Intelligent power distribution network equipment state sensing and abnormity diagnosis system

Abstract

The invention discloses an intelligent power distribution network equipment state perception and abnormity diagnosis system, and the system operation process specifically comprises the following steps: collecting the operation state data of power distribution network equipment in real time, carrying out the time-space alignment and feature fusion, and generating an equipment multi-dimensional state vector; inputting a pre-constructed equipment health dynamic baseline model, and outputting a real-time health deviation degree; when the real-time health deviation degree exceeds an early warning deviation threshold value, triggering an abnormal preliminary screening mechanism, and extracting abnormal feature fragments; inputting a multi-stage diagnosis knowledge graph model, and generating an abnormal cause hypothesis set; performing confidence ranking on the abnormal cause hypothesis set, and outputting first # imgabs0 diagnosis results and corresponding confidence weights; and generating an equipment maintenance strategy instruction set according to the diagnosis result. The method has the following advantages and effects: the dynamic baseline is adaptively generated from multi-source data, and a multi-stage diagnosis framework of a physical model, a power grid rule and a historical case is fused, so that the accuracy and timeliness of anomaly diagnosis are finally improved.

Description

Technical Field

[0001] The present invention relates to the field of power grid technology, and in particular to a system for sensing and diagnosing equipment status in an intelligent distribution network. Background Art

[0002] With the development of smart grids, monitoring the operating status and diagnosing anomalies in distribution network equipment have become critical to ensuring the reliability and safety of power systems. These devices operate continuously in complex grid environments and are susceptible to various factors, including electrical load fluctuations, thermal stress accumulation, and mechanical vibration. Any anomalies can lead to equipment failure or even cascading power outages. Therefore, efficient and accurate sensing of equipment status and timely diagnosis of anomaly causes are core requirements for distribution network operations and maintenance.

[0003] In existing technologies, monitoring of distribution network equipment primarily relies on sensor networks to collect multi-source heterogeneous data, combined with offline analysis of historical data or rule-based methods for anomaly detection. However, these methods face a core problem: due to the different physical properties and inconsistent spatiotemporal scales of multi-source heterogeneous sensor data, as well as the dynamic complexity of the power grid environment, existing systems struggle to achieve effective data fusion and reliable diagnosis, resulting in low anomaly detection accuracy, high false alarm rates, and an inability to adapt to real-time changing operating conditions. Specifically, Limitations of data fusion: The data generated by multi-source sensors are heterogeneous. Existing methods often use simple weighting or statistical averaging to process this data, lacking an effective spatiotemporal alignment mechanism. This makes feature extraction insufficient and unable to form a multidimensional vector that fully reflects the device status.

[0004] Unreliable diagnostic decisions: Abnormal diagnosis often relies on a single model, ignoring the constraints of grid topology, such as the influence of upstream equipment or the coupling effects of parallel devices. This results in diagnostic results that are out of sync with actual conditions. Furthermore, the diagnostic process lacks a mechanism for quantifying confidence, which can easily lead to low-confidence assumptions. This makes it impossible to provide a reliable basis for maintenance decisions, increasing operational risks.

[0005] This problem directly restricts the intelligence level of the distribution network. There is an urgent need for a multi-level diagnostic framework that can adaptively generate dynamic baselines from multi-source data and integrate physical models, grid rules and historical cases to improve the accuracy and timeliness of abnormal diagnosis.

[0006] The system of the present invention is intended to solve this core problem. Summary of the Invention

[0007] The purpose of the present invention is to provide a smart distribution network equipment status perception and abnormality diagnosis system to solve the problems raised in the background technology.

[0008] The above technical objectives of the present invention are achieved through the following technical solutions: A smart distribution network equipment status perception and abnormality diagnosis system, the system operation process specifically includes the following steps: S100, collecting operating status data of distribution network equipment in real time through a multi-source heterogeneous sensor cluster deployed at distribution network nodes; the operating status data at least includes current waveform data, voltage sag data, partial discharge amplitude data, temperature distribution matrix data, and vibration spectrum data; S200, performing spatiotemporal alignment and feature fusion on the operating status data to generate a multi-dimensional state vector of the device; S300: Input the multi-dimensional state vector of the device into a pre-built dynamic baseline model of device health, and output a real-time health deviation; wherein the dynamic baseline model of device health is generated by learning historical normal operation data of the device through a long short-term memory network and is dynamically updated; S400: When the real-time health deviation exceeds the warning deviation threshold, triggering the abnormality initial screening mechanism to extract abnormal feature fragments; S500: Input the abnormal feature fragment into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses; wherein the multi-level diagnostic knowledge graph model integrates the device physical model, the historical fault case library and the power grid topology constraint rules; S600: Rank the confidence levels of the anomaly cause hypothesis set and output Diagnosis results and corresponding confidence weights; S700: Generate a device maintenance strategy instruction set according to the diagnosis result.

[0009] By adopting the above technical solutions, holographic perception and closed-loop intelligent diagnosis of distribution network equipment status can be achieved; through multi-source heterogeneous sensor clusters, multi-physical quantity operation data such as electrical, mechanical, and thermodynamics are synchronously collected, breaking through the limitations of single-dimensional monitoring and providing a complete data basis for equipment health assessment; real-time health deviation calculation based on the dynamic baseline model of equipment health makes anomaly detection adaptive, effectively overcoming the problems of false alarms and missed alarms caused by fixed thresholds; the anomaly initial screening mechanism ensures that diagnostic resources are accurately positioned in effective anomaly fragments, avoiding waste caused by invalid analysis; the multi-level diagnostic knowledge graph model integrates the equipment physical model, historical fault case library and power grid topology constraint rules, so that the diagnostic results have both physical authenticity and feasibility; the confidence ranking mechanism outputs a decision basis for quantifiable evaluation, and the maintenance strategy instruction set realizes closed-loop management from status perception to operation and maintenance execution; the above process significantly improves the accuracy and timeliness of anomaly diagnosis in distribution network fault warning, and can minimize the risk of equipment downtime.

[0010] A further configuration is to perform spatiotemporal alignment and feature fusion on the operating status data to generate a multi-dimensional state vector of the device, including: S210, performing synchronous time window segmentation on the current waveform data, voltage sag data, partial discharge amplitude data, and vibration spectrum data to generate time-aligned original data blocks; S220: performing a cross-modal feature extraction operation on each of the original data blocks to obtain a cross-modal feature vector; S230, binding the temperature distribution matrix data to the physical coordinates of the device, identifying the abnormal temperature hotspot area based on thermodynamic properties, and extracting the temperature feature vector of the abnormal temperature hotspot area; S240: Concatenate the cross-modal feature vector and the temperature feature vector, and bind the concatenated fusion feature vector to the timestamp of the corresponding time window to generate a multi-dimensional state vector of the device.

[0011] By adopting the above technical solutions, the spatiotemporal consistency problem of multi-source heterogeneous sensor data is solved and deep feature fusion is achieved; data asynchronous errors caused by differences in sensor sampling frequencies are eliminated through time window segmentation, ensuring the causal correlation of transient signals such as current, voltage, and vibration; cross-modal features are used to extract the implicit coupling relationship between electrical parameters and mechanical vibrations, such as the mapping law between current harmonics and specific frequency vibrations; temperature data is bound to physical coordinates to realize spatial modeling of the equipment thermal field and improve the positioning accuracy of local overheating areas; a unified state representation space is constructed through feature vector splicing, and originally isolated indicators such as temperature anomalies, partial discharges, and voltage sags are converted into fusion vectors that can be jointly analyzed, thereby greatly improving the comprehensiveness of state assessment and avoiding monitoring blind spots caused by single sensor failure.

[0012] It is further configured that the step of identifying the abnormal temperature hotspot area is: A triangular mesh model is constructed with infrared pixels on the device surface as vertices, and a region growing algorithm is used to aggregate adjacent mesh units with similar thermodynamic properties to form a physical region. Calculate the temperature gradient modulus, area change rate, and distance from the edge of the equipment structure for each physical area; When the temperature gradient modulus, area change rate, and distance from the edge of the equipment structure of a physical area exceed corresponding preset thresholds, it is marked as a temperature anomaly hotspot area and the coordinates of the area center and the thermal stress diffusion rate are output; the thermal stress diffusion rate is calculated based on the absolute change in the area between adjacent sampling time points, the time interval between infrared image acquisition, the maximum temperature rise of the temperature anomaly hotspot area relative to the baseline temperature at the current moment, and the distance between the physical area and the edge of the equipment structure.

[0013] By adopting the above technical solutions, accurate identification of abnormal temperature hotspots can be achieved; the triangular mesh model accurately fits the geometric characteristics of the heat distribution on the equipment surface, avoiding the loss of edge information caused by traditional rectangular partitioning; the region growing algorithm automatically identifies physical regions with continuous thermodynamic properties and accurately segments local temperature rises caused by internal faults; the temperature gradient modulus calculation effectively distinguishes normal heat dissipation hotspots from faulty overheating areas, eliminating interference from ambient temperature fluctuations; area change rate monitoring captures the diffusion dynamics of thermal anomalies and realizes early warning of slow-changing faults (such as insulation degradation); by calculating the distance between the physical area and the edge of the equipment structure, the sensitivity to mechanical defects such as loose connectors and casing cracks is improved; the thermal stress diffusion rate provides a key parameter for predicting the life of insulation materials and supports preventive maintenance decisions.

[0014] A further configuration is to input the multi-dimensional state vector of the device into a pre-built device health dynamic baseline model and output a real-time health deviation, including: S310, performing time dimension slicing and normalization preprocessing on the multi-dimensional state vector of the device to generate a standardized time series state matrix; S320, inputting the standardized time series state matrix into the device health dynamic baseline model, and extracting spatiotemporal feature vectors through a three-layer long short-term memory network; S330, calculating the multi-scale distance between the spatiotemporal feature vector and the dynamic health reference vector to generate a real-time health deviation degree; S340: Perform health status confidence interval verification on the real-time health deviation, and output the verified real-time health deviation.

[0015] By adopting the above technical solutions, a dynamic assessment of equipment health status is established. Time dimension slicing captures the evolution trend of state parameters, and normalization eliminates the impact of dimensional differences on the model. A three-layer long-short-term memory network deeply extracts the spatiotemporal dependence characteristics of state parameters, such as the temporal correlation between vibration spectrum changes and temperature rise. The dynamic health benchmark vector reflects the normal aging pattern of equipment in real time, avoiding the misinterpretation of natural performance degradation as failure. Multi-scale distance calculation simultaneously identifies sudden anomalies and gradual degradation (such as increased contact resistance). Confidence interval verification filters instantaneous deviations caused by measurement noise, improving the reliability of health deviations. Ultimately, the equipment health assessment has the ability to continuously evolve and adapt to changes in operating mode caused by grid load changes and equipment modifications.

[0016] It is further configured that when the real-time health deviation exceeds the warning deviation threshold, the abnormality initial screening mechanism is triggered to extract abnormal feature fragments, including: S410: When the real-time health deviation exceeds the warning deviation threshold, adaptively setting the sliding time window length based on the magnitude of the real-time health deviation; S420: Extracting candidate abnormal feature segments aligned with the sliding time window from the multidimensional state vector of the device; wherein the candidate abnormal feature segments include a cross-modal feature vector and a temperature feature vector aligned with timestamps; S430: Perform a feature enhancement operation on the candidate abnormal feature segment to generate an enhanced feature vector; wherein the feature enhancement operation includes: performing principal component analysis dimensionality reduction on the cross-modal feature vector to eliminate feature redundancy; Performing thermal stress diffusion rate weighted aggregation on the temperature characteristic vector to highlight the influence of temperature anomaly hotspot areas; S440: Compare and verify the enhanced feature vector with the health status confidence interval, and output the abnormal feature segment after the comparison and verification pass.

[0017] By adopting the above technical solutions, accurate extraction and enhanced expression of abnormal features can be achieved; the adaptive sliding time window intelligently adjusts the analysis time according to the degree of deviation: a short window is used to quickly capture high-amplitude transient anomalies (such as short circuits), and a long window is used for deep tracking of low-amplitude slowly changing anomalies (such as bearing wear); the extraction of candidate abnormal feature fragments retains the temporal and spatial correlation of the original data to ensure the traceability of subsequent diagnosis; principal component analysis reduces the dimension and eliminates redundant information in cross-modal features (such as the correlation between current waveforms and voltage sags) to highlight the core abnormal pattern; weighted aggregation of thermal stress diffusion rate amplifies the influence weight of fault hotspots to avoid the masking effect of local overheating by the average temperature; health status confidence interval verification eliminates false alarm fragments, ensures the data quality of the input diagnosis module, and ultimately significantly improves the signal-to-noise ratio of abnormal features, providing a basis for accurate diagnosis.

[0018] A further configuration is to input the abnormal feature fragment into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses, including: S510: In the physical layer of the multi-level diagnostic knowledge graph model, performing a first-level analysis on the abnormal feature fragment based on the device physical model to generate a set of physical layer diagnostic factors; including: Matching the cross-modal feature vectors in the abnormal feature segment with the device physical equations to calculate the electrical parameter deviation matrix; Inputting the temperature characteristic vector into a thermodynamic simulation module to generate a thermal stress distribution anomaly coefficient; Correlation mapping is performed between the electrical parameter deviation matrix and the thermal stress distribution anomaly coefficient to output a set of physical layer diagnostic factors; the set of physical layer diagnostic factors at least includes an insulation aging index, an abnormal value of contact resistance, and a mechanical deformation probability; S520: In the rule layer of the multi-level diagnostic knowledge graph model, second-level reasoning is performed on the physical layer diagnostic factor set based on the power grid topology constraint rules to generate a rule layer hypothesis set; including: Obtain the grid topology adjacency of the current device, extract the health status of upstream devices and the operating parameters of parallel devices; Constructing a topology constraint decision tree based on the physical layer diagnostic factor set, and adding the anomaly cause hypothesis that meets the topology constraint conditions to the rule layer hypothesis set; S530: In the case layer of the multi-level diagnostic knowledge graph model, performing a third-level verification on the rule layer hypothesis set based on a historical fault case library to generate an abnormality cause hypothesis set; including: Calculating the feature similarity between each anomaly cause hypothesis in the rule-level hypothesis set and historical fault cases; If the feature similarity is greater than the preset case matching threshold, the case verification mechanism is activated and the verified anomaly cause hypothesis is added to the anomaly cause hypothesis set; If the feature similarity does not reach the case matching threshold, manual intervention is triggered.

[0019] By adopting the above technical solutions, physical layer analysis maps abnormal characteristics into quantifiable fault factors (such as insulation aging index), establishing a connection between abnormal phenomena and physical parameters of the equipment; the thermodynamic simulation module can reflect the coupling mechanism between temperature distribution anomalies and electrical parameters, such as the positive feedback process of partial discharge and hot spot formation; rule-layer reasoning introduces grid topology constraints to automatically eliminate pseudo-equipment anomalies caused by upstream faults (such as protection malfunction caused by bus voltage drop); topology constraint decision trees identify fault transmission paths between devices, such as resonant overvoltage caused by parallel capacitor faults; case-layer verification matches real-time abnormal characteristics with the historical fault database, activating the handling experience library of similar cases; finally, through the human intervention mechanism, the system's ability to learn new faults is ensured, and the knowledge graph coverage is continuously expanded; through the above three-level diagnostic system, the results simultaneously meet the physical laws, system operation rules and engineering practice experience.

[0020] A further configuration is to sort the confidence level of the hypothesis set of the abnormal cause and output the The diagnostic results and corresponding confidence weights include: S610: Calculate a basic confidence level based on the feature similarity and treatment measure effectiveness score of each anomaly cause hypothesis in the anomaly cause hypothesis set; S620: Perform time decay correction on the basic confidence level to generate a time-corrected confidence level. S630, applying spatial constraint correction based on the grid topology adjacency relationship to generate a final confidence level; S640: Divide the final confidence level of each anomaly cause hypothesis by the sum of the final confidence levels of all anomaly cause hypotheses to obtain a normalized confidence weight. S650, sort the anomaly cause hypothesis set in descending order according to the final confidence level, and select the first The abnormal cause hypothesis is used as the diagnosis result and the corresponding normalized confidence weight output is bound.

[0021] By adopting the above technical solutions, diagnostic decision support with quantifiable evaluation is provided; the basic confidence level comprehensively considers the feature matching degree and the effectiveness of historical treatment, such as the extension rate of equipment life after repair with the same insulation aging index; time decay correction reduces the weight of obsolete cases and gives priority to recommending diagnostic solutions that have been recently verified to be effective; spatial constraint correction identifies the fault correlation of equipment clusters, such as when parallel equipment has the same vibration anomaly, it improves the credibility of mechanical looseness diagnosis; normalized confidence weights intuitively present the credibility of each diagnostic result, supporting the optimal allocation of operation and maintenance resources. The multiple diagnostic outputs avoid the decision-making risk of a single conclusion and provide alternative treatment plans for complex faults; ultimately, the diagnostic results are traceable and more reliable.

[0022] A further setting is to apply spatial constraint correction based on the grid topology adjacency relationship to generate the final confidence, including: Retrieve the real-time health deviation of the current device's direct upstream device and parallel devices within a preset time window; if the real-time health deviation of the upstream device exceeds the preset topology impact threshold, multiply the time-corrected confidence of the anomaly cause hypothesis involving electrical coupling by the spatial enhancement factor; If there are unresolved similar anomaly cause hypotheses for parallel equipment, the time-corrected confidence level of the anomaly cause hypothesis involving mechanical or thermal coupling is multiplied by the spatial enhancement factor.

[0023] By adopting the above technical solutions, the health status of upstream equipment can be monitored to identify the source of fault propagation, such as harmonic amplification of distribution lines caused by transformer winding deformation; the setting of preset topology impact thresholds can automatically enhance correlation diagnosis (such as voltage fluctuations of local equipment caused by poor contact of upstream circuit breakers), avoiding the one-sidedness of isolated diagnosis; similar anomaly detection of parallel equipment can discover systemic defects (such as synchronous degradation of a batch of disconnectors), triggering group equipment maintenance plans; spatial enhancement factors dynamically adjust confidence distribution to ensure that the diagnostic conclusions are consistent with the actual operating status of the power grid; ultimately, it can significantly improve the early warning capability of cascading faults and prevent local faults from spreading into systemic accidents.

[0024] In summary, the present invention has the following beneficial effects: A dynamic baseline is adaptively generated from multi-source data, and a multi-level diagnostic framework is integrated with physical models, grid rules, and historical cases to ultimately improve the accuracy and timeliness of anomaly diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a main flow chart of the embodiment; Figure 2 A flowchart illustrating the steps of performing spatiotemporal alignment and feature fusion on operating status data to generate a multi-dimensional state vector of a device in an embodiment; Figure 3 Schematic diagram of the process of identifying abnormal temperature hotspots in the embodiment; Figure 4 This is a flowchart of the steps of triggering the abnormality initial screening mechanism and extracting abnormal feature fragments when the real-time health deviation exceeds the warning deviation threshold in the embodiment; Figure 5 A flowchart illustrating the steps of inputting abnormal feature fragments into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses in an embodiment; Figure 6 Schematic diagram of a flow chart of steps for applying spatial constraint correction based on grid topology adjacency to generate a final confidence level in an embodiment. DETAILED DESCRIPTION

[0026] The present invention will be further described in detail below with reference to the accompanying drawings.

[0027] As attached Figure 1-6 As shown; This embodiment discloses a system for sensing and diagnosing the status of smart distribution network equipment. The system operation process specifically includes the following steps: S100, using a multi-source heterogeneous sensor cluster deployed at a distribution network node to collect operating status data of distribution network equipment in real time; the operating status data includes at least current waveform data, voltage sag data, partial discharge amplitude data, temperature distribution matrix data, and vibration spectrum data; wherein the multi-source heterogeneous sensor cluster includes a current transformer, a voltage sensor, a partial discharge detector, an infrared thermal imager, and a vibration accelerometer; In the above steps, the current transformer is configured to collect power frequency The current waveform data of the voltage sensor is configured to capture the duration Periodic voltage sag data; the partial discharge detector is configured to record pulse resolution The infrared thermal imager generates spatial resolution of partial discharge amplitude data. / Pixel temperature distribution matrix data; vibration accelerometer acquisition Vibration spectrum data.

[0028] S200, performing spatiotemporal alignment and feature fusion on the operating status data to generate a multi-dimensional state vector of the device; S300, inputting the multi-dimensional state vector of the device into a pre-built device health dynamic baseline model, and outputting a real-time health deviation; wherein the device health dynamic baseline model is generated by learning the historical normal operation data of the device through a long short-term memory network and is dynamically updated; S400: When the real-time health deviation exceeds the warning deviation threshold, the abnormality initial screening mechanism is triggered to extract abnormal feature fragments; S500, inputting the abnormal feature fragment into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses; wherein the multi-level diagnostic knowledge graph model integrates the equipment physical model, the historical fault case library and the power grid topology constraint rules; S600, sort the confidence level of the anomaly cause hypothesis set and output Diagnosis results and corresponding confidence weights; S700: Generate a device maintenance strategy instruction set according to the diagnosis result.

[0029] In one possible implementation, the operating status data is subjected to spatiotemporal alignment and feature fusion to generate a multi-dimensional device state vector, including: S210, performing synchronous time window segmentation on the current waveform data, voltage sag data, partial discharge amplitude data, and vibration spectrum data to generate time-aligned raw data blocks; wherein the time window length is dynamically adjusted to an integer multiple of the power frequency period; S220. Perform a cross-modal feature extraction operation on each original data block to obtain a cross-modal feature vector. Specifically, perform a joint wavelet packet transform on the current waveform data and the voltage sag data to extract the energy entropy ratio feature from the power frequency to 5 kHz frequency band; perform pulse density clustering on the partial discharge amplitude data to generate a discharge phase distribution spectrum and a maximum pulse gradient feature; perform resonance frequency band demodulation on the vibration spectrum data to extract the bearing characteristic frequency amplitude ratio and the mechanical looseness index; S230: Bind the temperature distribution matrix data to the physical coordinates of the device, identify the abnormal temperature hotspot area based on thermodynamic properties, and extract the temperature feature vector of the abnormal temperature hotspot area; specifically, the temperature feature vector includes the center coordinates of the abnormal temperature hotspot area, the thermal stress diffusion rate, and the maximum temperature rise of the abnormal temperature hotspot area relative to the reference temperature at the current moment; S240: Concatenate the cross-modal feature vector and the temperature feature vector, and bind the concatenated fusion feature vector to the timestamp of the corresponding time window to generate a multi-dimensional state vector of the device.

[0030] In the above steps, the time window is dynamically adjusted to , align current waveform data, voltage sag data, partial discharge amplitude data and vibration spectrum data; Combine the current waveform and voltage sag data with wavelet packet transform to extract Frequency band energy entropy ratio ; The local discharge pulse density clustering generates the discharge phase distribution spectrum, and the maximum pulse gradient reaches ; Resonance band demodulation found that the bearing characteristic frequency amplitude ratio exceeded the standard times, mechanical looseness index .

[0031] Output temperature feature vector ; After concatenating the cross-modal feature vector and the temperature feature vector .

[0032] In one possible implementation, the steps for identifying the abnormal temperature hotspot area are: A triangular mesh model is constructed with infrared pixels on the device surface as vertices, and a region growing algorithm is used to aggregate adjacent mesh units with similar thermodynamic properties to form a physical region. Calculate the temperature gradient modulus, area change rate, and distance from the edge of the equipment structure for each physical area; When the temperature gradient modulus, area change rate and distance from the edge of the equipment structure of the physical area exceed the corresponding preset thresholds, it is marked as a temperature anomaly hotspot area and the coordinates of the area center and the thermal stress diffusion rate are output; wherein the thermal stress diffusion rate is calculated based on the absolute change in the area between adjacent sampling time points, the time interval of infrared image acquisition, the maximum temperature rise of the temperature anomaly hotspot area relative to the reference temperature at the current moment and the distance between the physical area and the edge of the equipment structure. Specifically, when the physical area satisfies the temperature gradient modulus average value at the same time , area change rate , equipment structural edge distance When , it is marked as a temperature anomaly hotspot area, and the center coordinates and thermal stress diffusion rate are output; The calculation process of thermal stress diffusion rate is: ; in, Indicates the absolute change in regional area between adjacent sampling time points; Indicates the time interval for infrared image acquisition; Indicates the maximum temperature rise of the temperature anomaly hotspot area relative to the reference temperature at the current moment; Indicates the distance of the physical area from the edge of the device structure.

[0033] In the above steps, a triangular mesh model of the transformer tank surface is constructed, and its mesh side length is ; In the process of identifying the temperature anomaly hotspot area, the average value of the temperature gradient modulus is The area change rate is , the distance from the tank weld is ; Identified as a hotspot area of ​​abnormal temperature.

[0034] The thermal stress diffusion rate was calculated to be 12.3.

[0035] In one possible implementation, a multi-dimensional state vector of a device is input into a pre-built dynamic baseline model of device health, and a real-time health deviation is output, including: S310. Perform time dimension slicing and normalization preprocessing on the multidimensional state vector of the device to generate a standardized time series state matrix; wherein the time dimension slice length strictly matches the input time window of the device health dynamic baseline model, and the normalization preprocessing adopts a mean-variance normalization method based on the historical normal operation data of the device; S320, input the standardized time series state matrix into the equipment health dynamic baseline model, and extract the spatiotemporal feature vector through the three-layer long short-term memory network; wherein, the equipment health dynamic baseline model is dynamically updated through the sliding time window mechanism: each time a new A historical normal operation data sample triggers incremental learning, and the Adam optimizer is used to adjust the network weights; Update frequency parameters for the preset model; Preferably, the model update mechanism is configured as follows: incremental learning is triggered every 200 new samples, and the Adam optimizer learning rate is 0.001.

[0036] S330. Calculate the multi-scale distance between the spatiotemporal feature vector and the dynamic health baseline vector to generate a real-time health deviation. The dynamic health baseline vector is generated by the activation function of the output layer of the device health dynamic baseline model and reflects the feature distribution center of the device under historical normal operating conditions. The multi-scale distance includes a weighted fusion result of the Euclidean distance and the dynamic time warping distance, with the weight ratio dynamically configured according to the device type. S340: Perform health status confidence interval verification on the real-time health deviation and output the verified real-time health deviation. The health status confidence interval is extracted from historical normal operation data using kernel density estimation. When the real-time health deviation exceeds the confidence interval, an immediate recalibration mechanism of the device health dynamic baseline model is triggered.

[0037] After the above steps, the value of health deviation is obtained .

[0038] In one possible implementation, when the real-time health deviation exceeds the warning deviation threshold, the abnormality initial screening mechanism is triggered to extract abnormal feature fragments, including: S410. When the real-time health deviation exceeds the warning deviation threshold, the sliding time window length is adaptively set based on the magnitude of the real-time health deviation. Specifically, if the real-time health deviation is between the warning deviation threshold and the emergency alarm threshold, the sliding time window length is set to a first integer multiple of the power frequency cycle. If the real-time health deviation exceeds the emergency alarm threshold, the sliding time window length is set to a second integer multiple of the power frequency cycle. The emergency alarm threshold is greater than the warning deviation threshold. S420: Extract candidate abnormal feature segments aligned with the sliding time window from the device multidimensional state vector; wherein the candidate abnormal feature segments include a cross-modal feature vector and a temperature feature vector aligned with timestamps; S430: Perform a feature enhancement operation on the candidate abnormal feature segment to generate an enhanced feature vector; wherein the feature enhancement operation includes: The cross-modal eigenvectors are subjected to principal component analysis and dimensionality reduction to eliminate feature redundancy. Specifically, the covariance matrix of the cross-modal eigenvectors is calculated, the eigenvalues ​​and eigenvectors of the covariance matrix are solved, and then the eigenvectors are sorted in descending order according to the eigenvalues. The top eigenvectors whose cumulative contribution rate exceeds the preset contribution rate threshold are selected. principal component directions; project the cross-modal eigenvectors to the front principal component directions to generate cross-modal feature sub-vectors after dimensionality reduction; Performing weighted aggregation of thermal stress diffusion rates on the temperature feature vector to highlight the impact of the temperature anomaly hotspot area; specifically, obtaining the temperature feature vector of the temperature anomaly hotspot area; calculating a normalized weight factor based on the thermal stress diffusion rate; more specifically, dividing the thermal stress diffusion rate by the sum of the thermal stress diffusion rates of all the temperature anomaly hotspot areas to generate the normalized weight factor; retaining the center coordinates and thermal stress diffusion rates in the temperature feature vector, performing a weighted operation on the maximum temperature rise element; more specifically, multiplying the maximum temperature rise by the normalized weight factor to generate a weighted maximum temperature rise; and outputting a weighted temperature feature subvector containing the center coordinates, the thermal stress diffusion rate, and the weighted maximum temperature rise; S440: Compare and verify the enhanced feature vector with the health status confidence interval, and output the abnormal feature fragment after the comparison and verification passes. Specifically, receive the enhanced feature vector, compare and verify the enhanced feature vector with the predefined health status confidence interval; wherein the health status confidence interval is extracted from historical normal operation data and stored in the equipment health dynamic baseline model; Calculate the Mahalanobis distance from the enhanced feature vector to the center of the health state confidence interval; Determine whether the Mahalanobis distance exceeds the radius threshold of the health state confidence interval; if so, the comparison check passes and the abnormal feature segment is output; otherwise, the comparison check fails and S410 is re-executed to adjust the sliding time window length.

[0039] In the above steps, the value of health deviation , greater than the warning deviation threshold , but less than the emergency alarm threshold , the time window length is adjusted to ; The value of the normalized weight factor is calculated as , multiplying the maximum temperature rise by the normalized weight factor to obtain , and finally output the weighted temperature feature vector .

[0040] In one possible implementation, the abnormal feature fragments are input into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses, including: S510: In the physical layer of the multi-level diagnostic knowledge graph model, first-level analysis is performed on the abnormal feature fragments based on the device physical model to generate a set of physical layer diagnostic factors; including: Match the cross-modal eigenvectors in the abnormal feature segments with the device's physical equations to calculate the electrical parameter deviation matrix. Specifically, the electrical parameter deviation matrix includes the current waveform distortion rate, voltage sag duration, partial discharge amplitude growth rate, and vibration spectrum energy offset. The temperature characteristic vector is input into the thermodynamic simulation module to generate the thermal stress distribution anomaly coefficient. Specifically, the thermal stress distribution anomaly coefficient is calculated by finite element analysis based on the center coordinates of the temperature anomaly hotspot area, the thermal stress diffusion rate, and the weighted maximum temperature rise. Correlate and map the electrical parameter deviation matrix with the thermal stress distribution anomaly coefficient to output a set of physical layer diagnostic factors; the set of physical layer diagnostic factors at least includes the insulation aging index, the contact resistance anomaly value, and the mechanical deformation probability; Specifically, the generation process of the physical layer diagnostic factor set satisfies: The calculation of the insulation aging index requires correlating the partial discharge amplitude growth rate with the distance from the center coordinates of the abnormal temperature hotspot area to the equipment insulation material; The calculation of abnormal contact resistance values ​​requires simultaneous verification of the current waveform distortion rate and the voltage sag duration; The calculation of mechanical deformation probability requires coupling the phase correlation between the vibration spectrum energy offset and the thermal stress distribution anomaly coefficient.

[0041] S520: In the rule layer of the multi-level diagnostic knowledge graph model, second-level reasoning is performed on the physical layer diagnostic factor set based on the power grid topology constraint rules to generate a rule layer hypothesis set; including: Obtain the grid topology adjacency of the current device, extract the health status of upstream devices and the operating parameters of parallel devices; Based on the physical layer diagnostic factor set, a topology constraint decision tree is constructed, and the anomaly cause hypotheses that meet the topology constraint conditions are added to the rule layer hypothesis set; S530: In the case layer of the multi-level diagnostic knowledge graph model, a third-level verification is performed on the rule layer hypothesis set based on the historical fault case library to generate an abnormality cause hypothesis set; including: Calculate the feature similarity between each anomaly cause hypothesis in the rule layer hypothesis set and historical fault cases. Feature similarity is measured using weighted Euclidean distance, with the following weight allocation strategy: electrical parameter deviation matrix weight accounts for 40%, thermal stress distribution anomaly coefficient weight accounts for 30%, and topology constraint matching weight accounts for 30%. If the feature similarity is greater than the preset case matching threshold, the case verification mechanism is activated and the verified anomaly cause hypothesis is added to the anomaly cause hypothesis set; the case verification mechanism is as follows: Retrieve the effectiveness scores of disposal measures in historical fault cases, and add the hypotheses with effectiveness scores higher than the preset effectiveness threshold to the abnormal cause hypothesis set; when there are multiple disposal measure scores in historical fault cases, weight the scores by implementation time and aggregate them into cause types, and only add the cause hypotheses with aggregate scores ≥ dynamic effectiveness threshold to the abnormal cause hypothesis set, where the dynamic effectiveness threshold is dynamically adjusted according to the newness of the case library to optimize the recently effective solutions.

[0042] If the feature similarity does not reach the case matching threshold, manual intervention is triggered. Specifically, the expert intervention interface is accessed to receive manually annotated anomaly cause hypotheses and add them to the anomaly cause hypothesis set.

[0043] Preferably, the topology constraint decision tree includes but is not limited to: If the insulation aging index exceeds the preset first threshold and there is harmonic injection in the upstream equipment, then the output harmonic insulation degradation is the abnormal cause hypothesis; If the abnormal value of the contact resistance exceeds the preset second threshold value and the temperature characteristic vector of the parallel device is abnormal, the output hypothesis is that the oxidative contact is poor and the abnormality is caused; If the mechanical deformation probability exceeds a preset third threshold and the vibration spectrum energy offset coincides with the natural frequency of the equipment, resonance fatigue is output as the cause of the abnormality hypothesis.

[0044] In the above steps, the first threshold is set to , the second threshold is set to , the mechanical deformation probability is set to .

[0045] In one possible implementation, the confidence level of the anomaly cause hypothesis set is sorted and the output is The diagnostic results and corresponding confidence weights include: S610. Calculate a basic confidence level based on the feature similarity and treatment measure effectiveness score of each anomaly cause hypothesis in the anomaly cause hypothesis set. Specifically, normalize the feature similarity and treatment measure effectiveness score separately, then perform a linear weighted fusion of the feature similarity and treatment measure effectiveness score to generate the basic confidence level. ; in, Indicates the basic confidence; Represents the normalized feature similarity; represents the normalized effectiveness score of the treatment measures; and are preset weight coefficients respectively.

[0046] S620: Apply time decay correction to the basic confidence level to generate a time-corrected confidence level. Specifically, obtain the timestamps of historical fault cases associated with each anomaly cause hypothesis and calculate the number of days between the current time and the time when the case occurred. Adjust the base confidence based on a negative exponential decay function: ; in, As the basic confidence level; Correct confidence for timeliness; Indicates the number of days between the current time and the time when the case occurred; is the preset attenuation coefficient; S630, applying spatial constraint correction based on the grid topology adjacency relationship to generate a final confidence level; S640: Divide the final confidence level of each anomaly cause hypothesis by the sum of the final confidence levels of all anomaly cause hypotheses to obtain a normalized confidence weight. S650, sort the anomaly cause hypothesis set in descending order according to the final confidence level, and select the first The abnormal cause hypothesis is used as the diagnosis result and the corresponding normalized confidence weight output is bound.

[0047] In the above steps, the normalized feature similarity , the normalized effectiveness score of the treatment measures , basic confidence ; Implement time decay correction on the basic confidence to generate time-corrected confidence ;in , .

[0048] In one possible implementation, spatial constraint correction is applied based on the grid topology adjacency relationship to generate a final confidence score, including: Retrieve the real-time health deviation of the current device's direct upstream device and parallel devices within a preset time window; if the real-time health deviation of the upstream device exceeds the preset topology impact threshold, multiply the time-corrected confidence of the anomaly cause hypothesis involving electrical coupling by the spatial enhancement factor; If there are unresolved hypotheses of the same type for the anomaly in parallel, the time-corrected confidence level for the anomaly hypothesis involving mechanical or thermal coupling is multiplied by the spatial enhancement factor. It is important to emphasize that the spatial enhancement factor is greater than 1.

[0049] In the above steps, the health deviation of the upstream equipment Greater than the preset topology impact threshold , the final confidence .

[0050] S700, output the diagnosis result to generate the equipment maintenance strategy instruction set; specifically, The diagnostic results with corresponding final confidence levels are mapped to the preset maintenance knowledge base to generate targeted disposal measures.

[0051] In the above steps, ;The output diagnostic results are shown in Table 1; Table 1 Diagnosis results

[0052] Mapping to maintenance knowledge base generation instruction set: including Power outage and cleaning of insulating sleeves within 72 hours; Tighten the busbar connecting bolts; Install a vibration suppressor.

[0053] This specific embodiment is merely an explanation of the present invention and is not intended to limit the present invention. After reading this specification, those skilled in the art may make non-creative modifications to this embodiment as needed. However, as long as such modifications are within the scope of the claims of the present invention, they are protected by patent law.

Claims

1. A smart distribution network equipment status perception and abnormality diagnosis system, characterized by: The system operation process specifically includes the following steps: S100, collecting operating status data of distribution network equipment in real time through a multi-source heterogeneous sensor cluster deployed at distribution network nodes; the operating status data at least includes current waveform data, voltage sag data, partial discharge amplitude data, temperature distribution matrix data, and vibration spectrum data; S200, performing spatiotemporal alignment and feature fusion on the operating status data to generate a multi-dimensional state vector of the device; S300: Input the multi-dimensional state vector of the device into a pre-built dynamic baseline model of device health, and output a real-time health deviation; wherein the dynamic baseline model of device health is generated by learning historical normal operation data of the device through a long short-term memory network and is dynamically updated; S400: When the real-time health deviation exceeds the warning deviation threshold, triggering the abnormality initial screening mechanism to extract abnormal feature fragments; S500: Input the abnormal feature fragment into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses; wherein the multi-level diagnostic knowledge graph model integrates the device physical model, the historical fault case library and the power grid topology constraint rules; S600: Rank the confidence levels of the anomaly cause hypothesis set and output Diagnosis results and corresponding confidence weights; S700: Generate a device maintenance strategy instruction set according to the diagnosis result.

2. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 1 is characterized by: Performing spatiotemporal alignment and feature fusion on the operating status data to generate a multi-dimensional state vector of the device, including: S210, performing synchronous time window segmentation on the current waveform data, voltage sag data, partial discharge amplitude data, and vibration spectrum data to generate time-aligned original data blocks; S220: performing a cross-modal feature extraction operation on each of the original data blocks to obtain a cross-modal feature vector; S230, binding the temperature distribution matrix data to the physical coordinates of the device, identifying the abnormal temperature hotspot area based on thermodynamic properties, and extracting the temperature feature vector of the abnormal temperature hotspot area; S240: Concatenate the cross-modal feature vector and the temperature feature vector, and bind the concatenated fusion feature vector to the timestamp of the corresponding time window to generate a multi-dimensional state vector of the device.

3. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 2 is characterized by: The steps for identifying the abnormal temperature hotspot area are: A triangular mesh model is constructed with infrared pixels on the device surface as vertices, and a region growing algorithm is used to aggregate adjacent mesh units with similar thermodynamic properties to form a physical region. Calculate the temperature gradient modulus, area change rate, and distance from the edge of the equipment structure for each physical area; When the temperature gradient modulus, area change rate, and distance from the edge of the equipment structure of a physical area exceed corresponding preset thresholds, it is marked as a temperature anomaly hotspot area and the coordinates of the area center and the thermal stress diffusion rate are output; the thermal stress diffusion rate is calculated based on the absolute change in the area between adjacent sampling time points, the time interval between infrared image acquisition, the maximum temperature rise of the temperature anomaly hotspot area relative to the baseline temperature at the current moment, and the distance between the physical area and the edge of the equipment structure.

4. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 1, characterized in that: The multi-dimensional state vector of the device is input into a pre-built dynamic baseline model of device health, and the real-time health deviation is output, including: S310, performing time dimension slicing and normalization preprocessing on the multi-dimensional state vector of the device to generate a standardized time series state matrix; S320, inputting the standardized time series state matrix into the device health dynamic baseline model, and extracting spatiotemporal feature vectors through a three-layer long short-term memory network; S330, calculating the multi-scale distance between the spatiotemporal feature vector and the dynamic health reference vector to generate a real-time health deviation degree; S340: Perform health status confidence interval verification on the real-time health deviation, and output the verified real-time health deviation.

5. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 1 is characterized by: When the real-time health deviation exceeds the warning deviation threshold, the abnormality initial screening mechanism is triggered to extract abnormal feature fragments, including: S410: When the real-time health deviation exceeds the warning deviation threshold, adaptively setting the sliding time window length based on the magnitude of the real-time health deviation; S420: Extracting candidate abnormal feature segments aligned with the sliding time window from the multidimensional state vector of the device; wherein the candidate abnormal feature segments include a cross-modal feature vector and a temperature feature vector aligned with timestamps; S430: Perform a feature enhancement operation on the candidate abnormal feature segment to generate an enhanced feature vector; wherein the feature enhancement operation includes: performing principal component analysis dimensionality reduction on the cross-modal feature vector to eliminate feature redundancy; Performing thermal stress diffusion rate weighted aggregation on the temperature characteristic vector to highlight the influence of temperature anomaly hotspot areas; S440: Compare and verify the enhanced feature vector with the health status confidence interval, and output the abnormal feature segment after the comparison and verification pass.

6. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 1, characterized in that: The abnormal feature fragments are input into a multi-level diagnostic knowledge graph model to generate a set of abnormal cause hypotheses, including: S510: In the physical layer of the multi-level diagnostic knowledge graph model, performing a first-level analysis on the abnormal feature fragment based on the device physical model to generate a set of physical layer diagnostic factors; including: Matching the cross-modal feature vectors in the abnormal feature segment with the device physical equations to calculate the electrical parameter deviation matrix; Inputting the temperature characteristic vector into a thermodynamic simulation module to generate a thermal stress distribution anomaly coefficient; Correlation mapping is performed between the electrical parameter deviation matrix and the thermal stress distribution anomaly coefficient to output a set of physical layer diagnostic factors; the set of physical layer diagnostic factors at least includes an insulation aging index, an abnormal value of contact resistance, and a mechanical deformation probability; S520: In the rule layer of the multi-level diagnostic knowledge graph model, second-level reasoning is performed on the physical layer diagnostic factor set based on the power grid topology constraint rules to generate a rule layer hypothesis set; including: Obtain the grid topology adjacency of the current device, extract the health status of upstream devices and the operating parameters of parallel devices; Constructing a topology constraint decision tree based on the physical layer diagnostic factor set, and adding the anomaly cause hypothesis that meets the topology constraint conditions to the rule layer hypothesis set; S530: In the case layer of the multi-level diagnostic knowledge graph model, performing a third-level verification on the rule layer hypothesis set based on a historical fault case library to generate an abnormality cause hypothesis set; including: Calculating the feature similarity between each anomaly cause hypothesis in the rule-level hypothesis set and historical fault cases; If the feature similarity is greater than the preset case matching threshold, the case verification mechanism is activated and the verified anomaly cause hypothesis is added to the anomaly cause hypothesis set; If the feature similarity does not reach the case matching threshold, manual intervention is triggered.

7. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 1 is characterized by: Sort the confidence of the hypothesis set of abnormal causes and output the The diagnostic results and corresponding confidence weights include: S610: Calculate a basic confidence level based on the feature similarity and treatment measure effectiveness score of each anomaly cause hypothesis in the anomaly cause hypothesis set; S620: Perform time decay correction on the basic confidence level to generate a time-corrected confidence level. S630, applying spatial constraint correction based on the grid topology adjacency relationship to generate a final confidence level; S640: Divide the final confidence level of each anomaly cause hypothesis by the sum of the final confidence levels of all anomaly cause hypotheses to obtain a normalized confidence weight. S650, sort the anomaly cause hypothesis set in descending order according to the final confidence level, and select the first The abnormal cause hypothesis is used as the diagnosis result and the corresponding normalized confidence weight output is bound.

8. The intelligent distribution network equipment status perception and abnormality diagnosis system according to claim 7, characterized in that: Apply spatial constraint corrections based on the grid topology adjacency relationship to generate the final confidence level, including: Retrieve the real-time health deviation of the current device's direct upstream device and parallel devices within a preset time window; if the real-time health deviation of the upstream device exceeds the preset topology impact threshold, multiply the time-corrected confidence of the anomaly cause hypothesis involving electrical coupling by the spatial enhancement factor; If there are unresolved similar anomaly cause hypotheses for parallel equipment, the time-corrected confidence level of the anomaly cause hypothesis involving mechanical or thermal coupling is multiplied by the spatial enhancement factor.

Images