collect multi-source digital monitoring data generated during the operation of power equipment, preprocess the multi-source digital monitoring data, and generate a standardized dataset;

[0009] Principal component analysis is performed on the standardized dataset to extract principal component eigenvectors and corresponding principal component score sequences. The temporal representation of principal components is then constructed based on the principal component score sequences.

[0010] Based on the principal component time series representation, the evolution trajectory of the principal component score sequence within a continuous time window is extracted, the allowable thermodynamic evolution interval of the principal component score sequence is constructed, and the evolution trajectory is compared with the allowable thermodynamic evolution interval to generate the principal component thermal evolution deviation result.

[0011] Based on the principal component score sequence, the temporal variation relationship of principal component disturbance between sensors is calculated, the propagation speed of principal component disturbance is determined, and the propagation speed is compared with the upper bound of propagation speed adaptively generated based on the thermal inertia parameters of the equipment to generate the thermal disturbance boundary judgment result.

[0012] The standardized dataset is reconstructed using principal component eigenvectors to obtain principal component reconstruction residuals. A residual induction graph is constructed, and the residual time series data is used as node feature input to improve the MTGNN model. The temporal propagation behavior of principal component reconstruction residuals on the residual induction graph is analyzed, and the abnormal residual propagation path and abnormal source node are output.

[0013] Based on the principal component thermal evolution deviation results, thermal disturbance boundary judgment results, abnormal residual propagation paths, and abnormal source nodes, the overheating fault warning level and potential overheating hot spot locations of power equipment are determined, and the corresponding overheating fault warning information is output.

[0014] Optionally, the multi-source digital monitoring data includes temperature monitoring data collected from different locations of the power equipment, current data, voltage data, and load rate data characterizing the operating status of the power equipment, as well as ambient temperature data and ambient humidity data related to the operating environment of the power equipment.

[0015] Optionally, the preprocessing of multi-source digital monitoring data includes synchronizing the multi-source digital monitoring data according to a unified timestamp, and performing abnormal sampling point removal, missing data completion, and unit unification processing on the synchronized data.

[0016] Optionally, the construction of the principal component temporal representation based on the principal component score sequence includes:

[0017] The mean of the standardized dataset is calculated on each feature dimension, and the mean is subtracted from the corresponding feature of each time sample to obtain the centered dataset;

[0018] Based on the centralized dataset, the covariance matrix is ​​calculated on the feature dimension, and the covariance matrix is ​​decomposed to obtain the feature vector group and the corresponding feature value sorted from high to low according to the variance contribution.

[0019] The number of principal components is determined based on the comparison between the cumulative variance contribution rate and the target proportion threshold, and the first few corresponding eigenvectors are selected to form a set of principal component eigenvectors.

[0020] The centralized dataset is linearly projected along the set of principal component feature vectors to obtain the principal component score value corresponding to each time sample, and arranged in the order of timestamps to form the principal component score sequence.

[0021] Based on the principal component score sequence, slices are generated according to the sampling time order and the continuous time window length and step size to generate a principal component time series representation containing timestamp, principal component score vector and window index.

[0022] Optionally, the generation of principal component thermal evolution deviation results includes:

[0023] Obtain the time-series representation of the principal components, read the load rate data and ambient temperature data corresponding to the timestamps from the standardized dataset, determine the operating condition label corresponding to each continuous time window, and read the thermal inertia parameters and thermal diffusivity parameters corresponding to the power equipment from the equipment parameter library.

[0024] The principal component time series representation is sliced ​​according to the window length and step size to obtain the evolution trajectory of the principal component score sequence within each window, and the change amount, change rate and change acceleration of the principal component score of adjacent sampling points are calculated to form the window thermal evolution characteristics.

[0025] Based on the principal component feature vector, the absolute values ​​of the loads of temperature monitoring data, current data, voltage data, load rate data, and environmental parameter data in each principal component are extracted as weights. Combined with thermal inertia parameters and thermal diffusion characteristic parameters, the thermal inertia mapping coefficient and thermal diffusion mapping coefficient of each principal component are calculated and associated with the operating condition label.

[0026] For each window and each principal component, the boundary of the rate of change and the boundary of the acceleration of change of the principal component score within the window are adaptively generated based on the thermal inertia mapping coefficient and the thermal diffusion mapping coefficient. Historical windows with consistent operating condition labels are selected from historical normal operation data. The upper and lower boundary ranges of the corresponding principal component thermal evolution characteristics are statistically analyzed. The boundary and the upper and lower boundary ranges are merged and constrained to construct the allowable thermodynamic evolution range of the principal component score sequence within the window.

[0027] The evolution trajectory and window thermal evolution characteristics are compared with the allowable thermodynamic evolution range to determine the timestamp set and principal component set that exceed the allowable thermodynamic evolution range. The principal component thermal evolution deviation results are generated and the corresponding window index and timestamp index are output.

[0028] Optionally, the generation of thermal disturbance boundary violation determination results includes:

[0029] Obtain the principal component score sequence, the physical connection relationship between sensors and the corresponding connection distance, and the thermal inertia parameter and thermal diffusion characteristic parameter in the device parameter library;

[0030] The principal component score sequence is differentially processed between adjacent sampling times to obtain the principal component perturbation sequence of each principal component at each sampling time. Based on the load values ​​of each sensor channel on each principal component in the principal component feature vector, the principal component perturbation sequence is mapped to the channel perturbation sequence corresponding to each sensor channel.

[0031] For each pair of physically connected adjacent sensors, the channel perturbation sequence is read separately, and the propagation delay of the two channel perturbation sequences is determined by cross-correlation alignment. The propagation delay is the time offset that makes the correlation between the two channel perturbation sequences reach its maximum value.

[0032] The propagation velocity of the principal component perturbation between adjacent sensors is calculated based on the connection distance and the propagation delay, forming a set of propagation velocities for each pair of adjacent sensors;

[0033] The propagation speed upper bound is generated based on thermal inertia parameters and thermal diffusion characteristic parameters. The propagation speed upper bound is adaptively updated with the operating condition label and time window. The propagation speed set is compared with the propagation speed upper bound one by one. When the propagation speed of any adjacent sensor pair exceeds the corresponding propagation speed upper bound, the thermal disturbance boundary violation judgment result is output.

[0034] Optionally, the output anomaly residual propagation path and anomaly source node include:

[0035] The principal component feature vectors are called to perform linear reconstruction on the standardized dataset. The original data and the reconstructed data are subtracted time-stamp by time to obtain the principal component reconstruction residual time series.

[0036] Based on the sensor layout diagram, the physical connection relationship of the nodes is determined, and the residual correlation between the nodes is calculated by combining the principal component reconstruction residual time series. The physical connection relationship and the residual correlation are fused to generate the adjacency relationship of the residual induced graph.

[0037] The residual time series reconstructed from principal components is used as the node input, and the adjacency relationship of the residual induced graph is used as the structure input. These are then fed into an improved MTGNN model. The improved MTGNN model consists of several sequentially connected graph-temporal alternation blocks. Each graph-temporal alternation block sequentially includes graph convolution processing and temporal convolution processing, and is embedded with the following structure:

[0038] A thermal topological attention fusion layer is inserted before graph convolution processing. The edge attention weights are calculated based on the correlation between the thermal resistance between nodes and the residual between nodes. The edge attention weights are then weighted and fused with the adjacency relationship to obtain the adjusted adjacency relationship. The adjusted adjacency relationship is then input into the graph convolution processing.

[0039] A heat source sensing gating unit is connected in parallel between the graph convolution processing output and the temporal convolution processing input. The node temperature baseline, material thermal conductivity and node-to-shell distance are read to generate a gating factor. The graph convolution processing output is modulated node by node. The modulation result is spliced ​​with the graph convolution processing output to form the temporal convolution input feature.

[0040] One-dimensional temporal convolution is performed on the temporal convolution input features to obtain the first temporal feature. An alternating dilation residual block is embedded between two adjacent one-dimensional temporal convolution layers. The alternating dilation residual block takes the first temporal feature as input and performs one-dimensional temporal convolution with different dilation coefficients in sequence. The two convolution features are added together and superimposed with the first temporal feature through a short-circuit residual connection to obtain the fused temporal feature. The fused temporal feature is then passed to the next figure - the temporal alternating block.

[0041] After the last graph-temporal alternation block, an output layer is set up to output the node residual propagation strength sequence and the edge propagation weight sequence to determine the abnormal source node and the abnormal residual propagation path;

[0042] Using principal component reconstruction of residual time series and residual induced graph as input features, historical normal operating condition samples and confirmed overheating operating condition samples are selected as training data to perform end-to-end training on the improved MTGNN model. The training output is the residual propagation intensity sequence of each node and the overheating probability sequence of each node.

[0043] Optionally, the output corresponding overheating fault warning information includes:

[0044] Read the principal component thermal evolution deviation results, extract the set of timestamps where the principal component thermal evolution deviation occurs, and generate the corresponding principal component thermal evolution deviation intensity sequence based on the magnitude of the principal component score sequence exceeding the allowable thermodynamic evolution range at the timestamp.

[0045] Read the thermal disturbance boundary violation determination result, extract the set of adjacent sensor pairs where thermal disturbance boundary violation occurred, and generate the corresponding thermal disturbance boundary violation intensity set according to the magnitude of the principal component disturbance propagation speed in each adjacent sensor pair exceeding the corresponding propagation speed upper bound.

[0046] Read the abnormal residual propagation path and abnormal source node, and for each abnormal residual propagation path, count the number of nodes in the path, the physical distance between the starting node and the ending node of the path, and the cumulative value of the residual propagation intensity of each node on the path to form a path propagation feature set.

[0047] Based on the principal component thermal evolution deviation intensity sequence, thermal disturbance boundary breach intensity set, and path propagation characteristic set, a comprehensive risk score is calculated according to a pre-set weight combination relationship. The comprehensive risk score is then compared with a multi-level early warning threshold set to determine the corresponding overheating fault early warning level.

[0048] The abnormal source node is taken as the potential hot spot location. The overheating fault warning level, potential hot spot location, abnormal residual propagation path and comprehensive risk score are uniformly packaged into overheating fault warning information and output.

[0049] The beneficial effects of this invention are:

[0050] This invention constructs an overheating fault early warning process based on principal component analysis by uniformly digitizing and analyzing multi-source operating data of power equipment. This enables the effective extraction and characterization of thermal evolution features implicit in multi-dimensional data during equipment operation. By compressing multi-source monitoring data into a physically meaningful principal component time-series representation, this invention avoids the information bias caused by single sensor threshold judgments, allowing overheating anomalies to be identified before they manifest as temperature increases, thus improving the timeliness of overheating fault early warning for power equipment.

[0051] This invention further combines principal component thermal evolution deviation judgment and thermal disturbance propagation analysis mechanisms to describe the diffusion process of abnormal thermal behavior within the internal structure of equipment. This allows the early warning results to not only reflect the existence of an anomaly but also its propagation characteristics and development trend. Through joint analysis of principal component disturbance propagation speed and equipment thermal inertia characteristics, this invention effectively reduces false alarms caused by operating condition fluctuations or instantaneous noise, and improves the stability and reliability of early warnings under conditions of strong correlation among multiple sensors and rapid changes in operating status.

[0052] This invention introduces a graph structure time-series modeling method based on residual-induced graphs to identify the propagation path and anomaly source nodes of principal component reconstruction residuals, giving overheat warning results clear spatial orientation and interpretability. The warning information not only provides the overheat risk level but also indicates the location of potential overheating hotspots and anomaly propagation paths, offering maintenance personnel directly referable decision-making basis. This helps improve the efficiency and accuracy of locating and handling overheating faults in power equipment, enhancing the intelligence level of power equipment operation and management. Attached Figure Description

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0054] Figure 1 This is a flowchart of a power equipment overheating fault early warning method based on principal component analysis proposed in this invention;

[0055] Figure 2 This is a schematic diagram of principal component disturbance propagation and boundary violation determination in a power equipment overheating fault early warning method based on principal component analysis proposed in this invention.

[0056] Figure 3 This is a schematic diagram of the improved MTGNN model structure for a power equipment overheating fault early warning method based on principal component analysis proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1 , Figure 2 and Figure 3 A method for early warning of overheating faults in power equipment based on principal component analysis, comprising:

[0059] Collect multi-source digital monitoring data generated during the operation of power equipment, preprocess the multi-source digital monitoring data, and generate a standardized dataset;

[0060] Principal component analysis is performed on the standardized dataset to extract principal component eigenvectors and corresponding principal component score sequences. The temporal representation of principal components is then constructed based on the principal component score sequences.

[0061] Based on the principal component time series representation, the evolution trajectory of the principal component score sequence within a continuous time window is extracted, the allowable thermodynamic evolution interval of the principal component score sequence is constructed, and the evolution trajectory is compared with the allowable thermodynamic evolution interval to generate the principal component thermal evolution deviation result.

[0062] Based on the principal component score sequence, the temporal variation relationship of principal component disturbance between sensors is calculated, the propagation speed of principal component disturbance is determined, and the propagation speed is compared with the upper bound of propagation speed adaptively generated based on the thermal inertia parameters of the equipment to generate the thermal disturbance boundary judgment result.

[0063] The standardized dataset is reconstructed using principal component eigenvectors to obtain principal component reconstruction residuals. A residual induction graph is constructed, and the residual time series data is used as node feature input to improve the MTGNN model. The temporal propagation behavior of principal component reconstruction residuals on the residual induction graph is analyzed, and the abnormal residual propagation path and abnormal source node are output.

[0064] Based on the principal component thermal evolution deviation results, thermal disturbance boundary judgment results, abnormal residual propagation paths, and abnormal source nodes, the overheating fault warning level and potential overheating hot spot locations of power equipment are determined, and the corresponding overheating fault warning information is output.

[0065] In this embodiment, the multi-source digital monitoring data includes temperature monitoring data collected from different locations of the power equipment, current data, voltage data, and load rate data characterizing the operating status of the power equipment, as well as ambient temperature data and ambient humidity data related to the operating environment of the power equipment.

[0066] In this embodiment, the preprocessing of multi-source digital monitoring data includes synchronizing the multi-source digital monitoring data according to a unified timestamp, and performing abnormal sampling point removal, missing data completion, and unit unification processing on the synchronized data.

[0067] In this embodiment, the construction of the principal component temporal representation based on the principal component score sequence includes:

[0068] The mean of the standardized dataset is calculated on each feature dimension, and the mean is subtracted from the corresponding feature of each time sample to obtain the centered dataset;

[0069] The covariance matrix is ​​calculated along the feature dimension based on the centralized dataset. Eigenvalue decomposition is then performed on the covariance matrix to obtain feature vector sets and corresponding eigenvalues ​​sorted by variance contribution from high to low. Specifically, the calculation of the covariance matrix along the feature dimension based on the centralized dataset involves:

[0070] The centralized dataset is represented as a multidimensional data matrix with time samples as rows and feature dimensions as columns, where each column of the data matrix corresponds to a sensor feature dimension.

[0071] For each feature dimension of the centralized data matrix, sample values ​​under the same time index are selected to form a feature sequence of the feature dimension in the full time sample range;

[0072] For any pair of feature dimensions, the joint change relationship of the corresponding feature sequence is calculated over the entire time sample range. The joint change relationship is obtained by summing the product of the values ​​of the two feature dimensions under each time sample and normalizing it according to the number of samples.

[0073] The joint change relationships between each feature dimension are filled into the corresponding positions of the matrix according to the feature dimension index order, and a covariance matrix with the feature dimension as the row and column index is constructed. The diagonal elements of the matrix represent the change intensity of the corresponding feature dimension itself, and the off-diagonal elements represent the degree of cooperative change between different feature dimensions.

[0074] The number of principal components is determined based on the comparison between the cumulative variance contribution rate and the target proportion threshold. The top few eigenvectors are then selected to form a set of principal component eigenvectors. Specifically, the determination of the number of principal components based on the comparison between the cumulative variance contribution rate and the target proportion threshold is as follows:

[0075] Based on the eigenvalue results, the variance contribution ratio of each eigenvector is calculated according to the sorting order of the corresponding eigenvectors, and then accumulated to form a cumulative variance contribution rate sequence.

[0076] Read the target proportion threshold from the preset target proportion threshold, compare the cumulative variance contribution rate sequence with the target proportion threshold, and determine the number of feature vectors corresponding to the first time when the cumulative variance contribution rate reaches or exceeds the target proportion threshold.

[0077] The number of feature vectors is determined to be the number of principal components, and the corresponding feature vectors that are ranked first are selected to form the principal component feature vector set;

[0078] The centralized dataset is linearly projected along the set of principal component feature vectors to obtain the principal component score value corresponding to each time sample, and arranged in time stamp order to form the principal component score sequence. Specifically, the linear projection of the centralized dataset along the set of principal component feature vectors involves:

[0079] Using a centralized dataset as input data, the multidimensional feature vector corresponding to each time sample is selected sequentially according to the time sample order;

[0080] For each time sample's multidimensional feature vector, a linear mapping operation is performed with each principal component feature vector in the principal component feature vector set to obtain the corresponding projection results of the time sample in each principal component direction;

[0081] The projection results of each time sample in each principal component direction are mapped one-to-one with the timestamps and arranged in order of timestamps to form the principal component score sequence.

[0082] Based on the principal component score sequence, slices are generated according to the sampling time order and the continuous time window length and step size to generate a principal component time series representation containing timestamp, principal component score vector and window index.

[0083] In this embodiment, the generation of principal component thermal evolution deviation results includes:

[0084] Obtain the time-series representation of the principal components, read the load rate data and ambient temperature data corresponding to the timestamps from the standardized dataset, determine the operating condition label corresponding to each continuous time window, and read the thermal inertia parameters and thermal diffusivity parameters corresponding to the power equipment from the equipment parameter library.

[0085] The principal component time series representation is sliced ​​according to the window length and step size to obtain the evolution trajectory of the principal component score sequence within each window. The change, rate of change, and acceleration of the principal component scores at adjacent sampling points are calculated to form the window thermal evolution characteristics. Specifically, the calculation of the change, rate of change, and acceleration of the principal component scores at adjacent sampling points is as follows:

[0086] Within each time window, the principal component scores of two adjacent sampling points are selected sequentially according to the timestamp order. By performing a difference operation between the principal component score of the later sampling point and the principal component score of the earlier sampling point, the change in principal component scores between adjacent sampling points is obtained.

[0087] Based on the change in principal component scores between adjacent sampling points, and combined with the fixed sampling time interval between adjacent sampling points, the change in principal component scores is normalized over time to obtain the rate of change of principal component scores within the window.

[0088] Based on the rate of change, the rate of change of adjacent sampling points is differentially calculated again to obtain the degree of change of the rate of change between adjacent sampling points. Then, the normalization process is performed in combination with the sampling time interval to form the acceleration of the principal component score change.

[0089] Based on the principal component eigenvectors, the absolute values ​​of the loads of temperature monitoring data, current data, voltage data, load rate data, and environmental parameter data in each principal component are extracted as weights. Combined with thermal inertia parameters and thermal diffusivity parameters, the thermal inertia mapping coefficient and thermal diffusivity mapping coefficient of each principal component are calculated and associated with the operating condition label. Specifically, the calculation of the thermal inertia mapping coefficient and thermal diffusivity mapping coefficient of each principal component is as follows:

[0090] For each principal component, the load values ​​of temperature monitoring data, current data, voltage data, load rate data, and environmental parameter data in the principal component feature vector are read respectively. The absolute value of the load value is taken and then normalized to obtain a set of parameter weights that reflect the degree of influence of various operating parameters on the principal component.

[0091] The set of parameter weights is weighted and fused with the thermal inertia parameters corresponding to the device. The thermal inertia parameters include the thermal capacity and thermal response characteristics determined by the device structure and material properties. The thermal inertia mapping coefficients corresponding to the principal components are calculated by weighting and accumulating the weights of each parameter with the corresponding thermal inertia parameters.

[0092] The set of parameter weights is weighted and fused with the thermal diffusion characteristic parameters corresponding to the device. The thermal diffusion characteristic parameters include structural and medium characteristics that characterize the ability of heat to spread inside the device. The thermal diffusion mapping coefficient corresponding to the principal component is calculated by weighting and accumulating the weights of each parameter with the corresponding thermal diffusion characteristic parameters.

[0093] For each window and each principal component, the boundary of the rate of change and acceleration of the principal component score within the window is adaptively generated based on the thermal inertia mapping coefficient and the thermal diffusion mapping coefficient. Historical windows with consistent operating condition labels are selected from historical normal operation data, and the upper and lower boundary ranges of the corresponding principal component thermal evolution characteristics are statistically analyzed. The boundary and the upper and lower boundary ranges are merged and constrained to construct the allowable thermodynamic evolution interval of the principal component score sequence within the window. Specifically, the boundary of the rate of change and acceleration of the principal component score within the window is adaptively generated based on the thermal inertia mapping coefficient and the thermal diffusion mapping coefficient.

[0094] For each principal component within the current time window, the corresponding thermal inertia mapping coefficient and thermal diffusion mapping coefficient are read as adjustment factors characterizing the slowness of the principal component's thermal response and the activity of its thermal propagation.

[0095] Based on the thermal inertia mapping coefficient, the rate of change of principal component scores is scaled, wherein principal components with larger thermal inertia mapping coefficients correspond to a lower allowable rate of change range, and principal components with smaller thermal inertia mapping coefficients correspond to a higher allowable rate of change range, thereby generating the rate of change boundary of principal components within the current window.

[0096] Based on the thermal diffusion mapping coefficient, the acceleration of principal component score change is scaled, wherein the principal component with a larger thermal diffusion mapping coefficient corresponds to a higher allowable acceleration range, and the principal component with a smaller thermal diffusion mapping coefficient corresponds to a lower allowable acceleration range, thereby generating the acceleration boundary of the principal component within the current window;

[0097] The evolution trajectory and window thermal evolution characteristics are compared with the allowable thermodynamic evolution range to determine the timestamp set and principal component set that exceed the allowable thermodynamic evolution range. The principal component thermal evolution deviation results are generated and the corresponding window index and timestamp index are output.

[0098] In this embodiment, the generation of thermal disturbance boundary violation determination results includes:

[0099] Obtain the principal component score sequence, the physical connection relationship between sensors and the corresponding connection distance, and the thermal inertia parameter and thermal diffusion characteristic parameter in the device parameter library;

[0100] The principal component score sequence is differentially processed between adjacent sampling times to obtain the principal component perturbation sequence of each principal component at each sampling time. Based on the load values ​​of each sensor channel on each principal component in the principal component feature vector, the principal component perturbation sequence is mapped to the channel perturbation sequence corresponding to each sensor channel.

[0101] For each pair of physically connected adjacent sensors, channel perturbation sequences are read separately. The propagation delay of the two channel perturbation sequences is determined using a cross-correlation alignment method. This propagation delay is the time offset that maximizes the correlation between the two channel perturbation sequences. Specifically, the cross-correlation alignment method for determining the propagation delay of the two channel perturbation sequences is as follows:

[0102] For each pair of physically connected adjacent sensors, the corresponding two channel perturbation sequences are extracted in time stamp order. The two channel perturbation sequences are then length-aligned and normalized within the same time window to eliminate the influence of dimensional differences on correlation calculation.

[0103] Within a preset time offset range, one of the channel perturbation sequences is gradually shifted in time, and at each time offset position, the correlation strength between the shifted channel perturbation sequence and the other channel perturbation sequence during the overlapping time period is calculated;

[0104] The correlation strength corresponding to each time offset position is compared to determine the time offset position corresponding to the maximum correlation strength, and the time offset position is used as the propagation delay between the adjacent sensors.

[0105] Based on the connection distance and the propagation delay, the propagation velocity of the principal component perturbation between adjacent sensors is calculated to form a set of propagation velocities for each pair of adjacent sensors. Specifically, the calculation of the propagation velocity of the principal component perturbation between adjacent sensors is as follows:

[0106] For each pair of physically connected adjacent sensors, read the physical connection distance between the adjacent sensor pairs and the propagation delay of the adjacent sensor pairs within the current time window.

[0107] Based on the physical connection distance and the propagation delay, the change in distance of the principal component disturbance propagation between adjacent sensor pairs per unit time is calculated to obtain the propagation speed of the principal component disturbance corresponding to the adjacent sensor pairs;

[0108] The propagation velocity of the principal component disturbance calculated by each adjacent sensor pair within the current time window is associated and recorded with the corresponding sensor pair identifier and time window index to form a set of propagation velocities of adjacent sensor pairs.

[0109] The propagation speed upper bound is generated based on thermal inertia parameters and thermal diffusion characteristic parameters. The propagation speed upper bound is adaptively updated with the operating condition label and time window. The propagation speed set is compared with the propagation speed upper bound one by one. When the propagation speed of any adjacent sensor pair exceeds the corresponding propagation speed upper bound, the thermal disturbance boundary violation judgment result is output.

[0110] In this embodiment, the output abnormal residual propagation path and abnormal source node include:

[0111] The principal component feature vectors are called to perform linear reconstruction on the standardized dataset. The original data and the reconstructed data are subtracted time-stamp by time to obtain the principal component reconstruction residual time series.

[0112] Based on the sensor layout diagram, the physical connection relationship of the nodes is determined, and the residual correlation between the nodes is calculated by combining the principal component reconstruction residual time series. The physical connection relationship and the residual correlation are fused to generate the adjacency relationship of the residual induced graph.

[0113] The residual time series reconstructed from principal components is used as the node input, and the adjacency relationship of the residual induced graph is used as the structure input. These are then fed into an improved MTGNN model. The improved MTGNN model consists of several sequentially connected graph-temporal alternation blocks. Each graph-temporal alternation block sequentially includes graph convolution processing and temporal convolution processing, and is embedded with the following structure:

[0114] A thermal topological attention fusion layer is inserted before graph convolution processing. The edge attention weights are calculated based on the correlation between the thermal resistance between nodes and the residual between nodes. The edge attention weights are then weighted and fused with the adjacency relationship to obtain the adjusted adjacency relationship. The adjusted adjacency relationship is then input into the graph convolution processing.

[0115] A heat source sensing gating unit is connected in parallel between the graph convolutional processing output and the temporal convolutional processing input. It reads the node temperature baseline, material thermal conductivity, and node-to-shell distance to generate a gating factor. The graph convolutional processing output is modulated node-by-node, and the modulation result is concatenated with the graph convolutional processing output to form the temporal convolutional input feature, where:

[0116] The node temperature baseline refers to the long-term stable operating temperature average of the corresponding sensor node under the same or similar load rate and environmental conditions when the equipment is in its historical normal operating conditions. The node temperature baseline is obtained by statistical analysis of historical normal operating data and is associated with the corresponding sensor node one by one.

[0117] The thermal conductivity of a material refers to the thermal conductivity parameter of the material corresponding to the component where the sensor node is located. The thermal conductivity of the material is read from the equipment design parameters or material property parameter table and is used to characterize the inherent characteristics of the component where the node is located in terms of heat transfer.

[0118] The node-to-casing distance refers to the shortest geometric distance between the sensor node's internal spatial location and the device's outer casing or heat dissipation boundary. The node-to-casing distance is calculated based on the device's structural layout diagram or three-dimensional structural model and is fixedly associated with the corresponding sensor node.

[0119] One-dimensional temporal convolution is performed on the temporal convolution input features to obtain the first temporal feature. An alternating dilation residual block is embedded between two adjacent one-dimensional temporal convolution layers. The alternating dilation residual block takes the first temporal feature as input and performs one-dimensional temporal convolution with different dilation coefficients in sequence. The two convolution features are added together and superimposed with the first temporal feature through a short-circuit residual connection to obtain the fused temporal feature. The fused temporal feature is then passed to the next figure - the temporal alternating block.

[0120] After the last graph-temporal alternation block, an output layer is set up to output the node residual propagation strength sequence and the edge propagation weight sequence to determine the abnormal source node and the abnormal residual propagation path;

[0121] Using principal component reconstruction of residual time series and residual induced graph as input features, historical normal operating condition samples and confirmed overheating operating condition samples are selected as training data to perform end-to-end training on the improved MTGNN model. The training output is the residual propagation intensity sequence of each node and the overheating probability sequence of each node.

[0122] This invention constructs an improved MTGNN model based on the original MTGNN's backbone structure of alternating graph convolution and temporal convolution processing. It uses the principal component reconstruction residual time series as node input and the adjacency relationships of the residual-induced graph as structural input. The improved MTGNN model consists of several sequentially connected graph-temporal alternating blocks. Three new structures are embedded in each alternating block at a fixed level. A thermal topological attention fusion layer is inserted before graph convolution processing to weight and adjust the adjacency relationships based on the correlation between node thermal resistance and residuals. The graph convolution output is then combined with the temporal convolution... A parallel heat source sensing gating unit is connected between the inputs to read the node temperature baseline, material thermal conductivity, and node-to-shell distance to generate a gating factor. The graph convolution output is then modulated node by node and concatenated with the original features to form a temporal convolution input feature. Alternating dilated residual blocks are embedded between two adjacent one-dimensional temporal convolution layers to sequentially perform one-dimensional temporal convolution with different dilation coefficients on the temporal features output by the one-dimensional temporal convolution of the previous layer. After being fused by short-circuit residual connection, it is used as the temporal convolution input, forming an improved MTGNN model that includes attention adjacency adjustment, graph convolution, gating concatenation, and temporal convolution.

[0123] In this embodiment, the output of the corresponding overheating fault warning information includes:

[0124] Read the principal component thermal evolution deviation results, extract the set of timestamps where the principal component thermal evolution deviation occurs, and generate the corresponding principal component thermal evolution deviation intensity sequence based on the magnitude of the principal component score sequence exceeding the allowable thermodynamic evolution range at the timestamp.

[0125] Read the thermal disturbance boundary violation determination result, extract the set of adjacent sensor pairs where thermal disturbance boundary violation occurred, and generate the corresponding thermal disturbance boundary violation intensity set according to the magnitude of the principal component disturbance propagation speed in each adjacent sensor pair exceeding the corresponding propagation speed upper bound.

[0126] Read the abnormal residual propagation path and abnormal source node, and for each abnormal residual propagation path, count the number of nodes in the path, the physical distance between the starting node and the ending node of the path, and the cumulative value of the residual propagation intensity of each node on the path to form a path propagation feature set.

[0127] Based on the principal component thermal evolution deviation intensity sequence, the thermal disturbance boundary breach intensity set, and the path propagation characteristic set, a comprehensive risk score is calculated according to a pre-set weighted combination relationship. The comprehensive risk score is then compared with a multi-level early warning threshold set to determine the corresponding overheating fault early warning level. Specifically, the calculation of the comprehensive risk score according to the pre-set weighted combination relationship is as follows:

[0128] The principal component thermal evolution deviation intensity sequence, thermal disturbance boundary breach intensity set, and path propagation characteristic set are read separately. The three types of indicators are processed with unified dimensions and converted into comparable standardized risk indicator values.

[0129] The weight combination relationship of three types of risk indicators is pre-defined, wherein the weight corresponding to the principal component thermal evolution deviation intensity is set to 0.4, the weight corresponding to the thermal disturbance boundary breach intensity is set to 0.35, and the weight corresponding to the path propagation characteristics is set to 0.25, and the sum of the weights is 1.

[0130] The standardized principal component thermal evolution deviation intensity, thermal disturbance boundary breach intensity, and path propagation characteristics are weighted and calculated with their corresponding weights, and the weighted results are accumulated to obtain a comprehensive risk score.

[0131] The abnormal source node is taken as the potential hot spot location. The overheating fault warning level, potential hot spot location, abnormal residual propagation path and comprehensive risk score are uniformly packaged into overheating fault warning information and output.

[0132] Example 1:

[0133] To verify the feasibility of this invention in practice, it was applied to an oil-immersed power transformer in a 110kV urban substation. This transformer operates under mixed residential and commercial load conditions, with significant daily load fluctuations during morning and evening peak hours, and the ambient temperature is significantly affected by the seasons. The equipment has been in operation for 9 years and has not experienced any obvious faults. However, during routine inspections in recent years, slight signs of aging were found at some bushing connection points, which are potential overheating risk areas.

[0134] The substation's existing monitoring system primarily relies on winding temperature, oil temperature, and bushing surface temperature for threshold alarms. However, the alarm thresholds are set too conservatively, typically triggering only when the temperature approaches or exceeds the operating limit. In previous operational experience, maintenance personnel discovered that some overheating hazards had persisted for some time before exceeding the temperature limit, but traditional methods were insufficient for effective identification, resulting in delayed warnings and unclear anomaly localization.

[0135] In this embodiment, temperature, current, voltage, and ambient temperature sensors are installed at key locations such as the top of the transformer windings, the sidewalls of the tank, the bushing roots, and the inlet and outlet of the radiator, forming a total of 14 sensor channels with a sampling period of 60 seconds. All collected data is transmitted to a computer processing platform through the station's data acquisition system.

[0136] The platform first performs time synchronization and standardization processing on the collected multi-source digital monitoring data to form a standardized multi-dimensional time-series dataset. Then, it performs principal component analysis on the standardized dataset to extract principal component feature vectors and corresponding principal component score sequences, and constructs a principal component time-series representation.

[0137] During operation, the system continuously analyzes the evolution trajectory of the principal component score sequence within a continuous time window and constructs the allowable thermodynamic evolution range of the principal component score sequence based on historical normal operating condition data. When the actual evolution trajectory exceeds this range in terms of rate of change and trend, the system generates principal component thermal evolution deviation results.

[0138] The system calculates the temporal variation relationship of principal component disturbances among various sensors based on the principal component score sequence. Combining transformer structural parameters and thermal inertia characteristics, it analyzes the propagation speed of principal component disturbances within the equipment. When the propagation speed is higher than the normal propagation characteristics of the equipment under similar operating conditions, a thermal disturbance breach determination result is generated.

[0139] The system reconstructs a standardized dataset using principal component eigenvectors to obtain the principal component reconstruction residual time series, and constructs a residual induced graph based on the physical connectivity of sensors. The principal component reconstruction residual time series is used as node input, and the residual induced graph is used as structural input, fed into an improved MTGNN model to model the propagation behavior of residuals within the device's internal structure, outputting the propagation path of abnormal residuals and the abnormal source nodes.

[0140] The system integrates the principal component thermal evolution deviation results, thermal disturbance boundary judgment results, abnormal residual propagation paths, and abnormal source nodes to determine the overheating fault early warning level and output the potential overheating hot spot locations and corresponding early warning information.

[0141] Table 1 Comparison of Key Transformer Operating Data and Early Warning Results

[0142] time Winding temperature (°C) Sleeve temperature (°C) Load rate (%) Principal component thermal evolution deviation Thermal disturbance boundary Warning Level 08-0400:40 74.8 61.2 68.5 no no none 08-0401:40 75.1 61.6 69.2 yes no Low 08-0402:10 75.4 62.0 70.1 yes yes middle 08-0403:00 75.9 62.6 71.3 yes yes middle 08-0408:20 78.2 65.8 73.5 yes yes high

[0143] As shown in Table 1, at 00:40 (08:00-04:00), the transformer winding temperature was 74.8℃, the bushing temperature was 61.2℃, and the load rate was 68.5%. All operating parameters were within the normal range. The results of the principal component thermal evolution deviation and thermal disturbance boundary violation judgments were both negative, and the system did not output any warning information, indicating that the equipment was operating stably. By 01:40, the load rate had increased slightly, and the winding and bushing temperatures had only changed slightly, still significantly lower than the traditional alarm thresholds. However, the system had detected the principal component thermal evolution deviation and issued a low-level warning, indicating that abnormal thermal evolution trends could be identified even before the temperature had increased significantly.

[0144] At 02:10, the load rate further increased to 70.1%, and the winding temperature and bushing temperature were still in a slow rising phase. Traditional threshold methods still could not trigger an alarm. However, while continuously identifying deviations in the thermal evolution of the principal components, the system detected thermal disturbance boundary breaches, and the warning level was upgraded to medium level. By 03:00, the changes in equipment operating parameters remained relatively gradual. The system maintained its judgment results on the deviations in the thermal evolution of the principal components and the thermal disturbance boundary breaches, and the warning level remained stable at medium level, demonstrating the continuity and stability of the warning results during the abnormal period.

[0145] At 08:20 on 08:04, the winding temperature rose to 78.2℃, the bushing temperature rose to 65.8℃, and the load rate reached 73.5%. The equipment's operating status was approaching the traditional alarm threshold level, and the system raised the warning level to a high level. The overall time evolution process shows that the method of this invention provided effective warnings for several hours before the temperature significantly increased, and the warning level gradually increased with the severity of the anomaly, providing maintenance personnel with sufficient time to respond in advance. This verifies the practical effectiveness of this method in early identification and warning classification of overheating faults.

[0146] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of overheating faults in power equipment based on principal component analysis, characterized in that, include: Collect multi-source digital monitoring data generated during the operation of power equipment, preprocess the multi-source digital monitoring data, and generate a standardized dataset; Principal component analysis is performed on the standardized dataset to extract principal component eigenvectors and corresponding principal component score sequences. The temporal representation of principal components is then constructed based on the principal component score sequences. Based on the principal component time series representation, the evolution trajectory of the principal component score sequence within a continuous time window is extracted, the allowable thermodynamic evolution interval of the principal component score sequence is constructed, and the evolution trajectory is compared with the allowable thermodynamic evolution interval to generate the principal component thermal evolution deviation result. Based on the principal component score sequence, the temporal variation relationship of principal component disturbance between sensors is calculated, the propagation speed of principal component disturbance is determined, and the propagation speed is compared with the upper bound of propagation speed adaptively generated based on the thermal inertia parameters of the equipment to generate the thermal disturbance boundary judgment result. The standardized dataset is reconstructed using principal component eigenvectors to obtain principal component reconstruction residuals. A residual induction graph is constructed, and the residual time series data is used as node feature input to improve the MTGNN model. The temporal propagation behavior of principal component reconstruction residuals on the residual induction graph is analyzed, and the abnormal residual propagation path and abnormal source node are output. Based on the principal component thermal evolution deviation results, thermal disturbance boundary judgment results, abnormal residual propagation paths and abnormal source nodes, the overheating fault warning level and potential overheating hot spot locations of power equipment are determined, and the corresponding overheating fault warning information is output. The generated principal component thermal evolution deviation results include: Obtain the time-series representation of the principal components, read the load rate data and ambient temperature data corresponding to the timestamps from the standardized dataset, determine the operating condition label corresponding to each continuous time window, and read the thermal inertia parameters and thermal diffusivity parameters corresponding to the power equipment from the equipment parameter library. The principal component time series representation is sliced ​​according to the window length and step size to obtain the evolution trajectory of the principal component score sequence within each window, and the change amount, change rate and change acceleration of the principal component score of adjacent sampling points are calculated to form the window thermal evolution characteristics. Based on the principal component feature vector, the absolute values ​​of the loads of temperature monitoring data, current data, voltage data, load rate data, and environmental parameter data in each principal component are extracted as weights. Combined with thermal inertia parameters and thermal diffusion characteristic parameters, the thermal inertia mapping coefficient and thermal diffusion mapping coefficient of each principal component are calculated and associated with the operating condition label. For each window and each principal component, the boundary of the rate of change and the boundary of the acceleration of change of the principal component score within the window are adaptively generated based on the thermal inertia mapping coefficient and the thermal diffusion mapping coefficient. Historical windows with consistent operating condition labels are selected from historical normal operation data. The upper and lower boundary ranges of the corresponding principal component thermal evolution characteristics are statistically analyzed. The boundary and the upper and lower boundary ranges are merged and constrained to construct the allowable thermodynamic evolution range of the principal component score sequence within the window. The evolution trajectory and window thermal evolution characteristics are compared with the allowable thermodynamic evolution range to determine the timestamp set and principal component set that exceed the allowable thermodynamic evolution range. The principal component thermal evolution deviation results are generated and the corresponding window index and timestamp index are output.

2. The method for early warning of overheating faults in power equipment based on principal component analysis according to claim 1, characterized in that, The multi-source digital monitoring data includes temperature monitoring data collected from different locations of the power equipment, current data, voltage data, and load rate data characterizing the operating status of the power equipment, as well as ambient temperature data and ambient humidity data related to the operating environment of the power equipment.

3. The method for early warning of overheating faults in power equipment based on principal component analysis according to claim 1, characterized in that, The preprocessing of multi-source digital monitoring data includes synchronizing the multi-source digital monitoring data according to a unified timestamp, and performing abnormal sampling point removal, missing data completion, and unit unification processing on the synchronized data.

4. The method for early warning of overheating faults in power equipment based on principal component analysis according to claim 1, characterized in that, The construction of the principal component temporal representation based on the principal component score sequence includes: The mean of the standardized dataset is calculated on each feature dimension, and the mean is subtracted from the corresponding feature of each time sample to obtain the centralized dataset; Based on the centralized dataset, the covariance matrix is ​​calculated on the feature dimension, and the covariance matrix is ​​decomposed to obtain the feature vector group and the corresponding feature value sorted from high to low according to the variance contribution. The number of principal components is determined based on the comparison between the cumulative variance contribution rate and the target proportion threshold, and the first few corresponding eigenvectors are selected to form a set of principal component eigenvectors. The centralized dataset is linearly projected along the set of principal component feature vectors to obtain the principal component score value corresponding to each time sample, and arranged in the order of timestamps to form the principal component score sequence. Based on the principal component score sequence, slices are generated according to the sampling time order and the continuous time window length and step size to generate a principal component time series representation containing timestamp, principal component score vector and window index.

5. The method for early warning of overheating faults in power equipment based on principal component analysis according to claim 1, characterized in that, The generated thermal disturbance boundary violation determination result includes: Obtain the principal component score sequence, the physical connection relationship between sensors and the corresponding connection distance, and the thermal inertia parameter and thermal diffusion characteristic parameter in the device parameter library; The principal component score sequence is differentially processed between adjacent sampling times to obtain the principal component perturbation sequence of each principal component at each sampling time. Based on the load values ​​of each sensor channel on each principal component in the principal component feature vector, the principal component perturbation sequence is mapped to the channel perturbation sequence corresponding to each sensor channel. For each pair of physically connected adjacent sensors, the channel perturbation sequence is read separately, and the propagation delay of the two channel perturbation sequences is determined by cross-correlation alignment. The propagation delay is the time offset that makes the correlation between the two channel perturbation sequences reach its maximum value. The propagation velocity of the principal component perturbation between adjacent sensors is calculated based on the connection distance and the propagation delay, forming a set of propagation velocities for each pair of adjacent sensors; The propagation speed upper bound is generated based on thermal inertia parameters and thermal diffusion characteristic parameters. The propagation speed upper bound is adaptively updated with the operating condition label and time window. The propagation speed set is compared with the propagation speed upper bound one by one. When the propagation speed of any adjacent sensor pair exceeds the corresponding propagation speed upper bound, the thermal disturbance boundary violation judgment result is output.

6. The method for early warning of overheating faults in power equipment based on principal component analysis according to claim 1, characterized in that, The output anomaly residual propagation path and anomaly source node include: The principal component feature vectors are called to perform linear reconstruction on the standardized dataset. The original data and the reconstructed data are subtracted time-stamp by time to obtain the principal component reconstruction residual time series. Based on the sensor layout diagram, the physical connection relationship of the nodes is determined, and the residual correlation between the nodes is calculated by combining the principal component reconstruction residual time series. The physical connection relationship and the residual correlation are fused to generate the adjacency relationship of the residual induced graph. The residual time series reconstructed from principal components is used as the node input, and the adjacency relationship of the residual induced graph is used as the structure input. These are then fed into an improved MTGNN model. The improved MTGNN model consists of several sequentially connected graph-temporal alternation blocks. Each graph-temporal alternation block sequentially includes graph convolution processing and temporal convolution processing, and is embedded with the following structure: A thermal topological attention fusion layer is inserted before graph convolution processing. The edge attention weights are calculated based on the correlation between the thermal resistance between nodes and the residual between nodes. The edge attention weights are then weighted and fused with the adjacency relationship to obtain the adjusted adjacency relationship. The adjusted adjacency relationship is then input into the graph convolution processing. A heat source sensing gating unit is connected in parallel between the graph convolution processing output and the temporal convolution processing input. The node temperature baseline, material thermal conductivity and node-to-shell distance are read to generate a gating factor. The graph convolution processing output is modulated node by node. The modulation result is spliced ​​with the graph convolution processing output to form the temporal convolution input feature. One-dimensional temporal convolution is performed on the temporal convolution input features to obtain the first temporal feature. An alternating dilation residual block is embedded between two adjacent one-dimensional temporal convolution layers. The alternating dilation residual block takes the first temporal feature as input and performs one-dimensional temporal convolution with different dilation coefficients in sequence. The two convolution features are added together and superimposed with the first temporal feature through a short-circuit residual connection to obtain the fused temporal feature. The fused temporal feature is then passed to the next figure - the temporal alternating block. After the last graph-temporal alternation block, an output layer is set up to output the node residual propagation strength sequence and the edge propagation weight sequence to determine the abnormal source node and the abnormal residual propagation path; Using principal component reconstruction of residual time series and residual induced graph as input features, historical normal operating condition samples and confirmed overheating operating condition samples are selected as training data to perform end-to-end training on the improved MTGNN model. The training output is the residual propagation intensity sequence of each node and the overheating probability sequence of each node.

7. The method for early warning of overheating faults in power equipment based on principal component analysis according to claim 1, characterized in that, The overheating fault warning information corresponding to the output includes: Read the principal component thermal evolution deviation results, extract the set of timestamps where the principal component thermal evolution deviation occurs, and generate the corresponding principal component thermal evolution deviation intensity sequence based on the magnitude of the principal component score sequence exceeding the allowable thermodynamic evolution range at the timestamp. Read the thermal disturbance boundary violation determination result, extract the set of adjacent sensor pairs where thermal disturbance boundary violation occurred, and generate the corresponding thermal disturbance boundary violation intensity set according to the magnitude of the principal component disturbance propagation speed in each adjacent sensor pair exceeding the corresponding propagation speed upper bound. Read the abnormal residual propagation path and abnormal source node, and for each abnormal residual propagation path, count the number of nodes in the path, the physical distance between the starting node and the ending node of the path, and the cumulative value of the residual propagation intensity of each node on the path to form a path propagation feature set. Based on the principal component thermal evolution deviation intensity sequence, thermal disturbance boundary breach intensity set, and path propagation characteristic set, a comprehensive risk score is calculated according to a pre-set weight combination relationship. The comprehensive risk score is then compared with a multi-level early warning threshold set to determine the corresponding overheating fault early warning level. The abnormal source node is taken as the potential hot spot location. The overheating fault warning level, potential hot spot location, abnormal residual propagation path and comprehensive risk score are uniformly packaged into overheating fault warning information and output.

Images