Search

CN-121999595-A - Power equipment overheat fault early warning method based on principal component analysis

CN121999595ACN 121999595 ACN121999595 ACN 121999595ACN-121999595-A

Abstract

The invention discloses a power equipment overheat fault early warning method based on principal component analysis, which relates to the technical field of power equipment state monitoring and comprises the steps of collecting multisource monitoring data, performing preprocessing and generating a standardized time sequence data set; the method comprises the steps of performing principal component analysis on standardized data, extracting score sequences and feature vectors, constructing a thermodynamic evolution interval, comparing evolution tracks to generate deviation judgment results, calculating disturbance propagation speed, dynamically modeling change relation to generate boundary breaking judgment results, reconstructing residual time sequence data, inputting improved MTGNN, identifying propagation paths and source nodes, integrating all early warning indexes, and outputting overheat grades and hot spot positions. According to the invention, the main component analysis is carried out on the multi-source operation data of the power equipment, and the graph structure time sequence modeling is combined, so that the abnormal heat evolution characteristics and the propagation behaviors of the abnormal heat evolution characteristics are subjected to joint analysis, and the early warning of the overheat fault of the power equipment and the accurate positioning of the potential overheat point are realized.

Inventors

  • LI SONGWEI
  • ZHU BO

Assignees

  • 哈尔滨理工大学

Dates

Publication Date
20260508
Application Date
20260409

Claims (8)

  1. 1. The utility model provides a power equipment overheat fault early warning method based on principal component analysis which is characterized in that the method comprises the following steps: The method comprises the steps of collecting multi-source digital monitoring data generated in the operation process of power equipment, preprocessing the multi-source digital monitoring data, and generating a standardized data set; Performing principal component analysis on the normalized dataset, extracting principal component feature vectors and corresponding principal component score sequences, and constructing a principal component timing representation based on the principal component score sequences; Based on the principal component time sequence representation, extracting an evolution track of a principal component score sequence in a continuous time window, constructing an allowable thermodynamic evolution interval of the principal component score sequence, comparing the evolution track with the allowable thermodynamic evolution interval, and generating a principal component thermal evolution deviation result; Based on the principal component score sequence, calculating a time sequence change relation of principal component disturbance among sensors, determining the propagation speed of the principal component disturbance, comparing the propagation speed with an upper bound of the propagation speed adaptively generated according to the thermal inertia parameters of the equipment, and generating a thermal disturbance boundary breaking judgment result; Reconstructing the standardized dataset by using the principal component feature vector to obtain a principal component reconstruction residual, constructing a residual induction diagram, inputting residual time sequence data as node features into an improved MTGNN model, analyzing time sequence propagation behaviors of the principal component reconstruction residual on the residual induction diagram, and outputting an abnormal residual propagation path and an abnormal source node; And determining overheat fault early warning grades and potential hot spot positions of the power equipment based on the main component heat evolution deviation result, the heat disturbance boundary breaking judgment result, the abnormal residual error propagation path and the abnormal source node, and outputting corresponding overheat fault early warning information.
  2. 2. The method for early warning of overheat faults of electrical equipment based on principal component analysis of claim 1 wherein the multi-source digital monitoring data comprises temperature monitoring data collected at different locations of the electrical equipment, current data representing the operating state of the electrical equipment, voltage data and load factor data, and environmental temperature data and environmental humidity data related to the operating environment of the electrical equipment.
  3. 3. The method for early warning of overheat faults of power equipment based on principal component analysis according to claim 1, wherein the preprocessing of the multi-source digital monitoring data comprises synchronizing the multi-source digital monitoring data according to a uniform time stamp, and performing abnormal sampling point rejection, missing data complement and dimension uniform processing on the synchronized data.
  4. 4. The method for early warning of overheat faults of electrical equipment based on principal component analysis of claim 1, wherein the constructing the principal component timing representation based on the principal component score sequence comprises: Calculating a mean value of the standardized data set on each characteristic dimension, and subtracting the mean value from the corresponding characteristic of each time sample to obtain a centralized data set; Calculating a covariance matrix on the characteristic dimension based on the centralized data set, and carrying out characteristic decomposition on the covariance matrix to obtain a characteristic vector group and corresponding characteristic values which are ordered from high to low according to the variance contribution; Determining the number of principal components according to the comparison result of the accumulated variance contribution rate and the target proportion threshold value, and selecting a plurality of corresponding feature vectors to form a principal component feature vector set; Linearly projecting the centralized data set along the principal component feature vector set to obtain principal component score values corresponding to each time sample one by one, and arranging the principal component score values according to the time stamp sequence to form a principal component score sequence; Based on the principal component score sequence, slicing is performed in accordance with the sampling time sequence in combination with the continuous time window length and step size to generate a principal component timing representation comprising a timestamp-principal component score vector-window index.
  5. 5. The method for early warning of overheat faults of electrical equipment based on principal component analysis according to claim 1, wherein the generating the principal component thermal evolution deviation result comprises: Acquiring a main component time sequence representation, reading load rate data and environment temperature data which are in one-to-one correspondence with time stamps from a standardized data set, determining a working condition label corresponding to each continuous time window, and simultaneously reading thermal inertia parameters and thermal diffusion characteristic parameters corresponding to power equipment from an equipment parameter library; Window slicing is carried out on the principal component time sequence representation according to window length and step length to obtain evolution tracks of principal component score sequences in each window, and the principal component score variable quantity, the change rate and the change acceleration of adjacent sampling points are calculated to form window thermal evolution characteristics; Based on the feature vectors of the main components, extracting absolute values of loads of temperature monitoring data, current data, voltage data, load rate data and environmental parameter data in the main components as weights, and calculating thermal inertia mapping coefficients and thermal diffusion mapping coefficients of the main components by combining thermal inertia parameters and thermal diffusion characteristic parameters, and recording the thermal inertia mapping coefficients and the thermal diffusion mapping coefficients in association with the working condition labels; aiming at each window and each main component, generating a main component score change rate boundary and a change acceleration boundary in the window in a self-adaptive manner based on a thermal inertia mapping coefficient and a thermal diffusion mapping coefficient, screening a history window with consistent working condition labels from history normal operation data, counting upper and lower boundary ranges of the thermal evolution characteristics of the corresponding main components, merging and restraining the boundary and the upper and lower boundary ranges, and constructing a allowable thermodynamic evolution interval of a main component score sequence in the window; comparing the evolution track with the window thermal evolution characteristics with the allowed thermodynamic evolution interval, determining a timestamp set and a principal component set which exceed the allowed thermodynamic evolution interval, generating a principal component thermal evolution deviation result, and outputting a corresponding window index and a timestamp index.
  6. 6. The method for early warning of overheat faults of electrical equipment based on principal component analysis according to claim 1, wherein the generating of the thermal disturbance boundary breaking determination result comprises: Acquiring a main component score sequence, acquiring a physical connection relation and a corresponding connection distance between sensors, and simultaneously acquiring a thermal inertia parameter and a thermal diffusion characteristic parameter in a device parameter library; Performing differential processing on the principal component score sequences between adjacent sampling moments to obtain principal component disturbance sequences of all principal components at each sampling moment, and mapping the principal component disturbance sequences into channel disturbance sequences corresponding to all sensor channels based on load values of all sensor channels on all principal components in principal component feature vectors; For each pair of adjacent sensors which are physically connected, respectively reading channel disturbance sequences, and determining the propagation delay of the two channel disturbance sequences in a cross-correlation alignment mode, wherein the propagation delay is the time offset for enabling the correlation of the two channel disturbance sequences to reach the maximum value; Calculating the disturbance propagation speed of the principal component between adjacent sensors based on the connection distance and the propagation delay to form a propagation speed set of each adjacent sensor pair; And generating a propagation speed upper bound based on the thermal inertia parameter and the thermal diffusion characteristic parameter, adaptively updating the propagation speed upper bound along with the working condition label and the time window, comparing the propagation speed set with the propagation speed upper bound one by one, and outputting a thermal disturbance boundary breaking judgment result when the propagation speed of any adjacent sensor pair exceeds the corresponding propagation speed upper bound.
  7. 7. The method for early warning of overheat fault of electrical equipment based on principal component analysis according to claim 1, wherein the outputting of the abnormal residual propagation path and the abnormal source node comprises: Invoking the principal component feature vector, linearly reconstructing the standardized data set, and subtracting the original data and the reconstructed data from each other according to time stamps to obtain a principal component reconstruction residual error time sequence; Determining a physical connection relation of nodes according to the sensor layout, calculating residual correlation among nodes by combining a main component reconstruction residual time sequence, and fusing the physical connection relation with the residual correlation to generate an adjacent relation of a residual induction diagram; The method comprises the steps of taking a main component reconstruction residual time sequence as a node input, taking an adjacency relation of a residual induction diagram as a structure input, and sending into a modified MTGNN model, wherein the modified MTGNN model is composed of a plurality of diagram-time sequence alternating blocks which are connected in sequence, each diagram-time sequence alternating block sequentially comprises diagram convolution processing and time sequence convolution processing, and the following structure is embedded: Before the graph rolling process, a thermal topology attention fusion layer is inserted, edge attention weights are calculated according to the correlation between thermal resistance among nodes and residual errors among nodes, the edge attention weights and the adjacent relations are subjected to weighted fusion to obtain adjusted adjacent relations, and the adjusted adjacent relations are input into the graph rolling process; A heat source sensing gating unit is connected in parallel between a graph convolution processing output and a time sequence convolution processing input, a gating factor is generated by reading a node temperature baseline, a material heat conductivity coefficient and a node-to-shell distance, the graph convolution processing output is modulated node by node, and a modulation result and the graph convolution processing output are spliced to form a time sequence convolution input characteristic; Performing one-dimensional time sequence convolution on the time sequence convolution input features to obtain first time features, embedding an alternate expansion residual block between two adjacent layers of one-dimensional time sequence convolutions, sequentially performing one-dimensional time sequence convolutions with different expansion coefficients by using the first time features as input by the alternate expansion residual block, adding the two sections of convolution features and the first time features, performing short-circuit residual connection superposition on the two sections of convolution features to obtain fusion time features, and transmitting the fusion time features to the next graph-time sequence alternate block; Setting an output layer after the last graph-time sequence alternating block, outputting a node residual error propagation intensity sequence and an edge propagation weight sequence, and determining an abnormal source node and an abnormal residual error propagation path; And (3) taking the main component reconstructed residual time sequence and the residual induction diagram as input features, selecting a historical normal working condition sample and a confirmed overheat working condition sample as training data, and performing end-to-end training on the improved MTGNN model, wherein the training output is a residual propagation intensity sequence of each node and a hot spot probability sequence of each node.
  8. 8. The method for early warning of overheat faults of power equipment based on principal component analysis according to claim 1, wherein the outputting of the corresponding overheat fault early warning information comprises: Reading a main component thermal evolution deviation result, extracting a time stamp set with main component thermal evolution deviation, and generating a corresponding main component thermal evolution deviation intensity sequence according to the amplitude of the main component score sequence exceeding a thermal power evolution allowable interval at the time stamp; reading a thermal disturbance boundary breaking judgment result, extracting adjacent sensor pair sets with thermal disturbance boundary breaking, and generating corresponding thermal disturbance boundary breaking strength sets according to the amplitude that the disturbance propagation speed of main components in each adjacent sensor pair exceeds the upper boundary of the corresponding propagation speed; reading an abnormal residual error propagation path and an abnormal source node, and counting the number of nodes contained in the path, the physical distance between a path starting node and a path ending node and the accumulated value of the residual error propagation intensity of each node on the path aiming at each abnormal residual error propagation path to form a path propagation feature set; Based on the main component thermal evolution deviation intensity sequence, the thermal disturbance boundary breaking intensity set and the path propagation characteristic set, calculating a comprehensive risk score according to a preset weight combination relation, comparing the comprehensive risk score with a multi-level early warning threshold set, and determining a corresponding overheat fault early warning level; And uniformly packaging the overheat fault early warning level, the potential hot point position, the abnormal residual error propagation path and the comprehensive risk score into overheat fault early warning information and outputting the overheat fault early warning information by taking the abnormal source node as the potential hot point position.

Description

Power equipment overheat fault early warning method based on principal component analysis Technical Field The invention relates to the technical field of power equipment state monitoring, in particular to a power equipment overheat fault early warning method based on principal component analysis. Background Along with the continuous expansion of the scale of a power system and the increasing complexity of the operation working conditions of equipment, key power equipment such as a transformer, a switch cabinet, a cable joint and the like are easy to generate local overheating phenomenon due to load fluctuation, environmental change or structural aging in the long-term operation process. In order to ensure safe and stable operation of the power system, sensors such as temperature, current and voltage are arranged at key positions of equipment in the prior art to continuously monitor the operation state of the equipment, and overheat fault early warning analysis is carried out based on collected operation data. The method has become an important technical means for monitoring the state of the power equipment in practical application. The existing overheat early warning method of the power equipment mostly depends on single sensor threshold judgment or a static model constructed based on historical statistical characteristics, for example, abnormality is identified by setting a temperature upper limit, a change rate threshold or adopting a simple time sequence analysis method. A part of methods are introduced into dimension reduction technologies such as principal component analysis and the like, and anomaly judgment is carried out after characteristic compression is carried out on multi-source monitoring data, but most of related technologies focus on numerical value deviation at a single moment or in a short time, and systematic modeling of coupling relations among multi-sensor data under complex working conditions is difficult. In an actual operating environment, an electrical equipment overheat fault often appears as a gradual diffusion process of heat in the internal structure of the equipment, the abnormal characteristics of which have been gradually revealed by joint changes of multidimensional operating data before the temperature is significantly raised. The prior art generally lacks the joint analysis capability of main component evolution process, thermal disturbance propagation behavior and abnormal source positioning, is difficult to accurately describe the formation trend and propagation path of overheat abnormality under the conditions of strong correlation of multiple sensors and rapid change of working conditions, easily causes early warning lag and higher false alarm rate, and the early warning result lacks clear abnormal propagation interpretation basis, so that the actual requirements of carrying out refined early warning and positioning on overheat faults of the power equipment are difficult to meet. Therefore, how to provide an overheat fault early warning method for power equipment based on principal component analysis is a problem that needs to be solved by those skilled in the art. Disclosure of Invention According to the method, the multi-source time sequence operation data of the power equipment are digitally processed and analyzed, key features reflecting the thermal state evolution of the equipment are extracted through the analysis of the main components, the spatial propagation and the time evolution of a main component reconstruction residual are further combined and described by combining with the time sequence modeling of a graph structure, and the comprehensive judgment of the overheat abnormality forming process, the propagation behavior and the abnormal source position of the power equipment is realized. According to the invention, by constructing the main component thermal evolution deviation judgment, the thermal disturbance propagation analysis and the abnormal residual error propagation path identification mechanism, the early warning of potential overheat risk is realized before the temperature rises, and the method has the advantages of high early warning timeliness, low false alarm rate, definite abnormal positioning and strong result interpretation, and is suitable for the refined intelligent early warning of the overheat fault of the power equipment under the complex working condition. According to the embodiment of the invention, the power equipment overheat fault early warning method based on principal component analysis comprises the following steps: The method comprises the steps of collecting multi-source digital monitoring data generated in the operation process of power equipment, preprocessing the multi-source digital monitoring data, and generating a standardized data set; Performing principal component analysis on the normalized dataset, extracting principal component feature vectors and corresponding principal component score sequences, and constructing