A power equipment health assessment method based on big data analysis

Abstract

This invention discloses a method for health assessment of power equipment based on big data analysis, belonging to the field of big data analysis technology. The method includes: collecting multi-source operating status data of the target power equipment and generating a sequence of equipment operating status vectors; constructing a comprehensive operating phase space trajectory and generating a Poincaré dynamic feature matrix; determining the boundaries of local attraction domains and the transition relationships within those domains, generating a dynamic attraction domain topology; constructing an improved BiGAN model to generate latent state drift and a 95% confidence interval; performing equipment status determination and generating target power equipment status identification results; generating a health index for the target power equipment and generating health levels and risk warning results. This invention, by introducing a Poincaré mapping and an improved BiGAN model, achieves modeling of the dynamic evolution of power equipment operating status, identification of latent degradation states, and accurate assessment of health risks.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to a method for assessing the health of power equipment based on big data analytics. Background Technology

[0002] With the continuous expansion of smart grid construction and the increasing complexity of power equipment operating environments, power equipment such as transformers, circuit breakers, cables, and rotating equipment continuously generate multi-source operating data, including current, voltage, temperature, vibration, partial discharge, and environmental parameters. Because these devices operate under conditions of load fluctuations, operating mode switching, and environmental changes, their operating states exhibit nonlinear, dynamic, and multi-factor coupling characteristics. To ensure the safe and stable operation of the power system, the industry typically analyzes equipment operating status using online monitoring, condition sensing, and health assessment technologies, and conducts equipment maintenance and operation management based on the analysis results. Therefore, equipment health status assessment based on multi-source operating data has become an important research direction in the field of intelligent operation and maintenance of power equipment.

[0003] Existing power equipment health assessment technologies typically involve collecting multi-source operational data, preprocessing the data, extracting features, constructing an operational status representation, and combining anomaly detection, status identification, and health evaluation methods to determine the equipment's condition. This approach primarily describes equipment status changes through trends in operating parameters, changes in statistical characteristics, and differences in feature spatial distribution. However, early-stage equipment degradation often involves gradual state changes, localized evolution, and dynamic migration. When operating parameters have not reached anomaly thresholds, some latent degradation information is difficult to accurately describe through static feature changes, leading to insufficient ability to identify early-stage equipment degradation.

[0004] Therefore, how to provide a method for health assessment of power equipment based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a power equipment health assessment method based on big data analysis. This invention utilizes multi-source operating state data analysis, comprehensive operating phase space modeling, Poincaré mapping analysis, dynamic attraction domain topology modeling, and an improved BiGAN model to detail the complete processing flow of power equipment health status assessment, from operating state data acquisition, dynamic evolution modeling of operating states, construction of local attraction domain structures, latent state distribution analysis, and the achievement of joint characterization of the dynamic evolution law and latent state change law of equipment operating states. It also completes the identification of latent degradation states, generation of health indices, and output of risk warnings. Compared with traditional methods based on parameter thresholds or static feature analysis, this invention can identify early degradation trends of equipment even when operating parameters are not significantly abnormal, possessing advantages such as strong dynamic state characterization capabilities, high accuracy in identifying latent degradation, and high accuracy in health assessment.

[0006] A method for assessing the health of power equipment based on big data analysis according to an embodiment of the present invention includes: Collect multi-source operating status data of the target power equipment during continuous operation, preprocess the multi-source operating status data, and generate a sequence of equipment operating status vectors; The operating state window is determined based on the equipment operating state vector sequence, a comprehensive operating phase space trajectory is constructed, a Poincaré cross section is set for rolling updates and Poincaré mapping is performed to generate a set of Poincaré mapping trajectory points, dynamic trajectory structure features are extracted, and a Poincaré dynamic feature matrix is ​​generated. Based on the Poincaré map trajectory point set, the local attraction domain boundary is determined according to the clustering region of the state vector of the crossing point in the phase space, and the attraction domain transition relationship is determined according to the crossing frequency between different local attraction domains, thus generating a dynamic attraction domain topology. Based on the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to the current running state window, an improved BiGAN model is constructed to generate the current latent state vector. The latent state drift is calculated by combining the healthy latent state baseline distribution and a 95% confidence interval is generated. Based on operating parameters, latent state drift, trajectory divergence, and attraction domain transition relationships, the equipment status is determined, and the target power equipment status identification result is generated. Based on the dynamic trajectory structure characteristics, dynamic attraction domain topology, latent state drift, and target power equipment status identification results, a health index for the target power equipment is generated, and health level and risk warning results are generated.

[0007] Optionally, the multi-source operating status data specifically includes the effective value of current, the effective value of voltage, the load rate, the winding temperature, the oil temperature, the partial discharge quantity, the vibration amplitude, the vibration frequency, the insulation resistance, the harmonic content, the gas component content, the ambient temperature, and the ambient humidity.

[0008] Optionally, the generation of the device operating state vector sequence includes: According to the unified sampling time axis, the time stamp alignment, missing sample value interpolation and completion, abnormal sample point median replacement and rated operating range normalization are performed on each operating status data in the multi-source operating status data to obtain the standard operating status data corresponding to each sampling time point; The standard operating status data corresponding to the same sampling time point are combined into a device operating status vector according to the order of the multi-source operating status data, and all device operating status vectors are arranged according to the sampling time order to generate a device operating status vector sequence.

[0009] Optionally, generating the Poincaré dynamic feature matrix includes: The load rate, partial discharge, vibration amplitude, oil temperature and harmonic content in the equipment operation status vector sequence are read, and the length of the operation status window is determined according to the load fluctuation amplitude. The comprehensive operation phase space trajectory is constructed according to the sampling time sequence. Based on the recent 30 days of health operation data of the target power equipment, a Poincaré cross section is set up and updated on a rolling basis. The state vectors of the forward crossing point of the comprehensive operation phase space trajectory crossing the Poincaré cross section and the reverse crossing point of the reverse crossing point are read and alternately arranged to form a sawtooth Poincaré mapping chain. Read the state vectors of consecutive crossing points in the sawtooth Poincaré mapping chain, map the state vector of the previous crossing point to the state vector of the next crossing point, and calculate the active power integral difference between adjacent crossing points. When the active power integral difference is within ±1% of the rated power integral value, retain the corresponding mapping edge and generate the Poincaré energy trajectory point set. The state vectors of the crossing points in the Poincaré energy trajectory point set are used as graph nodes, and the Poincaré discrete state mapping relationship between adjacent crossing point state vectors is used as directed edges to construct a graph structure Poincaré mapping graph. All crossing point state vectors are arranged in chronological order to form a 3rd order trajectory feature tensor, where the first dimension represents the trajectory feature dimension, the second dimension represents the trajectory state dimension, and the third dimension represents the mapping jump step size. Manifold convolution is performed on the 3rd order trajectory feature tensor to generate a manifold convolution Poincaré feature tensor. Construct a set of health mapping distributions based on the historical health Poincaré mapping trajectory point set, calculate the optimal transmission centroid distribution of all health mapping distributions, and calculate the optimal transmission distance between the current manifold convolutional Poincaré feature tensor distribution and the optimal transmission centroid distribution. Based on the graph structure Poincaré map, the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, trajectory bifurcation strength, and mapping edge stability are extracted, and combined with the optimal transmission distance to form the Poincaré dynamic feature matrix.

[0010] Optionally, generating the dynamic attraction domain topology includes: Read the state vectors of all crossing points in the Poincaré mapping trajectory point set, and extract the load rate, partial discharge, vibration amplitude, oil temperature and harmonic content corresponding to each crossing point state vector according to the sampling time order to form an attraction domain analysis point set; The set of points for the attraction domain analysis is projected onto the integrated operating phase space, and the integrated operating phase space is divided into phase space grids with a side length of 0.05. The proportion of crossing points in each phase space grid to the total number of crossing points is counted. Adjacent phase space grids with a proportion greater than or equal to 3% are merged into local attraction domains. The boundary of the local attraction domain is determined based on the outer grid outline of the local attraction domain. Read the local attraction domains to which the state vectors of adjacent crossing points belong in the order of sampling time. When the local attraction domains to which the state vectors of adjacent crossing points belong are different, the record of the previous local attraction domain pointing to the next local attraction domain is regarded as 1 attraction domain transition, and the crossing frequency between each local attraction domain is counted. Each local attraction domain is used as a topological node, and the transition relationship of the attraction domain with a crossing frequency greater than or equal to 5% of the total crossing frequency is used as a directed topological edge. The crossing frequency is written into the edge weight of the corresponding directed topological edge to generate a dynamic attraction domain topological structure.

[0011] Optionally, the step of calculating the latent state drift and generating a 95% confidence interval by combining the baseline distribution of the health latent state includes: An improved BiGAN model is constructed, which includes a diffusing bidirectional coding module, an optimal transport latent space module, a physical constraint generation module, and a discriminative adversarial module. The Poincaré dynamic feature matrix corresponding to the current running state window is input into the diffusion bidirectional encoding module. Diffusion perturbation processing is performed on each feature dimension of the Poincaré dynamic feature matrix to generate the diffusion Poincaré dynamic feature matrix. Feature mapping and latent state encoding processing are performed through the topology evolution encoder to generate the current latent state vector. Input the current latent state vector into the optimal transmission latent space module, read the set of latent state vectors corresponding to the historical healthy training samples, construct the healthy latent state baseline distribution, calculate the latent state offset between the current latent state vector and the healthy latent state baseline distribution, and perform distribution alignment processing to generate a latent state alignment vector. Read the density of crossing points of each local attraction domain in the dynamic attraction domain topology. When the density gradient of crossing points inside a local attraction domain is greater than a preset density threshold, perform recursive subdomain partitioning on the current local attraction domain and recalculate the attraction domain transition relationship between subdomains to generate a refined dynamic attraction domain topology. The latent state alignment vector and the refined dynamic attraction domain topology are input into the physical constraint generation module. The latent state alignment vector is then processed by a differentiable second-order optimization generator to generate the initial reconstructed Poincaré dynamic feature matrix. Read the local attraction domain boundaries, attraction domain transition relationships and crossing frequencies in the refined dynamic attraction domain topology, calculate the topology association weights, and combine them with the effective values ​​of current, voltage, winding temperature, oil temperature, partial discharge quantity and harmonic content to perform electromagnetic thermal coupling correction processing, and generate a physically reconstructed Poincaré dynamic feature matrix. The Poincaré dynamic feature matrix and the current latent state vector are combined to form the true feature combination, and the physically reconstructed Poincaré dynamic feature matrix and the latent state alignment vector are combined to form the reconstructed feature combination. These are then input into the discriminative adversarial module, and the discriminator calculates the energy difference and information entropy difference between the true feature combination and the reconstructed feature combination, respectively. Calculate the potential state drift between the current potential state vector and the healthy potential state baseline distribution, and generate a 95% confidence interval based on the healthy potential state baseline distribution. Calculate the energy difference weight and information entropy difference weight based on the current potential state drift and the current operating state window load rate. Perform weighted fusion processing on the energy difference and information entropy difference to generate a comprehensive discriminant value. The improved BiGAN model is trained by reading the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to historical healthy training samples. The comprehensive discriminant value, latent state drift, and reconstruction error combination between the real feature combination and the reconstructed feature combination are used as optimization targets. The parameters of the diffusion bidirectional coding module, the optimal transmission latent space module, the physical constraint generation module, and the discriminative adversarial module are continuously optimized. Training is stopped when the absolute value of the difference between the optimization target and the previous optimization target for 5 consecutive rounds of training is less than 0.001.

[0012] Optionally, generating the target power equipment status identification result includes: Read the multi-source operating status data and various alarm thresholds corresponding to the current operating status window, and compare them one by one to generate an operating parameter over-limit identifier; Read the latent state drift amount and the upper limit of the 95% confidence interval. When the latent state drift amount is greater than the upper limit of the 95% confidence interval for three consecutive running status windows, a latent state continuous drift indicator is generated. Read the trajectory divergence, the mean trajectory divergence of healthy training samples, and the attraction domain transition relationship in the dynamic attraction domain topology from the Poincaré dynamic feature matrix. When the trajectory divergence is greater than 1.30 times the mean trajectory divergence of healthy training samples and there is a transition edge from healthy local attraction domain to degenerate local attraction domain in the attraction domain transition relationship, generate a dynamic structure degradation label. When the operating parameter limit violation flag is 0, the latent state continuous drift flag is 1, and the power structure degradation flag is 1, it is identified as a latent degradation state. When the operating parameter limit violation flag is 1, it is identified as an explicit abnormal state. When all three flags are 0, it is identified as a healthy operating state, and the target power equipment status identification result is generated.

[0013] Optionally, the generation of health level and risk warning results includes: Read the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, and trajectory bifurcation intensity from the Poincaré dynamic feature matrix, calculate normalized values ​​with the mean of the corresponding healthy training samples, and calculate dynamic trajectory risk values ​​based on each normalized value; Based on the number of attraction domain transitions, the amount of latent state drift, and the state identification results of the target power equipment, the attraction domain transition risk value, the latent state drift risk value, and the state risk coefficient are calculated respectively. Combined with the dynamic trajectory risk value, the comprehensive health risk value is calculated, and a health index is generated. When the health index is greater than or equal to 85, a normal level and no warning result are generated. When the health index is greater than or equal to 70 and less than 85, a concern level and inspection prompt result are generated. When the health index is greater than or equal to 50 and less than 70, a latent degradation level and early warning result are generated. When the health index is less than 50, a high-risk level and shutdown maintenance warning result are generated.

[0014] The beneficial effects of this invention are: This invention proposes a power equipment health assessment method based on big data analysis. It unifies the processing of multi-source operating status data, including RMS current, RMS voltage, load factor, winding temperature, oil temperature, partial discharge, vibration amplitude, vibration frequency, insulation resistance, harmonic content, gas component content, ambient temperature, and ambient humidity, to construct a comprehensive operating phase space trajectory. Furthermore, it establishes the correlation between the dynamic evolution process of equipment operating status and latent state change process by combining the Poincaré map trajectory point set, dynamic attraction domain topology, and an improved BiGAN model. This allows the equipment operating status to be characterized not only by parameter numerical changes but also by state evolution laws, attraction domain transition relationships, and latent state change processes. Compared to traditional techniques based on single-parameter thresholds or static feature analysis, this invention improves the expressive power of operating status and reduces the impact of single-parameter fluctuations on health assessment results.

[0015] This invention further characterizes the local evolution of the operating state through the dynamic attraction domain topology and combines latent state drift analysis and dynamic structure change analysis to achieve equipment state identification. This allows for early identification of latent degradation states by utilizing latent state change trends and changes in the operating trajectory structure, even when operating parameters have not reached alarm thresholds. Simultaneously, a health index is constructed by combining dynamic trajectory risk values, attraction domain transition risk values, and latent state drift risk values ​​to output equipment health levels and risk warning results. Therefore, it possesses advantages such as strong latent degradation identification capability, high dynamic state perception capability, high accuracy of health assessment, and strong early risk warning capability. Attached Figure Description

[0016] 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: Figure 1 This is a flowchart of a power equipment health assessment method based on big data analysis proposed in this invention; Figure 2 This is a functional flowchart of the Poincaré map for a power equipment health assessment method based on big data analysis proposed in this invention. Figure 3 This is a schematic diagram of the structure of an improved BiGAN model for a power equipment health assessment method based on big data analysis proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figure 1 , Figure 2 and Figure 3 A method for health assessment of power equipment based on big data analysis, comprising: Collect multi-source operating status data of the target power equipment during continuous operation, preprocess the multi-source operating status data, and generate a sequence of equipment operating status vectors; The operating state window is determined based on the equipment operating state vector sequence, a comprehensive operating phase space trajectory is constructed, a Poincaré cross section is set for rolling updates and Poincaré mapping is performed to generate a set of Poincaré mapping trajectory points, dynamic trajectory structure features are extracted, and a Poincaré dynamic feature matrix is ​​generated. Based on the Poincaré map trajectory point set, the local attraction domain boundary is determined according to the clustering region of the state vector of the crossing point in the phase space, and the attraction domain transition relationship is determined according to the crossing frequency between different local attraction domains, thus generating a dynamic attraction domain topology. Based on the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to the current running state window, an improved BiGAN model is constructed to generate the current latent state vector. The latent state drift is calculated by combining the healthy latent state baseline distribution and a 95% confidence interval is generated. Based on operating parameters, latent state drift, trajectory divergence, and attraction domain transition relationships, the equipment status is determined, and the target power equipment status identification result is generated. Based on the dynamic trajectory structure characteristics, dynamic attraction domain topology, latent state drift, and target power equipment status identification results, a health index for the target power equipment is generated, and health level and risk warning results are generated.

[0019] In this embodiment, the multi-source operating status data specifically includes the effective value of current, the effective value of voltage, the load rate, the winding temperature, the oil temperature, the partial discharge quantity, the vibration amplitude, the vibration frequency, the insulation resistance, the harmonic content, the gas component content, the ambient temperature, and the ambient humidity.

[0020] In this embodiment, the generation of the device operating state vector sequence includes: Following a unified sampling timeline, timestamp alignment, missing sample value interpolation, median replacement of outlier sample points, and normalization to the rated operating range are performed on each operational status data point from the multi-source operational status data to obtain standard operational status data corresponding to each sampling time point, where: The standard operating status data corresponding to each sampling time point is obtained as follows: A unified sampling time axis is established according to a 1-second sampling interval. The timestamps corresponding to various operating status data are read. Data with a time difference of less than 0.5 seconds are mapped to the same sampling time. The valid sample values ​​and corresponding times before and after the missing sampling point are read. The proportion of the current time difference to the total time difference is multiplied by the difference between adjacent sample values, and then the previous valid sample value is added to obtain the missing position value. The 5 sample values ​​before and after the abnormal sampling point are read and sorted. The value at the middle position is taken to replace the abnormal sample value. The rated operating upper limit and rated operating lower limit values ​​corresponding to various operating status data are read. The rated operating lower limit value is subtracted from the current sample value and then divided by the difference between the upper and lower limits to obtain the normalized result. All normalized results at the same sampling time are arranged in the order of fields to generate standard operating status data. The standard operating status data corresponding to the same sampling time point are combined into a device operating status vector according to the order of the multi-source operating status data, and all device operating status vectors are arranged according to the sampling time order to generate a device operating status vector sequence.

[0021] In this embodiment, generating the Poincaré dynamic feature matrix includes: The load rate, partial discharge, vibration amplitude, oil temperature, and harmonic content are read from the equipment operating status vector sequence. The operating status window length is determined based on the load fluctuation amplitude. A comprehensive operating phase space trajectory is constructed according to the sampling time sequence, where: The length of the operating status window is determined based on the load fluctuation amplitude, specifically as follows: Read the load rate corresponding to 30 consecutive sampling points, calculate the absolute value of the difference between adjacent load rates, sum them up and divide by 29 to obtain the load fluctuation range. When the load fluctuation range is less than 0.05, set the operating status window length to 120 seconds. When the load fluctuation range is greater than or equal to 0.05 and less than 0.15, set the operating status window length to 60 seconds. When the load fluctuation range is greater than or equal to 0.15, set the operating status window length to 30 seconds. The integrated operational phase space trajectory is constructed according to the sampling time sequence, specifically as follows: Read the corresponding values ​​of load rate, partial discharge, vibration amplitude, oil temperature and harmonic content in the operating status window, arrange the current value, the value of the previous sampling time and the value of the previous two sampling times in sequence to form a state vector, and connect each state vector in sequence according to the sampling time to generate a comprehensive operating phase space trajectory. Based on the target power equipment's recent 30 days of healthy operation data, a Poincaré cross section is set up and updated on a rolling basis. The state vectors of the forward crossing points of the comprehensive operational phase space trajectory and the reverse crossing points of the Poincaré cross section are read and alternately arranged to form a sawtooth Poincaré mapping chain, where: The Poincaré section is set up based on the target power equipment's health operation data for the most recent 30 days, and is updated on a rolling basis. Read the load rate values ​​in the comprehensive operation phase space trajectory corresponding to the health operation data of the most recent 30 days, calculate the average value of all load rate values, take the average value as the center position of the Poincaré section, calculate the standard deviation of all load rate values, take 0.50 times the standard deviation as the half width of the section, and determine the positive and negative half width range of the section center position as the Poincaré section. When adding a new operation status window data, delete the earliest operation status window data and recalculate the section center position and the half width of the section to complete the Poincaré section update. Alternating arrangements form a sawtooth Poincaré map chain, specifically: Read the forward crossing point state vector and the reverse crossing point state vector according to the time order of the trajectory points crossing the Poincaré section. Read the first forward crossing point state vector, the first reverse crossing point state vector, the second forward crossing point state vector, and the second reverse crossing point state vector in sequence. Complete the arrangement of all crossing points in the same order to generate a zigzag Poincaré mapping chain. Read the state vectors of consecutive crossing points in the sawtooth Poincaré mapping chain, map the state vector of the previous crossing point to the state vector of the next crossing point, and calculate the active power integral difference between adjacent crossing points. When the active power integral difference is within ±1% of the rated power integral value, retain the corresponding mapping edge to generate the Poincaré energy trajectory point set, where: In a sawtooth Poincaré map chain, the state vector of consecutive crossing points refers to the state vectors of two adjacent crossing points in the sawtooth Poincaré map chain. Mapping the state vector of the previous crossing point to the state vector of the next crossing point is done as follows: Read the corresponding values ​​of load rate, partial discharge, vibration amplitude, oil temperature and harmonic content from the state vectors of two consecutive crossing points, take the corresponding value of the state vector of the previous crossing point as the starting state, take the corresponding value of the state vector of the next crossing point as the target state, and establish the connection relationship between the starting state and the target state in chronological order to generate a mapping edge. Calculate the integral difference of active power between adjacent crossing points, specifically as follows: Read the effective values ​​of current and voltage at each sampling time within the time interval between adjacent crossing points, calculate the product of the effective values ​​of current and voltage at each sampling time, sum all the product results and multiply by the 1-second sampling interval to obtain the active power integral value of the current time interval, calculate the absolute value of the difference between the active power integral values ​​of two adjacent time intervals, and generate the active power integral difference value. Rated power integral value refers to the cumulative result of multiplying the effective current value and the effective voltage value at all sampling times within a unit operating state window when the target power equipment is operating under rated conditions. The Poincaré energy trajectory point set is generated as follows: Read the active power integral difference corresponding to all mapping edges, retain the state vectors of the corresponding mapping edges and associated crossing points within ±1% of the rated power integral value, and combine the retained results to generate the Poincaré energy trajectory point set; Using the state vectors of the crossing points in the Poincaré energy trajectory point set as graph nodes, and the Poincaré discrete state mapping relationships between adjacent crossing point state vectors as directed edges, a Poincaré mapping graph is constructed, where: Constructing a graph structured Poincaré map, specifically: Read all the state vectors of the crossing points in the Poincaré energy trajectory point set, use the state vector number of each crossing point as the graph node identifier, read the corresponding mapping edge between adjacent crossing point state vectors, use the starting crossing point number of the mapping edge as the starting point of the directed edge, and use the ending crossing point number of the mapping edge as the ending point of the directed edge to generate the graph structure Poincaré mapping graph. All crossing point state vectors are arranged in chronological order to form a 3rd-order trajectory feature tensor, where the first dimension represents the trajectory feature dimension, the second dimension represents the trajectory state dimension, and the third dimension represents the mapping jump step size. Manifold convolution is then performed on the 3rd-order trajectory feature tensor to generate a manifold convolution Poincaré feature tensor, where: The state vectors of all crossing points are arranged in chronological order to form a third-order trajectory feature tensor, specifically: Read all crossing point state vectors in chronological order, and read the load rate, partial discharge, vibration amplitude, oil temperature and harmonic content values ​​in each crossing point state vector. Determine the load rate, partial discharge, vibration amplitude, oil temperature and harmonic content as trajectory feature dimensions, determine the arrangement order of consecutive crossing point state vectors as trajectory state dimensions, determine the number of crossing point intervals as mapping jump step size dimensions, and arrange and combine them according to dimensions to form a third-order trajectory feature tensor. Perform manifold convolution on the 3rd-order trajectory feature tensor, specifically: Read the current crossing point state vector and the state vectors of adjacent crossing points from the 3rd order trajectory feature tensor. Read the corresponding values ​​of the current crossing point state vector and the state vectors of adjacent crossing points in terms of load rate, partial discharge, vibration amplitude, oil temperature and harmonic content respectively. Calculate the square of the difference between the corresponding values ​​for each item. Sum the squares of the five differences and calculate the square root to obtain the Euclidean distance. Sort the Euclidean distances from smallest to largest and read the state vectors of the eight adjacent crossing points with the smallest distance. Add the corresponding values ​​of the eight adjacent crossing point state vectors one by one and divide by 8. Combine the values ​​with the current crossing point state vector to generate the manifold convolution Poincaré feature tensor. A set of health mapping distributions is constructed based on the historical health Poincaré mapping trajectory point set. The optimal transmission centroid distribution of all health mapping distributions is calculated, and the optimal transmission distance between the current manifold convolutional Poincaré feature tensor distribution and the optimal transmission centroid distribution is calculated, where: Historical health Poincaré mapping trajectory point set refers to the set of Poincaré mapping trajectory points generated under the continuous 30-day healthy operation status of the target power equipment; A health mapping distribution set is constructed based on the historical health Poincaré map trajectory point set, specifically as follows: Read all trajectory point state vectors in the historical health Poincaré mapping trajectory point set, divide the trajectory point set according to the running status window, take all trajectory point state vectors in each running status window as 1 health mapping distribution, and combine all health mapping distributions to form a health mapping distribution set. The optimal transmission centroid distribution for all health mappings is calculated as follows: Read the state vectors of the trajectory points corresponding to all health mapping distributions, calculate the absolute value of the difference between the state vectors of each trajectory point between any two health mapping distributions, sum all the absolute values ​​of the difference to form the distribution distance value, read all the distribution distance values, calculate the average result of all the distribution distance values, and determine the distribution corresponding to the average result as the optimal transmission centroid distribution. Calculate the optimal transmission distance between the distribution corresponding to the current manifold convolutional Poincaré feature tensor and the optimal transmission centroid distribution, specifically as follows: Read the state vectors of the trajectory points corresponding to the current manifold convolutional Poincaré feature tensor and the state vectors of the trajectory points corresponding to the optimal transmission centroid distribution, calculate the absolute value of the numerical difference between the corresponding trajectory points and sum them up to obtain the optimal transmission distance; Based on the graph structure of the Poincaré map, the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability region area, trajectory density entropy, trajectory bifurcation strength, and mapping edge stability are extracted, and combined with the optimal transmission distance to form the Poincaré dynamic feature matrix, where; Based on the graph structure Poincaré map, the following parameters are extracted: trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability region area, trajectory density entropy, trajectory bifurcation strength, and mapping edge stability. Read the state vectors of all crossing points in the Poincaré map of the graph structure, calculate the average values ​​of load rate, partial discharge, vibration amplitude, oil temperature and harmonic content, and combine them in this order to form the coordinates of the current trajectory center. Read the coordinates of the center of the healthy trajectory corresponding to the healthy training sample, calculate the squared difference between the current trajectory center coordinates and the center of the healthy trajectory coordinates in 5 dimensions, sum the 5 squared results and calculate the square root to generate the trajectory center offset. Read the state vector of each crossing point one by one, calculate the squared difference between the state vector of each crossing point and the coordinates of the current trajectory center in five dimensions, sum the five squared results and calculate the square root to generate the corresponding crossing point center distance, sum all crossing point center distances and divide by the number of crossing points to generate the trajectory dispersion. Read the state vectors of adjacent crossing points in chronological order, and read the corresponding values ​​of load rate, partial discharge, vibration amplitude, oil temperature and harmonic content of two adjacent crossing points respectively. Calculate the square of the difference between the corresponding values ​​one by one, and calculate the square root after summing them to generate the Euclidean distance between the current two adjacent crossing point state vectors. Then read the two adjacent sets of Euclidean distances, calculate the absolute value of the difference between the two sets of Euclidean distances, and sum all the absolute values ​​of the differences and divide by the number of differences to generate the trajectory divergence. Read the state vectors of the first crossing point and the last crossing point. Read the corresponding values ​​of the first crossing point and the last crossing point in terms of load rate, partial discharge, vibration amplitude, oil temperature and harmonic content respectively. Calculate the square of the difference between the corresponding values ​​one by one, and calculate the square root after summing them to generate the Euclidean distance between the first and last crossing points. Add the Euclidean distance to the value 1, and then divide the value 1 by the sum to generate the trajectory closure. Read the phase space grid containing the state vectors of all crossing points, determine the outermost grid boundary containing the crossing points, count the number of grids inside the outermost grid boundary, multiply the number of grids by the area of ​​a single grid, and generate the area of ​​the trajectory stability domain. The number of crossing points in each phase space grid is counted. The number of crossing points in each grid is divided by the total number of crossing points to generate the proportion of each grid. The negative of the product of the logarithmic value of each grid proportion and the grid proportion is calculated and accumulated to generate the trajectory density entropy. Read the number of outgoing edges of all graph nodes, mark the graph nodes with a number of outgoing edges greater than 1 as branch nodes, divide the number of branch nodes by the total number of graph nodes to generate the trajectory branch strength. Read the current running state window mapping edge set and the previous running state window mapping edge set, count the number of mapping edges that appear in both mapping edge sets at the same time, divide the number of mapping edges by the total number of mapping edges in the current running state window, and generate the mapping edge stability. The Poincaré dynamic feature matrix is ​​formed by combining the optimal transmission distance, specifically as follows: The trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, trajectory bifurcation strength, mapping edge stability, and optimal transmission distance are read and arranged in order to form feature vectors. The feature vectors corresponding to all running state windows are combined to generate the Poincaré dynamic feature matrix.

[0022] In this embodiment, generating the dynamic attraction domain topology includes: Read the state vectors of all crossing points in the Poincaré map trajectory point set, and extract the load rate, partial discharge, vibration amplitude, oil temperature, and harmonic content corresponding to the state vectors of each crossing point according to the sampling time order to form an attraction domain analysis point set, where: The set of points for attraction domain analysis is formed as follows: Read the load rate, partial discharge, vibration amplitude, oil temperature and harmonic content values ​​corresponding to the state vector of all crossing points. Arrange the five state values ​​corresponding to the same crossing point in order to form a state point. Combine all state points in order of sampling time to generate an attraction domain analysis point set. The analysis point set of the attraction domain is projected onto the integrated operating phase space, and the integrated operating phase space is divided into phase space grids with a side length of 0.05. The proportion of crossing points in each phase space grid to the total number of crossing points is counted. Adjacent phase space grids with a proportion greater than or equal to 3% are merged into local attraction domains. The boundary of the local attraction domain is determined based on the outer grid contour of the local attraction domain, where: The integrated operating phase space refers to the state space coordinate region composed of load rate, partial discharge quantity, vibration amplitude, oil temperature, and harmonic content. The attraction domain analysis point set is projected onto the integrated operational phase space, specifically as follows: Read the load rate, partial discharge, vibration amplitude, oil temperature and harmonic content values ​​corresponding to all state points in the attraction domain analysis point set, and write the corresponding values ​​into the corresponding coordinate positions in the comprehensive operation phase space to form the spatial distribution results of state points. The integrated operational phase space is divided into a phase space grid with a side length of 0.05, specifically as follows: Read the maximum and minimum values ​​of each dimension of the integrated phase space, calculate the difference between the maximum and minimum values ​​of each dimension, divide the difference by 0.05 to obtain the number of grid divisions in each dimension, divide the grid boundaries according to the number of grid divisions in each dimension, and generate the phase space grid; The boundary of the local attraction domain is determined based on the outer mesh profile of the local attraction domain, specifically as follows: Read all phase space grids within the local attraction domain, and read each adjacent grid one by one. When an adjacent grid does not belong to the current local attraction domain, record the corresponding outer boundary of the current grid as a boundary line segment. Combine all boundary line segments in the connection order to generate the boundary of the local attraction domain. The local attraction domains of adjacent crossing point state vectors are read sequentially according to the sampling time. When the local attraction domains of adjacent crossing point state vectors are different, the record where the previous local attraction domain points to the next local attraction domain is regarded as one attraction domain transition. The crossing frequency between each local attraction domain is counted, where: The frequency of crossings between different local attraction domains is calculated as follows: Read the local attraction domains to which the state vectors of adjacent crossing points belong in the order of sampling time. When the local attraction domains corresponding to two adjacent crossing points are different, record the transition from the previous local attraction domain to the next local attraction domain as one attraction domain transition. After reading all crossing points, count the number of records between any two local attraction domains to obtain the crossing frequency between the corresponding local attraction domains. Each local attraction domain is used as a topological node, and the transition relationships of attraction domains with a traversal frequency greater than or equal to 5% of all traversals are used as directed topological edges. The traversal frequency is written into the edge weight of the corresponding directed topological edge to generate a dynamic attraction domain topology structure, where: Generate a dynamic attraction domain topology, specifically as follows: Read all local attraction domains as topological nodes, read the traversal frequency and the total number of traversals, divide the traversal frequency by the total number of traversals to get the traversal ratio, retain the attraction domain transition relationships with a traversal ratio greater than or equal to 5% as directed topological edges, and write the corresponding traversal frequency into the weight of the directed topological edge. Combine all topological nodes and directed topological edges to generate a dynamic attraction domain topological structure.

[0023] In this embodiment, the step of calculating the latent state drift and generating a 95% confidence interval by combining the baseline distribution of the health latent state includes: An improved BiGAN model is constructed, comprising a diffusing bidirectional coding module, an optimal transport latent space module, a physical constraint generation module, and a discriminative adversarial module, wherein: The improved BiGAN model is constructed as follows: The traditional BiGAN model includes an encoder, generator, and discriminator. This improved model adds a dynamic attraction domain topology to the traditional encoder and incorporates a diffusion perturbation process, transforming the traditional encoder into a topological evolution encoder, forming a diffusion bidirectional encoding module. A new connection structure is added between the output of the traditional encoder and the input of the generator, along with a latent space distribution alignment process, forming an optimal transport latent space module. Second-order gradient updates and topological association weights are added to the traditional generator, transforming it into a differentiable second-order optimization generator, forming a physically constrained generation module. Finally, joint discrimination based on energy difference and information entropy difference is added to the traditional discriminator, improving it into a discriminative adversarial module, generating an improved BiGAN model, where: The diffusion bidirectional coding module includes: Disturbance parameter register: stores the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, trajectory bifurcation strength, mapping edge stability, and the disturbance amplitude corresponding to the optimal transmission distance; Characteristic perturbator: Performs diffusion perturbation processing on the Poincaré dynamic characteristic matrix; Topology Evolution Encoder: Stores node connectivity relationships, attraction domain transition relationships, and topological association relationships, and outputs the current latent state vector; Latent state buffer: stores the current latent state vector; The optimal transmission latent space module includes: Health latent state register: stores the set of latent state vectors corresponding to historical health training samples; Health potential state baseline distribution buffer: stores the health potential state baseline distribution; Offset Calculator: Calculates latent state offset; Distribution aligner: generates latent state alignment vectors; The physical constraint generation module includes: Latent state register: stores the latent state alignment vector; Differentiable second-order optimization generator: stores weight values, gradient values, and gradient change values, and outputs the initial reconstructed Poincaré dynamic feature matrix; Topology Association Weight Calculator: Calculates topology association weights; Electromagnetic thermal coupling corrector: performs electromagnetic thermal coupling correction processing; Reconstruction cache: stores the physically reconstructed Poincaré dynamic feature matrix; The discriminative adversarial module includes: True Feature Register: Stores true feature combinations; Reconstruction Feature Register: Stores combinations of reconstructed features; Discriminator: Stores the joint distribution information corresponding to the true feature combination and the reconstructed feature combination, and outputs the energy difference and information entropy difference; Weight register: stores energy difference weights and information entropy difference weights; Fusion Calculator: Generates comprehensive discriminant values; The Poincaré dynamic feature matrix corresponding to the current running state window is input into the diffusion bidirectional encoding module. Diffusion perturbation processing is performed on each feature dimension of the Poincaré dynamic feature matrix to generate a diffusion Poincaré dynamic feature matrix. Feature mapping and latent state encoding processing are then performed through the topological evolution encoder to generate the current latent state vector, where: The diffusion perturbation process is applied to each eigendimensional of the Poincaré dynamic eigenma, specifically as follows: Read the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, trajectory bifurcation strength, mapping edge stability, and optimal transmission distance from the Poincaré dynamic feature matrix. Calculate the average value of each feature value, then calculate the squared difference between each feature value and the average value. Sum all the squared results, divide by the number of features, and calculate the square root to obtain the standard deviation of the corresponding feature. Determine 0.10 times the standard deviation as the perturbation amplitude. Accumulate positive and negative random perturbation values ​​on the corresponding feature values ​​to generate the diffusion Poincaré dynamic feature matrix. Topology evolution encoders refer to encoding structures that introduce dynamic attraction domain topology information on the basis of traditional BiGAN encoders, and incorporate node connection relationships, attraction domain transition relationships, and topological association relationships into the feature encoding process. Feature mapping and latent state encoding are performed through a topological evolution encoder, specifically as follows: Read the feature values ​​corresponding to each node in the diffusion Poincaré dynamic feature matrix, read the topology of the dynamic attraction domain, read all the connected nodes corresponding to the current node one by one, sum the feature values ​​corresponding to all connected nodes and divide by the number of connected nodes to obtain the adjacency aggregation feature of the current node, read the original feature value and the adjacency aggregation feature value of the current node, add the two sets of values ​​according to the corresponding positions, and then divide by 2 to obtain the fused feature value, arrange all the fused feature values ​​of the nodes in order and divide them into data segments of the same length, calculate the average value of each data segment, combine all the average values ​​to generate the current latent state vector; The current latent state vector is input into the optimal transmission latent space module. The set of latent state vectors corresponding to historical healthy training samples is read, a healthy latent state baseline distribution is constructed, the latent state offset between the current latent state vector and the healthy latent state baseline distribution is calculated, and distribution alignment processing is performed to generate a latent state alignment vector, where: The set of latent state vectors corresponding to historical health training samples refers to the set of all latent state vectors output by the topology evolution encoder under the target power equipment's continuous 30-day healthy operation state. Construct a baseline distribution for health potential states, specifically as follows: Read all historical health latent state vectors, read the latent state values ​​one dimension at a time, divide the values ​​into intervals of 0.05, count the number of times the latent state value appears in each interval, divide the number of times each interval appears by the total number of samples to obtain the probability value of the corresponding interval, arrange the probability values ​​of each interval in order of interval to form a latent state probability sequence of each dimension, and then combine all latent state probability sequences in order of latent state dimensions to generate the baseline distribution of health latent states. Calculate the latent state offset between the current latent state vector and the baseline distribution of healthy latent states, specifically as follows: Read the values ​​of each dimension of the current latent state vector and the baseline distribution of the healthy latent state, calculate the absolute value of the difference between the corresponding dimension values, sum all the absolute values ​​of the difference and divide by the number of latent state dimensions to generate the latent state offset. Generate the latent state alignment vector, specifically: Read the values ​​of each dimension of the current latent state vector and the corresponding values ​​of the healthy latent state baseline distribution, calculate the difference between the corresponding dimension values, and then adjust the values ​​of each dimension of the current latent state vector to be 0.50 times the difference in the direction of the corresponding values ​​of the healthy latent state baseline distribution to generate a latent state alignment vector. Read the density of crossing points of each local attraction domain in the dynamic attraction domain topology. When the density gradient of crossing points within a local attraction domain is greater than a preset density threshold, perform recursive subdomain partitioning on the current local attraction domain and recalculate the attraction domain transition relationships between subdomains to generate a refined dynamic attraction domain topology, where: In a dynamic attraction domain topology, the crossing point density of each local attraction domain refers to the number of crossing points contained within a unit phase space grid area. The density threshold is specifically 0.30; Perform recursive subdomain partitioning on the current local attraction domain, specifically as follows: Read the state vectors of all crossing points in the current local attraction domain, re-divide the phase space grid with a side length of 0.02, count the number of crossing points in each grid, divide the number of crossing points in the grid by the total number of crossing points in the current local attraction domain to obtain the percentage value, merge adjacent grids with a percentage value greater than or equal to 5% to form a sub-attraction domain, repeat the crossing point count, percentage calculation and grid merging until the density gradient of crossing points in all sub-attraction domains is less than the density threshold; Recalculate the attraction domain transition relationships between subdomains, specifically: Read the sub-attraction domains to which the state vectors of adjacent crossing points belong in the order of sampling time. When the sub-attraction domains corresponding to adjacent crossing points are different, record the transition from the previous sub-attraction domain to the next sub-attraction domain as 1 transition. Count the number of records between any two sub-attraction domains and generate the attraction domain transition relationship between sub-attraction domains. The latent state alignment vector and the refined dynamic attraction domain topology are input into the physical constraint generation module. A differentiable second-order optimization generator performs feature reconstruction processing on the latent state alignment vector to generate an initial reconstructed Poincaré dynamic feature matrix, where: Differentiable second-order optimization generator refers to a generation structure that adds second-order gradient update and topological association information writing to the traditional BiGAN generator, and maps the latent state alignment vector to the Poincaré dynamic feature space. Feature reconstruction is performed on the latent state alignment vector using a differentiable second-order optimization generator, specifically as follows: Read the values ​​of each dimension of the latent state alignment vector, multiply the values ​​of two adjacent dimensions one by one to form the feature combination value. When performing reconstruction for the first time, initialize all weight values ​​to 1, calculate the product of the reconstruction error and the feature combination value corresponding to the current feature combination value, generate the current gradient value, calculate the difference between the current gradient value and the previous gradient value, generate the gradient change value, multiply the gradient change value by 0.10 and add it to the current weight value to obtain the updated weight. Use the updated weight to multiply all feature combination values ​​one by one and accumulate them to generate the initial reconstruction Poincaré dynamic feature matrix. The local attraction domain boundaries, attraction domain transition relationships, and crossover frequencies in the refined dynamic attraction domain topology are read. Topological association weights are calculated, and electromagnetic thermal coupling correction is performed by combining the effective values ​​of current, voltage, winding temperature, oil temperature, partial discharge, and harmonic content to generate a physically reconstructed Poincaré dynamic feature matrix, where: The topological association weights are calculated as follows: Read the local attraction domain boundary, attraction domain transition relationship and crossing frequency, calculate the proportion of the current attraction domain crossing frequency to the total number of crossings, calculate the proportion of the current attraction domain boundary length to the total boundary length, multiply the crossing frequency proportion by 0.60, multiply the boundary length proportion by 0.40, and then sum the two results to generate the topological association weight. The electromagnetic thermal coupling correction process is executed as follows: The initial reconstructed Poincaré dynamic feature matrix, current RMS value, voltage RMS value, winding temperature, oil temperature, partial discharge quantity, and harmonic content are read. The product of current RMS value and voltage RMS value is calculated to obtain the power value. The difference between winding temperature and oil temperature is calculated to obtain the temperature rise value. The cumulative result of partial discharge quantity and harmonic content is calculated. The power value is multiplied by 0.40, the temperature rise value is multiplied by 0.35, and the cumulative result of partial discharge and harmonic content is multiplied by 0.25 and then summed to obtain the coupling correction value. The coupling correction value is multiplied by the topology association weight and then added to the corresponding position of the initial reconstructed Poincaré dynamic feature matrix to generate the physical reconstructed Poincaré dynamic feature matrix. The Poincaré dynamic feature matrix and the current latent state vector are combined to form the true feature combination, and the physically reconstructed Poincaré dynamic feature matrix and the latent state alignment vector are combined to form the reconstructed feature combination. These are then input into the discriminative adversarial module. The discriminator calculates the energy difference and information entropy difference between the true feature combination and the reconstructed feature combination, respectively. A discriminator is a discriminative structure that performs joint distribution difference calculation on the true feature combination and the reconstructed feature combination; The discriminator calculates the energy difference and information entropy difference between the true feature combination and the reconstructed feature combination, respectively, as follows: Read the corresponding values ​​of each dimension of the real feature combination and the reconstructed feature combination, calculate the squared difference of the corresponding dimension values, sum all the squared results to generate the energy difference, count the frequency of each interval value in the real feature combination and the reconstructed feature combination, divide the frequency by the number of samples to obtain the probability value, calculate the product of each probability value and the logarithm of the probability value and sum them to obtain the information entropy difference. Calculate the latent state drift between the current latent state vector and the healthy latent state baseline distribution, and generate a 95% confidence interval based on the healthy latent state baseline distribution. Calculate the energy difference weight and information entropy difference weight based on the current latent state drift and the current operating state window load rate. Perform weighted fusion processing on the energy difference and information entropy difference to generate a comprehensive discriminant value, where: Calculate the latent state drift between the current latent state vector and the baseline distribution of healthy latent states, specifically as follows: Read the values ​​of each dimension of the current latent state vector and the baseline distribution of the healthy latent state, calculate the absolute value of the difference between the corresponding dimension values, sum all the absolute values ​​of the difference and divide by the number of latent state dimensions to generate the latent state drift. A 95% confidence interval is generated based on the baseline distribution of health potential states, specifically: Read the mean and standard deviation of each dimension of the baseline distribution of health potential. Calculate the mean minus 1.96 times the standard deviation as the lower limit and the mean plus 1.96 times the standard deviation as the upper limit to generate a 95% confidence interval. The energy difference weight and information entropy difference weight are calculated based on the current latent state drift and the current operating state window load rate, specifically as follows: Read the current latent state drift and the current operating state window load rate, multiply the latent state drift by 0.70, multiply the load rate by 0.30 and accumulate them to obtain the energy difference weight. Subtract the energy difference weight from 1 to generate the information entropy difference weight. The comprehensive discriminant value is generated as follows: Read the energy difference, information entropy difference, energy difference weight, and information entropy difference weight. Multiply the energy difference by the corresponding weight and the information entropy difference by the corresponding weight. Then sum the two results to generate a comprehensive discrimination value. The improved BiGAN model is trained by reading the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to historical healthy training samples. The optimization objective is to use the combined discriminant value, latent state drift, and reconstruction error combination between the true feature combination and the reconstructed feature combination as the optimization target. The parameters of the diffusion bidirectional coding module, the optimal transport latent space module, the physical constraint generation module, and the discriminative adversarial module are continuously optimized. Training stops when the absolute value of the difference between the optimization target and the previous optimization target for five consecutive training rounds is less than 0.001. The improved BiGAN model is trained as follows: Read the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to the historical healthy training samples. Input the Poincaré dynamic feature matrix and dynamic attraction domain topology into the improved BiGAN model to generate the current latent state vector, latent state alignment vector, physical reconstruction Poincaré dynamic feature matrix and comprehensive discriminant value. Read the absolute value of the numerical difference between the current latent state vector and the baseline distribution of the healthy latent state and calculate the average to generate the latent state drift. Read the squared result of the numerical difference between the corresponding positions of the Poincaré dynamic feature matrix and the physical reconstruction Poincaré dynamic feature matrix and calculate the average to generate the reconstruction error. The comprehensive discriminant value is multiplied by 0.40, the latent state drift is multiplied by 0.30, and the reconstruction error is multiplied by 0.30, and then summed to generate the optimization target. The corresponding parameters of the diffusion bidirectional coding module, the optimal transmission latent space module, the physical constraint generation module, and the discriminant adversarial module are adjusted one by one. When the optimization target decreases after parameter adjustment, the adjustment result is retained. When the absolute value of the difference between the corresponding optimization target and the previous optimization target is less than 0.001 for 5 consecutive rounds of training, training is stopped.

[0024] In this embodiment, generating the target power equipment status identification result includes: Read the multi-source operating status data and various alarm thresholds corresponding to the current operating status window, compare them item by item to generate an out-of-limit indicator for the operating parameters, where: Each alarm threshold refers to the upper limit of permissible variation of various operating parameters under normal operating conditions of the target power equipment. Specifically, the alarm threshold for RMS current is 1.10 times the rated current; the alarm threshold for RMS voltage is 1.08 times the rated voltage; the alarm threshold for load rate is 0.90; the alarm threshold for winding temperature is 95 degrees Celsius; the alarm threshold for oil temperature is 85 degrees Celsius; the alarm threshold for partial discharge is 50 picocutos; the alarm threshold for vibration amplitude is 1.20 mm; the alarm threshold for vibration frequency is 120 Hz; the alarm threshold for insulation resistance is 0.70 times the average value of the healthy training sample; the alarm threshold for harmonic content is 8.00%; the alarm threshold for gas component content is 1.50 times the average value of the healthy training sample; the alarm threshold for ambient temperature is 45 degrees Celsius; and the alarm threshold for ambient humidity is 90%. The process involves comparing each parameter to generate an out-of-limit flag, specifically: Read the current RMS value, voltage RMS value, load rate, winding temperature, oil temperature, partial discharge, vibration amplitude, vibration frequency, insulation resistance, harmonic content, gas component content, ambient temperature and ambient humidity corresponding to the current operating status window, and read the corresponding alarm threshold. Compare the current operating parameters with the corresponding alarm thresholds one by one. When any operating parameter is greater than the corresponding alarm threshold, set the operating parameter over-limit flag to 1. When all operating parameters are within the corresponding alarm threshold, set the operating parameter over-limit flag to 0. Read the latent state drift amount and the upper limit of the 95% confidence interval. When the latent state drift amount is greater than the upper limit of the 95% confidence interval for three consecutive running state windows, a latent state persistent drift indicator is generated, where: Generate latent state persistent drift markers, specifically: Read the current running status window and the corresponding latent state drift of the previous two running status windows in chronological order, and read the corresponding upper limit of the 95% confidence interval item by item. When the latent state drift of three consecutive running status windows is greater than the corresponding upper limit of the 95% confidence interval, set the latent state continuous drift flag to 1; otherwise, set the latent state continuous drift flag to 0. The trajectory divergence, the mean trajectory divergence of healthy training samples, and the attraction domain transition relationships in the dynamic attraction domain topology are read from the Poincaré dynamic feature matrix. When the trajectory divergence is greater than 1.30 times the mean trajectory divergence of healthy training samples and there is a transition edge in the attraction domain transition relationship pointing from a healthy local attraction domain to a degenerate local attraction domain, a dynamic structure degradation identifier is generated, where: The mean divergence of the trajectory of the health training samples is the average value obtained by summing the divergence values ​​of the trajectories corresponding to all health training samples and dividing by the number of health training samples. Generate dynamic structure degradation indicators, specifically: Read the current trajectory divergence and the average divergence of the healthy training sample trajectory. Calculate 1.30 times the average divergence of the healthy training sample trajectory as the divergence threshold. Read all the attraction domain transition relationships in the dynamic attraction domain topology. Count the number of transition edges from the healthy local attraction domain to the degenerate local attraction domain. When the current trajectory divergence is greater than the divergence threshold and the number of transition edges is greater than 0, set the dynamic structure degradation flag to 1; otherwise, set the dynamic structure degradation flag to 0. When the operating parameter limit violation flag is 0, the latent state continuous drift flag is 1, and the power structure degradation flag is 1, it is identified as a latent degradation state. When the operating parameter limit violation flag is 1, it is identified as an explicit abnormal state. When all three flags are 0, it is identified as a healthy operating state, and the target power equipment status identification result is generated.

[0025] In this embodiment, generating health level and risk warning results includes: The trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability region area, trajectory density entropy, and trajectory bifurcation intensity are read from the Poincaré dynamic feature matrix. Normalized values ​​are calculated by comparing these values ​​with the mean of the corresponding healthy training samples. Based on these normalized values, a dynamic trajectory risk value is calculated, where: Calculate the normalized value by comparing it with the mean of the corresponding health training samples, specifically: Read the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, and trajectory bifurcation intensity, and read the mean of the corresponding healthy training samples. Calculate the absolute value of the difference between the current feature value and the mean of the corresponding healthy training samples, and then divide it by the mean of the corresponding healthy training samples to generate the normalized value of each feature. The dynamic trajectory risk value is calculated based on each normalized value, specifically as follows: Read the normalized values ​​of trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, and trajectory bifurcation intensity. Multiply the normalized values ​​of trajectory center offset, trajectory dispersion, trajectory closure, trajectory density entropy, and trajectory bifurcation intensity by 0.15, trajectory dispersion by 0.10, trajectory divergence by 0.25, trajectory closure by 0.10, trajectory stability domain area by 0.15, trajectory density entropy by 0.10, and trajectory bifurcation intensity by 0.15. Then sum all the results to generate a dynamic trajectory risk value. Based on the number of attraction domain transitions, latent state drift, and target power equipment state identification results, the attraction domain transition risk value, latent state drift risk value, and state risk coefficient are calculated respectively. Combined with the dynamic trajectory risk value, a comprehensive health risk value is calculated, generating a health index, where: The calculation of the attraction domain transition risk value, latent state drift risk value, and state risk coefficient is as follows: Read the current operating status window's attraction domain transition count and the historical healthy sample's average attraction domain transition count. Subtract the historical healthy sample's average attraction domain transition count from the current attraction domain transition count, and then divide by the historical healthy sample's average attraction domain transition count to generate the attraction domain transition risk value. Read the current latent state drift amount and the upper limit of the 95% confidence interval. Divide the latent state drift amount by the upper limit of the 95% confidence interval to generate the latent state drift risk value. Read the target power equipment's status identification result. When the status identification result is a healthy operating state, the status risk coefficient is 0; when the status identification result is a latent degradation state, the status risk coefficient is 0.60; when the status identification result is a manifest abnormal state, the status risk coefficient is 1.00. The comprehensive health risk score is calculated by combining the dynamic trajectory risk value, specifically as follows: Read the dynamic trajectory risk value, attraction domain transition risk value, latent state drift risk value, and state risk coefficient. Multiply the dynamic trajectory risk value by 0.35, the attraction domain transition risk value by 0.20, the latent state drift risk value by 0.20, and the state risk coefficient by 0.25. Then sum all the results to generate a comprehensive health risk value. Subtract the comprehensive health risk value multiplied by 100 from 100 to generate a health index. When the health index is greater than or equal to 85, a normal level and no warning result are generated. When the health index is greater than or equal to 70 and less than 85, a concern level and inspection prompt result are generated. When the health index is greater than or equal to 50 and less than 70, a latent degradation level and early warning result are generated. When the health index is less than 50, a high-risk level and shutdown maintenance warning result are generated.

[0026] Example 1: To verify the feasibility of this invention in practice, it was applied to a health assessment scenario of an oil-immersed power transformer. The device has a rated capacity of 63,000 kVA and a rated voltage of 110 kV. The online monitoring system collects RMS current, RMS voltage, load factor, winding temperature, oil temperature, partial discharge, vibration amplitude, vibration frequency, insulation resistance, harmonic content, gas component content, ambient temperature, and ambient humidity at 1-second intervals. First, 30 consecutive days of healthy operation data of the equipment were collected, yielding a total of 2,592,000 sets of sampled data. The effective current value mainly ranged from 298 amps to 336 amps, the load rate mainly ranged from 0.43 to 0.76, the winding temperature mainly ranged from 55.2°C to 72.8°C, the oil temperature mainly ranged from 44.9°C to 61.6°C, the partial discharge mainly ranged from 6.5 picocubes to 18.9 picocubes, the vibration amplitude mainly ranged from 0.17 mm to 0.43 mm, and the harmonic content mainly ranged from 1.6% to 3.9%. Based on these healthy samples, the system trained and improved the BiGAN model. Before training, the reconstruction error was 0.184, and after training, the reconstruction error decreased to 0.037. The mean latent state drift stabilized at 0.071, with a standard deviation of 0.038. The upper limit of the 95% confidence interval was calculated to be 0.145.

[0027] During a certain continuous operation segment, it was found that the traditional monitoring interface did not generate any alarms, but the partial discharge of the equipment gradually increased from 22.8 picocubes to 34.7 picocubes, the vibration amplitude gradually increased from 0.49 mm to 0.76 mm, the harmonic content gradually increased from 4.2% to 5.8%, the winding temperature increased from 72.6 degrees Celsius to 78.4 degrees Celsius, and the oil temperature increased from 59.4 degrees Celsius to 64.1 degrees Celsius. Since the alarm thresholds for partial discharge, vibration amplitude, and harmonic content in the traditional system are 50 picocubes, 1.20 mm, and 8.00%, respectively, and none of the above parameters exceeded the limits, the traditional method still marked the equipment as being in a healthy operating state. Further examination of the state vector sequence output by the system of this invention revealed that the system had replaced the median of the two abnormal mutation sampling points in this segment, interpolated and completed the three missing sampling points, and normalized the various operating parameters to form a continuous operating state vector.

[0028] The system then reads the load rate corresponding to 30 consecutive sampling points, calculates the average absolute value of the difference between adjacent load rates to be 0.048, and therefore sets the operating state window length to 120 seconds. Within one operating state window, the system reads the load rate, partial discharge, vibration amplitude, oil temperature, and harmonic content to form the operating trajectory. The load rate increases from 0.66 to 0.68, the partial discharge increases from 22.8 picocubes to 25.1 picocubes, the vibration amplitude increases from 0.49 mm to 0.56 mm, the oil temperature increases from 59.4 degrees Celsius to 60.7 degrees Celsius, and the harmonic content increases from 4.2% to 4.6%. Based on the mean load rate of 0.61 and standard deviation of 0.08 for healthy samples, the system sets the Poincaré cross section within the load rate range of 0.57 to 0.65. After the operating trajectory crosses the cross section within this window, the system records 18 forward crossing points and 17 reverse crossing points, arranged alternately according to crossing time to form a mapping chain.

[0029] In the analysis of adjacent crossing points, the system found that the average effective current value between a set of crossing points changed from 328 Amperes to 331 Amperes, and the average effective voltage value changed from 110.2 kV to 110.0 kV. The corresponding active power integral values ​​for the two time intervals were 4337.47 kWh / s and 4369.20 kWh / s, respectively, with a difference of 31.73 kWh / s. The system read the rated power integral value of 4320.00 kWh / s and calculated its ±1% range to be 4276.80 kWh / s to 4363.20 kWh / s. It was found that the latter integral value exceeded the upper limit, so the mapping edge was deleted. A total of 34 candidate mapping edges were formed within the same window, and the system ultimately retained 29 of them, forming the Poincaré energy trajectory point set. Traditional methods at this stage still only see that the parameters have not exceeded the limits, while this invention has discovered that there is an unstable energy mapping phenomenon within the running trajectory.

[0030] In the subsequent running window, the system uses the retained crossing points as graph nodes and the state mapping relationships between crossing points as directed edges, resulting in 31 graph nodes and 27 directed edges. The system continues to calculate trajectory structure indices, obtaining trajectory center offset of 0.067, trajectory dispersion of 0.162, trajectory divergence of 0.061, trajectory closure of 0.69, trajectory stability domain area of ​​0.36, trajectory density entropy of 1.43, trajectory bifurcation strength of 0.13, and mapping edge stability of 0.68. Compared with healthy samples, the average trajectory divergence of healthy samples is 0.052, while the current trajectory divergence has increased. The system further counts the number of crossing points within the phase space grid and finds that three adjacent grids contain 7, 8, and 6 crossing points respectively, accounting for 14.58%, 16.67%, and 12.50% of all 48 crossing points, all exceeding 3%, and therefore merges them into a local attraction domain. The system continues to read the attraction domains of adjacent crossing points and finds that there are 6 transitions from healthy local attraction domains to weakly degenerate local attraction domains, accounting for 12.50% of all crossings, which exceeds the 5% retention condition. Therefore, an attraction domain transition edge in the degenerate direction is generated.

[0031] In the improved BiGAN model analysis phase, the system inputs the Poincaré dynamic feature matrix and dynamic attraction domain topology of the current window into the model. The system first adds a perturbation of 0.10 times the standard deviation to the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stable domain area, trajectory density entropy, trajectory bifurcation strength, mapping edge stability, and optimal transmission distance. Then, the topology evolution encoder reads the connection relationships of the attraction domain nodes, averages the features of adjacent nodes, and fuses them with the features of the current node to generate the current latent state vector. In three consecutive operating state windows, the latent state drift values ​​obtained by the system were 0.168, 0.181, and 0.193, respectively, all exceeding the upper limit of the 95% confidence interval of 0.145. Simultaneously, the trajectory divergence increased to 0.083, reaching 1.60 times the average divergence of 0.052 for healthy samples, and the number of degenerate direction transition edges increased to 5. The system therefore generates a latent state continuous drift flag and a power structure degradation flag. Since the flag for exceeding the operating parameter limit is still 0, the system does not determine it as an explicit anomaly, but instead identifies the equipment as a latent degradation state.

[0032] The system then generates a health assessment result. In this window, the normalized values ​​for trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability area, trajectory stability region area, trajectory density entropy, and trajectory bifurcation intensity are 0.52, resulting in a dynamic trajectory risk value of 0.51. The window shows 47 attraction domain transitions, while the average number of attraction domain transitions for healthy samples is 25, resulting in an attraction domain transition risk value of 0.88. The latent state drift is 0.193, and the upper limit of the 95% confidence interval is 0.145, resulting in a latent state drift risk value of 1.33. Since the state identification result is a latent degradation state, the state risk coefficient is set to 0.60. After comprehensive calculation, the system obtains a health index of 57.9 and generates a latent degradation level and early warning results. Traditional methods still output the health status within the same window, with the health score remaining at 88.5, and no warning is generated.

[0033] At the end of the operating segment, the partial discharge level continued to rise to 52.3 picocoos, exceeding the 50 picocoo alarm threshold for the first time, at which point the traditional method would output an abnormal alarm. At this point, the present invention had already issued a latent degradation warning when the partial discharge level was 34.7 picocoos, the vibration amplitude was 0.76 mm, and the harmonic content was 5.8%. Statistical analysis of 360 assessment windows within the same segment showed that the traditional method identified 22 abnormal windows, with a latent degradation recall rate of 19.4% and an overall accuracy of 72.8%. The present invention identified 168 healthy windows, 74 windows of concern, 95 latent degradation windows, and 23 high-risk windows, with a latent degradation recall rate of 86.7% and an overall accuracy of 90.6%, with the warning time approximately 126 minutes earlier than the traditional method. As this embodiment demonstrates, the present invention does not merely determine whether parameters exceed limits, but rather identifies latent degradation of equipment before parameters show obvious abnormalities by jointly identifying operational trajectory deviations, abnormal attraction domain transitions, and continuous latent state drift, thereby improving the early identification capability and warning accuracy of power equipment health assessment.

[0034] 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 health assessment of power equipment based on big data analysis, characterized in that, include: Collect multi-source operating status data of the target power equipment during continuous operation, preprocess the multi-source operating status data, and generate a sequence of equipment operating status vectors; The operating state window is determined based on the equipment operating state vector sequence, a comprehensive operating phase space trajectory is constructed, a Poincaré cross section is set for rolling updates and Poincaré mapping is performed to generate a set of Poincaré mapping trajectory points, dynamic trajectory structure features are extracted, and a Poincaré dynamic feature matrix is ​​generated. Based on the Poincaré map trajectory point set, the local attraction domain boundary is determined according to the clustering region of the state vector of the crossing point in the phase space, and the attraction domain transition relationship is determined according to the crossing frequency between different local attraction domains, thus generating a dynamic attraction domain topology. Based on the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to the current running state window, an improved BiGAN model is constructed to generate the current latent state vector. The latent state drift is calculated by combining the healthy latent state baseline distribution and a 95% confidence interval is generated. Based on operating parameters, latent state drift, trajectory divergence, and attraction domain transition relationships, the equipment status is determined, and the target power equipment status identification result is generated. Based on the dynamic trajectory structure characteristics, dynamic attraction domain topology, latent state drift, and target power equipment status identification results, a health index for the target power equipment is generated, and health level and risk warning results are generated.

2. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The multi-source operating status data specifically includes the effective value of current, the effective value of voltage, the load rate, the winding temperature, the oil temperature, the partial discharge quantity, the vibration amplitude, the vibration frequency, the insulation resistance, the harmonic content, the gas component content, the ambient temperature, and the ambient humidity.

3. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The generated device operating state vector sequence includes: According to the unified sampling time axis, the time stamp alignment, missing sample value interpolation and completion, abnormal sample point median replacement and rated operating range normalization are performed on each operating status data in the multi-source operating status data to obtain the standard operating status data corresponding to each sampling time point; The standard operating status data corresponding to the same sampling time point are combined into a device operating status vector according to the order of the multi-source operating status data, and all device operating status vectors are arranged according to the sampling time order to generate a device operating status vector sequence.

4. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The generation of the Poincaré dynamic feature matrix includes: The load rate, partial discharge, vibration amplitude, oil temperature and harmonic content in the equipment operation status vector sequence are read, and the length of the operation status window is determined according to the load fluctuation amplitude. The comprehensive operation phase space trajectory is constructed according to the sampling time sequence. Based on the recent 30 days of health operation data of the target power equipment, a Poincaré cross section is set up and updated on a rolling basis. The state vectors of the forward crossing point of the comprehensive operation phase space trajectory crossing the Poincaré cross section and the reverse crossing point of the reverse crossing point are read and alternately arranged to form a sawtooth Poincaré mapping chain. Read the state vectors of consecutive crossing points in the sawtooth Poincaré mapping chain, map the state vector of the previous crossing point to the state vector of the next crossing point, and calculate the active power integral difference between adjacent crossing points. When the active power integral difference is within ±1% of the rated power integral value, retain the corresponding mapping edge and generate the Poincaré energy trajectory point set. The state vectors of the crossing points in the Poincaré energy trajectory point set are used as graph nodes, and the Poincaré discrete state mapping relationship between adjacent crossing point state vectors is used as directed edges to construct a graph structure Poincaré mapping graph. All crossing point state vectors are arranged in chronological order to form a 3rd order trajectory feature tensor, where the first dimension represents the trajectory feature dimension, the second dimension represents the trajectory state dimension, and the third dimension represents the mapping jump step size. Manifold convolution is performed on the 3rd order trajectory feature tensor to generate a manifold convolution Poincaré feature tensor. Construct a set of health mapping distributions based on the historical health Poincaré mapping trajectory point set, calculate the optimal transmission centroid distribution of all health mapping distributions, and calculate the optimal transmission distance between the current manifold convolutional Poincaré feature tensor distribution and the optimal transmission centroid distribution. Based on the graph structure Poincaré map, the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, trajectory bifurcation strength, and mapping edge stability are extracted, and combined with the optimal transmission distance to form the Poincaré dynamic feature matrix.

5. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The generation of the dynamic attraction domain topology includes: Read the state vectors of all crossing points in the Poincaré mapping trajectory point set, and extract the load rate, partial discharge, vibration amplitude, oil temperature and harmonic content corresponding to each crossing point state vector according to the sampling time order to form an attraction domain analysis point set; The set of points for the attraction domain analysis is projected onto the integrated operating phase space, and the integrated operating phase space is divided into phase space grids with a side length of 0.

05. The proportion of crossing points in each phase space grid to the total number of crossing points is counted. Adjacent phase space grids with a proportion greater than or equal to 3% are merged into local attraction domains. The boundary of the local attraction domain is determined based on the outer grid outline of the local attraction domain. Read the local attraction domains to which the state vectors of adjacent crossing points belong in the order of sampling time. When the local attraction domains to which the state vectors of adjacent crossing points belong are different, the record of the previous local attraction domain pointing to the next local attraction domain is regarded as 1 attraction domain transition, and the crossing frequency between each local attraction domain is counted. Each local attraction domain is used as a topological node, and the transition relationship of the attraction domain with a crossing frequency greater than or equal to 5% of the total crossing frequency is used as a directed topological edge. The crossing frequency is written into the edge weight of the corresponding directed topological edge to generate a dynamic attraction domain topological structure.

6. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The calculation of latent state drift based on the baseline distribution of health latent states and the generation of a 95% confidence interval include: An improved BiGAN model is constructed, which includes a diffusing bidirectional coding module, an optimal transport latent space module, a physical constraint generation module, and a discriminative adversarial module. The Poincaré dynamic feature matrix corresponding to the current running state window is input into the diffusion bidirectional encoding module. Diffusion perturbation processing is performed on each feature dimension of the Poincaré dynamic feature matrix to generate the diffusion Poincaré dynamic feature matrix. Feature mapping and latent state encoding processing are performed through the topology evolution encoder to generate the current latent state vector. Input the current latent state vector into the optimal transmission latent space module, read the set of latent state vectors corresponding to the historical healthy training samples, construct the healthy latent state baseline distribution, calculate the latent state offset between the current latent state vector and the healthy latent state baseline distribution, and perform distribution alignment processing to generate a latent state alignment vector. Read the density of crossing points of each local attraction domain in the dynamic attraction domain topology. When the density gradient of crossing points inside a local attraction domain is greater than a preset density threshold, perform recursive subdomain partitioning on the current local attraction domain and recalculate the attraction domain transition relationship between subdomains to generate a refined dynamic attraction domain topology. The latent state alignment vector and the refined dynamic attraction domain topology are input into the physical constraint generation module. The latent state alignment vector is then processed by a differentiable second-order optimization generator to generate the initial reconstructed Poincaré dynamic feature matrix. Read the local attraction domain boundaries, attraction domain transition relationships and crossing frequencies in the refined dynamic attraction domain topology, calculate the topology association weights, and combine them with the effective values ​​of current, voltage, winding temperature, oil temperature, partial discharge quantity and harmonic content to perform electromagnetic thermal coupling correction processing, and generate a physically reconstructed Poincaré dynamic feature matrix. The Poincaré dynamic feature matrix and the current latent state vector are combined to form the true feature combination, and the physically reconstructed Poincaré dynamic feature matrix and the latent state alignment vector are combined to form the reconstructed feature combination. These are then input into the discriminative adversarial module, and the discriminator calculates the energy difference and information entropy difference between the true feature combination and the reconstructed feature combination, respectively. Calculate the potential state drift between the current potential state vector and the healthy potential state baseline distribution, and generate a 95% confidence interval based on the healthy potential state baseline distribution. Calculate the energy difference weight and information entropy difference weight based on the current potential state drift and the current operating state window load rate. Perform weighted fusion processing on the energy difference and information entropy difference to generate a comprehensive discriminant value. The improved BiGAN model is trained by reading the Poincaré dynamic feature matrix and dynamic attraction domain topology corresponding to historical healthy training samples. The comprehensive discriminant value, latent state drift, and reconstruction error combination between the real feature combination and the reconstructed feature combination are used as optimization targets. The parameters of the diffusion bidirectional coding module, the optimal transmission latent space module, the physical constraint generation module, and the discriminative adversarial module are continuously optimized. Training is stopped when the absolute value of the difference between the optimization target and the previous optimization target for 5 consecutive rounds of training is less than 0.

001.

7. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The generated target power equipment status identification result includes: Read the multi-source operating status data and various alarm thresholds corresponding to the current operating status window, and compare them one by one to generate an operating parameter over-limit identifier; Read the latent state drift amount and the upper limit of the 95% confidence interval. When the latent state drift amount is greater than the upper limit of the 95% confidence interval for three consecutive running status windows, a latent state continuous drift indicator is generated. Read the trajectory divergence, the mean trajectory divergence of healthy training samples, and the attraction domain transition relationship in the dynamic attraction domain topology from the Poincaré dynamic feature matrix. When the trajectory divergence is greater than 1.30 times the mean trajectory divergence of healthy training samples and there is a transition edge from healthy local attraction domain to degenerate local attraction domain in the attraction domain transition relationship, generate a dynamic structure degradation label. When the operating parameter limit violation flag is 0, the latent state continuous drift flag is 1, and the power structure degradation flag is 1, it is identified as a latent degradation state. When the operating parameter limit violation flag is 1, it is identified as an explicit abnormal state. When all three flags are 0, it is identified as a healthy operating state, and the target power equipment status identification result is generated.

8. The method for health assessment of power equipment based on big data analysis according to claim 1, characterized in that, The generation of health level and risk warning results includes: Read the trajectory center offset, trajectory dispersion, trajectory divergence, trajectory closure, trajectory stability domain area, trajectory density entropy, and trajectory bifurcation intensity from the Poincaré dynamic feature matrix, calculate normalized values ​​with the mean of the corresponding healthy training samples, and calculate dynamic trajectory risk values ​​based on each normalized value; Based on the number of attraction domain transitions, the amount of latent state drift, and the state identification results of the target power equipment, the attraction domain transition risk value, the latent state drift risk value, and the state risk coefficient are calculated respectively. Combined with the dynamic trajectory risk value, the comprehensive health risk value is calculated, and a health index is generated. When the health index is greater than or equal to 85, a normal level and no warning result are generated. When the health index is greater than or equal to 70 and less than 85, a concern level and inspection prompt result are generated. When the health index is greater than or equal to 50 and less than 70, a latent degradation level and early warning result are generated. When the health index is less than 50, a high-risk level and shutdown maintenance warning result are generated.

Images