Switch cabinet partial discharge intelligent monitoring system for fault identification

Abstract

The invention relates to the technical field of switch cabinet fault monitoring, and discloses a switch cabinet partial discharge intelligent monitoring system for fault identification. The system comprises a signal dynamic capture layer, a feature reconstruction mapping layer, a heterogeneous data collaboration layer and a self-adaptive diagnosis decision layer. The signal dynamic capture layer obtains a multi-source partial discharge signal based on a spatial perception mechanism, and generates a full-dimensional discharge intensity distribution spectrogram; the feature reconstruction mapping layer deploys a distributed electromagnetic sensing array according to a high-risk area, and generates an insulation defect three-dimensional positioning map through a directional detection pulse and phase analysis algorithm; the heterogeneous data collaboration layer establishes space-time association, and after time reference is aligned, comprehensive risk confidence is generated through a multi-channel fusion network; and the adaptive diagnosis decision layer converts the comprehensive risk confidence into an executable monitoring instruction set, and issues the executable monitoring instruction set to an edge computing unit through an industrial internet of things protocol. According to the system, omnibearing monitoring and accurate diagnosis of partial discharge of the switch cabinet are realized, and the fault identification and response capability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of switch cabinet fault monitoring, in particular to a switch cabinet partial discharge intelligent monitoring system for fault identification. BACKGROUND

[0002] In the operation of the power system, the switch cabinet as the key equipment of the power transmission and distribution network, its operation state is directly related to the stability and safety of the power system. The partial discharge phenomenon is an important sign of insulation deterioration of the switch cabinet, if it cannot be found and handled in time, it may cause insulation breakdown, equipment damage and even large-area power failure and other serious accidents. Therefore, the effective monitoring of the partial discharge of the switch cabinet has always been the research focus in the field of operation and maintenance of power equipment.

[0003] At present, the monitoring methods for the partial discharge of the switch cabinet mainly include single technical methods such as ultrasonic detection, ultrahigh frequency detection and infrared thermal imaging. These methods have obvious limitations in practical application: the ultrasonic detection is easily disturbed by environmental noise, and the signal recognition degree is low under complex working conditions; the ultrahigh frequency detection has strict requirements for the deployment position of the sensor, and it is difficult to fully cover the key areas inside the switch cabinet; the infrared thermal imaging can only reflect the surface temperature change, and cannot directly associate with the essential characteristics of insulation defects. In addition, the existing monitoring system mostly adopts offline data analysis mode, lacks real-time response capability, and is difficult to meet the demand of dynamic perception of equipment state of the smart grid.

[0004] With the development of the power system towards intelligence and automation, the deficiencies of the traditional monitoring technology in multi-source information fusion, defect accurate positioning and risk rapid evaluation are increasingly prominent. The partial discharge signals obtained by single sensing technology have limited dimensions, which is difficult to construct a comprehensive equipment state atlas, resulting in one-sidedness of fault identification; the existing system lacks effective spatio-temporal correlation analysis mechanism, and cannot organically integrate monitoring data from different sources, so that the positioning accuracy of insulation defects and the accuracy of risk evaluation are greatly discounted. At the same time, in the industrial Internet of Things environment, how to realize the real-time transmission and rapid processing of monitoring data and the seamless docking with the edge computing unit is also a technical problem to be solved in the field of partial discharge monitoring of switch cabinet. SUMMARY

[0005] The purpose of the present application is to provide a switch cabinet partial discharge intelligent monitoring system for fault identification to solve the problems raised in the background.

[0006] To achieve the above purpose, the present application provides a switch cabinet partial discharge intelligent monitoring system for fault identification, which comprises:

[0007] A signal dynamic capture layer obtains multi-source partial discharge signals of the switch cabinet based on a spatial perception mechanism, applies a space-time coding network to the multi-source partial discharge signals, generates a full-dimensional discharge intensity distribution spectrum, and performs high-risk area identification on the full-dimensional discharge intensity distribution spectrum.

[0008] A feature reconstruction mapping layer deploys a distributed electromagnetic sensor array according to the identified high-risk area, transmits a directional probe pulse to penetrate the internal structure of the switch cabinet, analyzes internal reflection signals by a phase analysis algorithm, and generates a three-dimensional positioning map of insulation defects of the switch cabinet.

[0009] A heterogeneous data collaboration layer establishes a space-time association between the multi-source partial discharge signals and the three-dimensional positioning map of insulation defects, aligns a time reference by using a precise time synchronization protocol, integrates the full-dimensional discharge intensity distribution spectrum and the three-dimensional positioning map of insulation defects by a multi-channel fusion network, and generates a comprehensive risk confidence.

[0010] An adaptive diagnostic decision layer converts the comprehensive risk confidence into an executable monitoring instruction set based on a multi-criteria optimization engine, and real-time issues the executable monitoring instruction set to an edge computing unit through an industrial Internet of Things protocol.

[0011] Preferably, the system further comprises:

[0012] A closed-loop optimization layer continuously tracks the state evolution of the switch cabinet after the edge computing unit executes the executable monitoring instruction set, calculates the deviation degree of the state evolution data and a preset state evolution threshold, generates a monitoring strategy performance index, and dynamically iterates the switch cabinet partial discharge monitoring strategy until the monitoring strategy performance index converges to a stable interval.

[0013] Preferably, the method for obtaining multi-source partial discharge signals of the switch cabinet based on the spatial perception mechanism comprises:

[0014] Deploying three types of sensing devices, i.e., ultra-high frequency sensors, transient ground voltage sensors, and radio frequency current transformers, at key monitoring points of the switch cabinet to collect multi-source partial discharge signals of the switch cabinet body; the multi-source partial discharge signals include ultra-high frequency electromagnetic radiation waveforms, transient ground voltage pulse sequences, and radio frequency current spectrum characteristics.

[0015] Simulating the non-uniform response mechanism of a biological perception system to dynamically configure the sensitivity and sampling frequency of the sensors; based on a switch cabinet structure model, marking cable joints, insulator surfaces, and bus bar connection points of the switch cabinet as high-sensitivity focus areas; marking the remaining structure areas of the switch cabinet as low-sensitivity background areas; dynamically constructing a sensitivity function to divide the monitoring space of the switch cabinet into high-sensitivity focus areas and low-sensitivity background areas.

[0016] The transient characteristics in the multi-source partial discharge signal are analyzed in real time through an abnormal response network, the input of the abnormal response network is the multi-source partial discharge signal, and the output is a transient characteristic active area; a focusing area range is dynamically updated according to the transient characteristic active area, if a new transient characteristic active area is detected, the new transient characteristic active area is upgraded to a high-sensitivity focusing area; if it is detected that there is no transient characteristic activity in the original high-sensitivity focusing area, the original high-sensitivity focusing area is downgraded to a low-sensitivity background area; the high-sensitivity focusing area adopts millisecond-level continuous sampling, the low-sensitivity background area adopts second-level interval sampling, the sensor working mode is dynamically regulated through a configuration instruction, and an optimized multi-source partial discharge signal is obtained.

[0017] Preferably, the method for applying the space-time coding network to the multi-source partial discharge signal to generate a full-dimensional discharge intensity distribution spectrum includes:

[0018] The kurtosis normalization processing is performed on the multi-source partial discharge signal, and the normalized multi-source partial discharge signal is integrated into a five-dimensional tensor according to a signal source dimension, the dimensions of the five-dimensional tensor include a time axis, a frequency axis, an amplitude axis, a phase axis and a sensor channel; a space-time coding network architecture is constructed, the space-time coding network architecture includes a forward feature extraction path, a reverse feature reconstruction path and a cross-scale connection;

[0019] The multi-source partial discharge signal is input into the space-time coding network architecture, in the forward feature extraction path, a deep convolution network is used to perform feature compression on the input signal to generate a multi-scale feature mapping set, and high-order discharge mode representations are extracted layer by layer, the multi-scale feature mapping set covers microscopic transient characteristics, mesoscopic pattern characteristics and macroscopic trend characteristics; in the reverse feature reconstruction path, the macroscopic trend characteristics are gradually restored in time resolution through interpolation operation; the cross-scale connection is established between the forward feature extraction path and the reverse feature reconstruction path, and the same scale feature mapping is associated and fused;

[0020] A convolution kernel is applied to each output node of the space-time coding network architecture to reduce the feature dimension, the multi-scale feature mapping set output by the space-time coding network architecture is superimposed and reconstructed to generate a full-dimensional discharge intensity distribution spectrum.

[0021] Preferably, the method for identifying a high-risk area from the full-dimensional discharge intensity distribution spectrum includes:

[0022] The discharge risk score of each monitoring unit in the full-dimensional discharge intensity distribution spectrum is calculated, the ultra-high frequency electromagnetic radiation waveform, the transient voltage pulse sequence and the radio frequency current spectrum characteristics contained in the multi-source partial discharge signal are nonlinearly weighted and aggregated to obtain the discharge risk score;

[0023] The preset discharge risk score primary threshold and the discharge risk score advanced threshold are compared with the discharge risk score, if the discharge risk score is lower than the discharge risk score primary threshold, the unit corresponding to the score is marked as a safe domain; if the discharge risk score is higher than the discharge risk score primary threshold and lower than the discharge risk score advanced threshold, the unit corresponding to the score is marked as a warning domain; if the discharge risk score is higher than the discharge risk score advanced threshold, the unit corresponding to the score is marked as a high-risk domain.

[0024] The high-risk area of the full-dimensional discharge intensity distribution spectrum is identified by using different color coding, different colors correspond to different risk levels, the safe domain is defined as a risk-free area, the warning domain is defined as a potential risk area, and the high-risk domain is defined as a high-probability fault area.

[0025] Preferably, the method for generating a three-dimensional positioning map of an insulation defect of a switch cabinet comprises:

[0026] According to the identified high-risk area, a gradient density electromagnetic sensor is deployed at a selected position of the switch cabinet shell; the electromagnetic sensor is a distributed electromagnetic sensor array, and a differentiated detection strategy is implemented for different risk areas; a directional detection pulse is emitted by the electromagnetic sensor to the inside of the switch cabinet to perform an initial full-domain scanning, and the detection beam angle is dynamically adjusted according to the high-risk area of the full-dimensional discharge intensity distribution spectrum, wherein the detection beam angle interval of the high-risk area is smaller than that of the warning domain, and the detection beam angle interval of the warning domain is smaller than that of the safe domain; a beam forming algorithm is applied to real-time regulate the detection beam angle, a wave speed correction model of multi-physical field coupling is introduced, and the propagation parameters in the wave speed correction model are continuously optimized by an iterative phase analysis algorithm, and the iteration is terminated when the convergence condition is met; an electromagnetic wave propagation path matrix is constructed based on a ray path tracking algorithm, and the defect spatial distribution is calculated by a regularization inverse problem solving algorithm, and finally a three-dimensional positioning map of the insulation defect of the switch cabinet with a risk level identifier is output.

[0027] Preferably, the method for establishing the time-space correlation between the multi-source partial discharge signal and the three-dimensional positioning map of the insulation defect comprises:

[0028] A global space coordinate system is established with the center point of the bottom surface of the switch cabinet as the coordinate origin, the X-axis and the Y-axis are parallel to the bottom surface of the switch cabinet, and the Z-axis is perpendicular to the bottom surface of the switch cabinet; a reference signal source is used to calibrate the ultra-high frequency sensor to obtain the position parameters and direction parameters of the ultra-high frequency sensor; the sensor coordinates of the full-dimensional discharge intensity distribution spectrum are converted to the device coordinate system through the position parameters of the ultra-high frequency sensor, and then the device coordinate system is converted to the global coordinate system through the direction parameters, and the discharge intensity distribution data in the global coordinate system are obtained; the origin of the global coordinate system is used as the reference point to calibrate the position of the distributed electromagnetic sensor array to obtain the spatial parameters of the distributed electromagnetic sensor array.

[0029] The spatial parameters of the distributed electromagnetic sensor array are used to convert the three-dimensional positioning atlas of the insulation defect of the switch cabinet into a global coordinate system, and then the defect positioning data in the global coordinate system is obtained; in the global coordinate system, the discharge intensity distribution data and the defect positioning data are spatially registered to establish the spatial correlation between the multi-source partial discharge signal and the three-dimensional positioning atlas of the insulation defect; the clock source connected to the ultra-high frequency sensor is set as a master clock, and the clock source connected to the distributed electromagnetic sensor array is set as a slave clock, and the nanosecond-level time alignment is performed through the precise time synchronization protocol to establish the time correlation between the multi-source partial discharge signal and the three-dimensional positioning atlas of the insulation defect.

[0030] Preferably, the method for integrating the full-dimensional discharge intensity distribution spectrum and the three-dimensional positioning atlas of the insulation defect through the multi-channel fusion network comprises:

[0031] According to the discharge intensity distribution data and the defect positioning data in the global coordinate system, a heterogeneous correlation graph is constructed; a multi-head graph neural network based on a feature cross mechanism is used to build a multi-channel fusion network to perform feature-level fusion on the discharge intensity distribution data and the defect positioning data in the global coordinate system; the multi-channel fusion network comprises a data input layer, a feature conversion layer, a graph attention operation layer, a cross-modal interaction layer and a confidence output layer; the heterogeneous correlation graph is input into the data input layer of the multi-channel fusion network, and the comprehensive risk confidence is generated through the confidence output layer.

[0032] Preferably, the method for constructing the heterogeneous correlation graph comprises:

[0033] Each high-risk area in the discharge intensity distribution data in the global coordinate system is taken as a discharge node, and a multi-dimensional feature vector of each high-risk area is extracted as a discharge node feature; each defect cluster in the defect positioning data in the global coordinate system is taken as a positioning node, and a spatial distribution feature vector of each defect cluster is extracted as a positioning node feature; all discharge nodes and positioning nodes are collected to form a node set;

[0034] All discharge nodes are traversed, and the spatial distance between any two discharge nodes in the global coordinate system is calculated; a discharge node distance threshold is preset, if the spatial distance between any two discharge nodes is less than the discharge node distance threshold, a bidirectional connection edge is established between the corresponding discharge nodes; if the spatial distance between any two discharge nodes is greater than or equal to the discharge node distance threshold, no connection is established;

[0035] Traverse all positioning nodes, calculate the spatial distance of any two positioning nodes in the global coordinate system; preset the positioning node distance threshold, if the spatial distance of any two positioning nodes is less than the positioning node distance threshold, then a bidirectional connection edge is established between the corresponding positioning nodes; if the spatial distance of any two positioning nodes is greater than or equal to the positioning node distance threshold, then no connection is established; through spatial nearest neighbor search, the nearest positioning node is matched for each discharge node, and a bidirectional connection edge between the discharge node and the corresponding positioning node is established; all bidirectional connection edges are collected to form an edge set; and a heterogeneous association graph is constructed based on the node set and the edge set.

[0036] Preferably, the method for converting the comprehensive risk confidence into an executable monitoring instruction set by the multi-criteria optimization engine comprises:

[0037] The internal space of the switch cabinet is discretized into micro-element grids, and each micro-element grid records the current discharge intensity value and the comprehensive risk confidence; meanwhile, all configurable monitoring strategy parameters are enumerated, and the monitoring strategy parameters include the adjustment range of each parameter, the calculation resource consumption and the contribution degree to defect positioning in history;

[0038] Three optimization objectives are set, and the optimization objectives include a risk suppression objective, a state awareness objective and a resource economy objective; two types of constraint conditions are configured, and the constraint conditions include a physical invariable constraint and a strategy adjustable constraint;

[0039] A multi-criteria optimization engine is called, N groups of monitoring strategy parameter optimization schemes are randomly generated, the achievement degrees of the three optimization objectives are evaluated for each scheme, the achievement degree scores of the three optimization objectives are obtained, the three achievement degree scores are weighted and aggregated, a comprehensive evaluation value of the scheme is obtained, the scheme with the highest comprehensive evaluation value is selected from the N groups of schemes as the optimal monitoring strategy scheme; the selected optimal monitoring strategy scheme is converted into an actual executable monitoring instruction set; the executable monitoring instruction set includes a collection frequency instruction, a sensor configuration instruction and a diagnosis trigger instruction;

[0040] The achievement degree score of the three optimization objectives comprises:

[0041] The maximum discharge intensity drop in the switch cabinet is counted as the achievement degree score of the risk suppression objective; the information entropy of the comprehensive risk confidence of different micro-element grids is calculated as the achievement degree score of the state awareness objective; and the total amount of calculation resources caused by the adjustment of all monitoring strategy parameters is aggregated as the achievement degree score of the resource economy objective.

[0042] Compared with the prior art, the present application has the following beneficial effects:

[0043] Through the multi-level collaborative architecture, the switch cabinet partial discharge phenomenon is realized all-around, high-precision monitoring and diagnosis. The signal dynamic capture layer obtains multi-source partial discharge signals based on a space perception mechanism, generates a full-dimensional discharge intensity distribution spectrum combining a space-time coding network, can break through the limitations of traditional single sensing technology, fully capture the discharge characteristics of different regions at different times, and focus on key monitoring parts through high-risk area identification to reduce invalid data interference.

[0044] The feature reconstruction mapping layer deploys a distributed electromagnetic sensor array according to the high-risk area, uses a directional detection pulse to penetrate the internal structure of the switch cabinet, and generates a three-dimensional positioning map with a phase analysis algorithm to solve the problem of accurately positioning internal insulation defects that traditional methods cannot solve. Through the design of a distributed array, the uniformity and penetration ability of the signal coverage are improved, and the phase analysis algorithm further improves the spatial accuracy of defect positioning, which can clearly present the specific position and morphological characteristics of the defect, providing intuitive basis for fault analysis.

[0045] The heterogeneous data collaborative layer establishes the space-time association of multi-source signals and three-dimensional positioning maps, aligns the time reference using the precise time synchronization protocol, generates an integrated risk confidence through a multi-channel fusion network, and realizes the deep integration of different types and dimensions of monitoring data. Time synchronization ensures the consistency of data in the time dimension, and the multi-channel fusion network can mine the internal relationship between data, avoiding the one-sidedness of single data evaluation. The generation of integrated risk confidence makes the equipment state evaluation more comprehensive and objective, and can more accurately reflect the overall operation risk of the switch cabinet.

[0046] The adaptive diagnostic decision layer converts the integrated risk confidence into an executable monitoring instruction set based on a multi-criteria optimization engine, and transmits it to the edge computing unit in real time through the industrial Internet of Things protocol, realizing closed-loop management from state perception to decision execution. The multi-criteria optimization engine can generate monitoring instructions that meet actual needs by combining factors such as equipment operating conditions and historical data, and the industrial Internet of Things protocol ensures the real-time and reliability of instruction transmission, enabling the edge computing unit to respond quickly and take appropriate measures, enhancing the system's dynamic response capability to faults. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 The working principle diagram of the switch cabinet partial discharge intelligent monitoring system for fault identification described in the present application;

[0048] Figure 2 The flowchart of the closed-loop optimization layer iterating the monitoring strategy;

[0049] Figure 3 The transient feature active area map;

[0050] Figure 4a full-dimensional discharge intensity distribution spectrum;

[0051] Figure 5 a three-dimensional insulation defect positioning map;

[0052] Figure 6 a flowchart for generating a full-dimensional discharge intensity distribution spectrum;

[0053] Figure 7 a flowchart for generating a three-dimensional insulation defect positioning map;

[0054] Figure 8 a flowchart for integrating a multi-channel fusion network;

[0055] Figure 9 a monitoring strategy effectiveness index. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0057] Please refer to Figure 1 The present application provides a switch cabinet partial discharge intelligent monitoring system for fault identification, which comprises:

[0058] The intelligent monitoring system performs fault identification on the partial discharge of the switch cabinet. The signal dynamic capture layer collects multi-source partial discharge signals based on a spatial perception mechanism. This layer uses a space-time coding network to process signals, outputs a full-dimensional discharge intensity distribution spectrum, and identifies high-risk areas. The feature reconstruction mapping layer deploys a distributed electromagnetic sensor array according to the high-risk areas, transmits a directional detection pulse to penetrate the internal structure of the switch cabinet; analyzes the reflected signals through a phase analysis algorithm to generate a three-dimensional insulation defect positioning map. The heterogeneous data coordination layer aligns the multi-source partial discharge signals and the three-dimensional insulation defect positioning map in the time dimension, and uses a precise time synchronization protocol to unify the reference; a multi-channel fusion network integrates the discharge intensity distribution spectrum and the defect positioning map, and outputs a comprehensive risk confidence. The adaptive diagnostic decision layer converts the risk confidence into an executable monitoring instruction set based on a multi-criteria optimization engine, and transmits it to the edge computing unit in real time through an industrial Internet of Things protocol. The whole system realizes real-time monitoring of the switch cabinet fault through the cooperation of each layer.

[0059] Embodiment 1: Please refer to Figure 2Three types of sensing devices are deployed in the key structural areas of the switchgear to complete the signal acquisition network construction: a UHF sensor is installed in the central position of the front wall of the cabinet, with a detection frequency range set at 300 MHz to 3 GHz to capture electromagnetic radiation waveforms; a transient ground voltage sensor is arranged at the grounding bus connection point of the cabinet, with a range covering 0.1 mV to 10V to collect the ground voltage pulse sequence; and a radio frequency current transformer is sleeved on the cable shielding layer grounding wire, with a fixed sampling rate of 100 MS / s to obtain the current spectral characteristics. The three types of devices cooperatively output multi-source partial discharge signals containing time-domain waveforms, frequency-domain components, and energy distribution.

[0060] The sensor network simulates the non-uniform response mechanism of biological perception and dynamically configures working parameters. Based on the three-dimensional structural model of the switchgear, the key nodes of the physical structure are identified: a 5 cm x 5 cm rectangular monitoring area is defined for the cable joint area, a 10 cm wide strip monitoring area is divided along the axial direction for the insulator surface, and a 3 cm diameter circular monitoring area is established around the bolt for the busbar connection. These areas are defined as high-sensitivity focus areas with a sensitivity coefficient set to the reference value of 1.0. The remaining structural areas of the switchgear are marked as low-sensitivity background areas with a fixed sensitivity coefficient of 0.2. A dynamic sensitivity function is constructed through a spatial coordinate conversion algorithm to real-time divide the high and low sensitivity monitoring spatial ranges.

[0061] The abnormal response network uses a third-order convolution module to process the original signal. The first layer of the network is set with a sliding time window of 10 μs, and a transient feature scan is performed every 20 ms to extract specific waveforms with a pulse rising slope greater than 50 V / μs and a duration shorter than 5 μs as feature points. When the number of feature points in three consecutive sampling windows exceeds the preset threshold of 15, the spatial coordinates are output as transient feature active areas. The system dynamically updates the focus area distribution based on the active area: when a new active area is added, a 3 cm x 3 cm square high-sensitivity focus area is expanded around the point, and its sensitivity coefficient is immediately increased to 1.0; if an existing focus area has no transient feature activity for 50 consecutive scans, it is switched to a low-sensitivity background area with a sensitivity coefficient reduced to 0.2.

[0062] The signal acquisition mode is based on dynamic adjustment of the partition. The high-sensitivity focus area is configured with millisecond-level continuous sampling, with a data acquisition period set at 2 ms and the sensor power consumption mode improved to the high-speed state; the low-sensitivity background area uses second-level interval sampling with an acquisition period of 1.5 s and the sensor switches to a low-power mode. The configuration command is transmitted through the RS485 bus, and a new sensor parameter configuration table is generated after each partition update, which is issued to each sensor node in binary instruction format. The sensor node completes the mode switching within 200 ms after receiving the new configuration, achieving a dynamic balance between acquisition power consumption and data accuracy.

[0063] The state evolution data is collected in real time by the DAC module of the edge computing unit. The system acquires the switch cabinet core state parameter sequence every 10 minutes: the temperature parameter reads the average value of 6 distribution points, the vibration intensity records the RMS value, and the medium loss angle collects three-phase data. The preset state evolution threshold is stored in the FLASH memory, including the temperature upper limit value 75°C, the vibration intensity threshold 4 mm / s, and the medium loss angle limit value 0.02. The state deviation calculation uses the weighted relative error algorithm, with temperature weight 0.4, vibration weight 0.3, and medium loss angle weight 0.3. When the cumulative error exceeds 8%, the system triggers a warning.

[0064] The monitoring strategy performance index is composed of three dimensions: state deviation conversion value accounts for 40% of the index weight, historical strategy execution success rate (effective execution ratio of the previous 24 instructions) accounts for 35%, and hardware resource occupancy rate (memory usage rate and CPU load average value) accounts for 25%. The performance index is updated every 30 minutes, and the numerical value range is standardized to 0 to 100. When the index fluctuation amplitude of 5 consecutive cycles is less than ±2, it is determined to be convergent, triggering the strategy freezing mechanism. The dynamic iteration process uses the gradient descent algorithm, with ±5% adjustment of the collection frequency, ±2 adjustment of the number of sensor starts and stops, and ±10% adjustment of the scanning area, until the performance index enters the preset stable interval of 85 to 95.

[0065] The closed-loop optimization module establishes a double-buffer storage architecture. The current execution strategy is stored in Buffer A, and the newly generated strategy in Buffer B is temporarily stored. When the performance index continuously drops below 80 for 1 hour, the system switches to the new strategy in Buffer B, and the historical data in Buffer A is emptied. After the edge computing unit executes each instruction set, it feeds back the switch cabinet voltage fluctuation curve, partial discharge repetition rate curve, and equipment temperature rise curve through the CAN bus. These data are simultaneously written into the SD card history record, forming the data basis for closed-loop iteration. The optimization instruction issuance uses the Modbus-TCP protocol, and the instruction is packaged as a 16-byte data packet, including sensor address code, operation code, and parameter configuration value. The single transmission time is controlled within 150 ms, ensuring the timeliness of strategy execution.

[0066] Referring to Figure 3 , the transient feature active area detected by the abnormal response network is shown. The figure includes three types of sensor data: ultra-high frequency sensor signal (top), transient ground voltage signal (middle), and radio frequency current signal (bottom). The red highlighted area represents the detected transient feature active area, which has a pulse rising edge slope greater than 50 V / μs and a duration shorter than 5 μs. When the number of feature points in three consecutive sampling windows exceeds the threshold, the system marks the area as a transient feature active area and dynamically adjusts the sensitivity partition. As can be seen in the figure, multiple active areas are successfully identified.

[0067] Referring to Figure 4, which shows the full-dimensional discharge intensity distribution spectrum generated by the spatio-temporal encoding network. The color from blue (low risk) to red (high risk) represents the discharge risk score. Three main risk areas are clearly visible in the figure: a high-risk area (red) at the center coordinate (6.0, 5.0), a medium-risk area (yellow) at the coordinate (2.5, 3.0), and a low-risk area (green) at the coordinate (8.5, 2.0). This spectrum integrates the ultra-high frequency electromagnetic radiation waveform, transient ground voltage pulse sequence, and radio frequency current spectral characteristics, achieving the visualization of the spatial distribution of discharge intensity inside the switchgear.

[0068] Referring to Figure 5 , this figure shows the three-dimensional positioning map of insulation defects generated by the feature reconstruction mapping layer. Each point represents a detected insulation defect, and the color represents the risk level (blue for low risk and red for high risk). It can be seen in the figure that there is a high-risk defect cluster near the area (3, 4, 2). This spectrum generates directional detection pulses through a distributed electromagnetic sensor array, applies a phase analysis algorithm and wave speed correction model, and achieves accurate spatial positioning of insulation defects inside the switchgear.

[0069] Referring to Figure 9 , this figure shows the strategy optimization process of the adaptive diagnostic decision layer. The blue curve represents the change of the monitoring strategy performance index over time, the yellow area represents the target performance interval (85-95), and the green area represents the strategy convergence area. The system uses the gradient descent algorithm to dynamically adjust the acquisition frequency, the number of sensor start-stop, and the scanning area, and completes the strategy optimization within the first 15 hours, with the performance index entering the stable interval (85-95). When the performance index fluctuates less than ±2 for 5 consecutive cycles, the system triggers the strategy freezing mechanism.

[0070] Example 2: Referring to Figure 6 , the multi-source partial discharge signal preprocessing uses a sliding window kurtosis normalization algorithm. A sliding time window with a length of 100 ms is set, and it moves along the time axis with a step size of 50 ms. The amplitude peak value of the ultra-high frequency electromagnetic radiation waveform of each sensor channel in the window is calculated, the pulse density of the transient ground voltage pulse sequence is counted, and the main frequency energy proportion of the radio frequency current spectrum is extracted; a dynamic range compression function is used to perform the following calculation on each signal dimension:

[0071]

[0072] wherein, is the original sampling value, is the arithmetic mean of all sampling points in the current sliding window, is the standard deviation in the window, is the kurtosis statistic. The kurtosis factor compensation represents a normalization of the output values, linearizing the signal amplitude for non-Gaussian distributions. The calibration signal stream is generated from the normalized data of each channel independently.

[0073] The normalized signals are reorganized in a five-dimensional tensor space. The time axis is divided into discrete time bins of 10 ms scale; the frequency axis is divided into 20 equal-width frequency bands by the FFT analysis, with a bandwidth of 150 MHz; the amplitude axis is quantized into 256 levels of voltage steps; the phase axis is discretized into 32 equal parts in the interval of 0°-360°; the sensor channel dimension is assigned three independent index channel identifiers. Each five-dimensional data unit stores the integrated energy value under the combination of the time-space-frequency-phase channel. The initial inactive unit is set to zero value, and when the existence of valid signals in the dimension combination is detected, the energy integration result is filled to form a sparse energy distribution tensor.

[0074] The spatio-temporal coding network architecture is designed as a four-order feature pyramid. The forward feature extraction path implements four steps of dimension reduction operation: the first level of convolution kernel size 3×3×3×3×1, 32 channel output, spatial step (2, 2, 2, 2, 1), compressing the time and space dimensions; the second level of convolution kernel size 3×3×3×1×1, 64 channel output, step (2, 2, 2, 1, 1); the third level of convolution kernel size 3×3×1×1×1, 128 channel output, step (2, 2, 1, 1, 1); the fourth level of convolution kernel size 3×1×1×1×1, 256 channel output, step (2, 1, 1, 1, 1). The output of each level is connected to batch normalization and ReLU activation. The backward feature reconstruction path implements three orders of dimension increase: starting from the 256 channel feature map, the first deconvolution layer size 3×1×1×1×1, 128 channel output, step (2, 1, 1, 1, 1); the second deconvolution layer size 3×3×1×1×1, 64 channel output, step (2, 2, 1, 1, 1); the third deconvolution layer size 3×3×3×1×1, 32 channel output, step (2, 2, 2, 1, 1). The cross-scale connection concatenates the third level output in the forward path to the second level input in the backward path in the frequency domain dimension, and the second level output in the amplitude dimension is concatenated to the first level input in the backward path. The final output layer uses a 1×1×1×1×3 convolution kernel to aggregate channel information.

[0075] The network output is processed by feature reconstruction. Micro-transient features are extracted from the first stage output of the forward path, recording the characteristics of the nanosecond pulse front; meso-mode features are extracted from the third stage output of the forward path, representing the millisecond discharge cycle mode; macro-trend features are obtained at the third stage output of the reverse path, describing the trend of the second discharge intensity. In the cross-scale fusion stage, the micro-feature matrix is interpolated and aligned to the macro-feature matrix coordinate system according to the timestamp, and the amplitude channel is consistent through bilinear interpolation, and finally stacked into an 800x600 resolution image matrix. After Gamma correction (coefficient 2.2), the matrix is mapped to a full-dimensional discharge intensity distribution spectrum, with the horizontal axis representing the switchgear X-axis spatial position (pixel coordinates 0-800 corresponding to 0-1 m physical size), and the vertical axis representing the Y-axis position (pixel coordinates 0-600 corresponding to 0-0.75 m), and the color channel H value is linearly mapped to the discharge intensity range of 0-1000 pC.

[0076] The discharge risk scoring unit uses a dimensional dispersion algorithm. For each 2cmx2cm grid cell, calculate the three-dimensional feature dispersion :

[0077] Ultra-high frequency signal dimension: Calculate the energy distribution variance of each 20MHz frequency band in the range of 30-3000MHz with 20MHz as the unit

[0078] Transient ground voltage dimension: Count the number of positive and negative pulses in a 100μs time window

[0079] Radio frequency current dimension: Extract the amplitude ratio of the third harmonic to the fundamental at 1MHz resolution

[0080]

[0081] wherein, is the current grid risk score, with a value range of 0-1. The primary risk threshold is , and the advanced risk threshold is . The spectrum visualization processing uses RGB three-channel rendering: Region fill RGB(80,200,80) green; Region fill RGB(255,200,50) yellow; Region fill RGB(220,50,60) red. Each grid cell is labeled with a numerical risk score value, embedded in the center of the grid in Arial 9 font. The spectrum metadata area records the current scan time, maximum risk value coordinates, and the cumulative high-risk area proportion.

[0082] Example 3: see Figure 7, the distributed electromagnetic sensor array is deployed in a ring configuration outside the switchgear housing. For high-risk areas (areas with a risk score ≥ 0.7), 64 electromagnetic transceiving probes are arranged along the circumference of the outer surface of the cabinet, with a probe spacing ≤ 5 cm; 32 probes are configured in the warning area (0.35 ≤ risk score < 0.7) with a spacing of 10 cm; and 16 probes are configured in the safe area (risk score < 0.35) with a spacing of 20 cm. Each probe includes a microstrip antenna with a frequency range of 200 MHz-1.5 GHz and an ultra-wideband signal generator. The initial scan performs global coverage detection: the array controller activates the probe pairs in the preset sequence, transmits a directional detection pulse with a pulse width of 2 ns, and sets the pulse repetition frequency to 1 kHz in the safe area, 5 kHz in the warning area, and 10 kHz in the high-risk area.

[0083] The beam steering system embeds a dynamic angle compensation algorithm. When the spatial coordinates of a high-risk area are , the adjacent beam angle interval is set to Δθ = 0.8°, and the beam tilt angle Φ is controlled within ±60°; the beam interval in the warning area is Δθ = 2.5°, and the beam interval in the safe area is Δθ = 5°. The beam forming algorithm uses coherent synthesis technology, and the phase delay correction model is:

[0084]

[0085] wherein, represents the propagation delay compensation value of probe i to j, is the physical distance (unit: meters), is the speed of light in vacuum ( ), is the equivalent refractive index of the switchgear composite material (value 1.2-2.1), is the device inherent delay (typical value 1.2 ns). The model parameters are optimized by iteration: initially set , and after collecting the reflected signal, calculate the phase consistency index :

[0086] when the current model is maintained

[0087] when adjusted step ±0.1

[0088] when reset and calibrated

[0089] The iteration termination condition is set to three consecutive , usually 10-15 iterations to achieve convergence.

[0090] Defect positioning is based on the ray path matrix calculation. The propagation path matrix dimension is denotes the number of probes, denotes the number of spatial voxels, the matrix element is:

[0091]

[0092] wherein, is the wavelength of the path, is the path length, is the propagation delay. The reflected signal vector is solved by least square regularization:

[0093]

[0094] wherein, is the defect spatial distribution vector, is the Tikhonov regularization coefficient (default value 0.03). The calculation result is mapped to a 512x512x256 voxel three-dimensional atlas, and each voxel is labeled with a risk level: the original high-risk area is encoded in red, the warning area is yellow, and the safe area is green.

[0095] The spatial reference is established using a standard coordinate system. The geometric center of the bottom surface of the switch cabinet is taken as the origin , the long side direction of the bottom surface is taken as the X axis, the short side direction is taken as the Y axis, and the vertical ground direction is taken as the Z axis to establish a right-handed rectangular coordinate system. The position calibration of the ultra-high frequency sensor is realized by a laser tracker: three reference marker points are set on the sensor shell, and the three-dimensional coordinates are measured with as the reference; the direction parameters are recorded by the compass module to record the yaw angle , the pitch angle , and the roll angle . The coordinate conversion of the full-dimensional discharge intensity distribution spectrum is divided into two steps:

[0096] Device coordinate system conversion: let the local coordinate of the sensor be , and the conversion matrix be ;

[0097] Global mapping: .

[0098] wherein, is the cabinet installation posture correction matrix, which contains the levelness compensation amount.

[0099] The distributed array calibration adopts an acoustic auxiliary method. A piezoelectric transducer is installed on the array installation support to emit a 40 kHz pulse signal; the array probe receives the time difference of arrival of the acoustic wave, and calculates the three-dimensional coordinates of the probe in combination with the temperature and humidity compensation. The array position parameters are recorded as ​​The global coordinate formula of the three-dimensional positioning map of insulation defects is converted into a voxel:

[0100]

[0101] wherein, is the global coordinate of the starting corner point of the map, is the voxel resolution.

[0102] The spatial correlation performs double-precision alignment. The spatial registration adopts the ICP algorithm: 100 high-risk feature point coordinate sets are extracted from the full-dimensional spectrum , and the corresponding position point set is extracted from the defect map , the rotation matrix and the translation vector are solved, and the objective function is minimized:

[0103]

[0104] The iteration ends when the point position mean square error is less than 0.1 mm. The time reference synchronization adopts the PTP precise time protocol: the built-in atomic clock of the ultra-high frequency sensor is the master clock, and generates a 10MHz synchronization signal; the distributed array is equipped with an IEEE1588v2 protocol slave clock module, and transmits the time stamp through the optical fiber. The master clock sends a Sync message every 5 seconds, and the slave clock records the receiving time , and the master clock records the sending time ; the slave clock replies a Delay_Req message, and the master clock records the receiving time , and the slave clock records the sending time . The time delay correction value is:

[0105]

[0106] The clock deviation compensation is within ±2ns, and the cross-system data frame time stamp alignment is realized. The timing data is marked as a 64-bit binary number: the high 32 bits are the second count (from the Unix epoch), and the low 32 bits are the nanosecond count. The discharge intensity data points in the global coordinate system are associated with the defect voxels through the spatial coordinate mapping table, each association pair is marked with the same time stamp index, and a space-time correlation database is formed.

[0107] Example 4: refer to Figure 8 , the following describes the heterogeneous data collaboration process with a 40.5kV switch cabinet monitoring example. The origin of the global coordinate system is located at the center of the bottom surface of the cabinet, and the coordinate system range is: X axis ±0.75m, Y axis ±0.5m, Z axis 0-1.8m. The full-dimensional discharge intensity distribution spectrum data extracts the high-risk area coordinate point set, including 3 high-risk areas; the three-dimensional positioning map of insulation defects extracts the defect cluster center point, and identifies 2 main defect clusters. The data basic information is shown in Table 1:

[0108] Table 1: Monitoring instance data graph for a certain 40.5 kV switchgear.

[0109]

[0110] Heterogeneous association graph construction process: node set initialization includes three discharge nodes F1-F3 and two positioning nodes D1-D2. Traversal of discharge node distance: F1 and F2 spatial distance 0.87 m (exceeding threshold 0.2 m), no connection established; F1 and F3 distance 0.94 m, no connection; F2 and F3 distance 0.76 m, no connection. Traversal of positioning node distance: D1 and D2 are 1.05 m apart, exceeding the threshold, no connection. Spatial proximity matching stage: F1 node coordinates (0.52, 0.32, 0.85) and D1 node (0.55, 0.48, 1.20) calculate the Euclidean distance 0.38 m; and D2 node distance 1.03 m, so bind F1→D1 connection edge. F2 node coordinates (0.15, -0.28, 1.25) and D1 distance 0.97 m, and D2 distance 1.08 m, no matching node. F3 node coordinates (-0.41, 0.18, 0.65) and D1 distance 1.06 m, and D2 distance 0.07 m (below threshold), establish F3→D2 connection edge. Finally generate a heterogeneous association graph topology structure of 5 nodes and 2 connection edges.

[0111] Multi-channel fusion network operation process: the data input layer receives the heterogeneous association graph and performs node feature vectorization encoding: the discharge node feature vector includes risk score normalized value, main frequency band position index, pulse width quantization value, phase distribution entropy and other parameters; the positioning node feature vector includes three-dimensional spatial coordinate offset, volume logarithm value, surface curvature statistical quantity and other parameters. The feature conversion layer uses a three-layer fully connected network to process the input vector: the first layer of 256 neurons outputs, which compresses the 15-dimensional discharge feature to 48-dimensional hidden layer expression; the second layer of 128 neurons outputs, which maps the 8-dimensional positioning feature to 48-dimensional; the third layer of 64 neurons unifies the feature dimension. The graph attention operation layer calculates the connection edge weight: the F3-D2 edge weight factor is 0.92 based on feature similarity calculation, and the F1-D1 edge weight factor is 0.84. The cross-modal interaction layer implements feature fusion: performs feature cross splicing on F3 and D2 nodes to generate a 96-dimensional mixed vector; filters key feature components through a gating mechanism to output the fusion result. The confidence output layer performs binary classification: the F1-D1 association body inputs the multi-layer perceptron, and outputs the comprehensive risk confidence 0.88 through the Sigmoid function; the F3-D2 association body outputs the confidence 0.96. The final output layer generates a comprehensive risk confidence matrix.

[0112] Instance data conversion details:

[0113] Feature vector encoding of discharge node F1:

[0114] Position parameter: X-axis coordinate value 0.52 is mapped to [0.693] (normalized value).

[0115] Spectrum parameter: dominant frequency 128 MHz is marked as one-hot vector [000000010000000] at spectrum index 8th band region.

[0116] Time domain parameter: pulse width 1.2 μs is logarithmically compressed to 0.512.

[0117] Statistical parameter: phase distribution standard deviation 0.35 in 32 equal intervals.

[0118] Feature vector encoding of positioning node D2:

[0119] Spatial parameter: ΔX = -0.03, ΔY = 0.03, ΔZ = -0.05 (relative coordinate offset).

[0120] Morphology parameter: volume 28 mm³ is log10 transformed to 1.447.

[0121] Structural parameter: surface curvature sampling variance 0.28 in 16 directions.

[0122] Network operation parameter configuration: 8 independent head structures are set for graph attention layer, and each head is allocated a 64-dimensional query vector space. The double control mechanism is adopted for the gated cross mechanism: the discharge feature channel is set to a forgetting gate threshold of 0.6, and the positioning feature channel is set to a forgetting gate threshold of 0.4. The multi-layer perception hidden layer configuration is: [96-dimensional input → 128 neurons ReLU → 64 neurons ReLU → 1 neuron Sigmoid]. The weight initialization of each layer adopts He normal distribution, and the initial value of the bias term is 0.01. The Adam optimizer is used in the training process, the learning rate is fixed at 0.001, and the batch size is set to 8 groups of associated body data. Each inference calculation takes about 12 ms, and the output confidence is accurate to three decimal places.

[0123] At the data storage level, the heterogeneous association graph adopts an adjacency list structure for storage: the node information table records the feature vector and type identifier, and the edge information table records the source node ID, target node ID, and weight value. The node feature data is stored in the DDR4 memory pool in single-precision floating-point array format, and the data index is rebuilt each time the graph is updated. The comprehensive risk confidence output matrix is stored as a two-dimensional array structure, with the row number corresponding to the discharge node index, the column number corresponding to the positioning node index, and the confidence value saved in percentage format. This matrix is transmitted to the decision system buffer area through the DMA channel, serving as the input basis for the subsequent optimization engine.

[0124] Example 5: Discretization of the internal space of the switchgear divides the three-dimensional area into a standard cubic microelement grid unit. The unit size is set to 2 cm x 2 cm x 2 cm, with 75 units along the X-axis, 50 units along the Y-axis, and 90 units along the Z-axis, generating a total of 337,500 grids. Each grid stores two key data: the discharge intensity numerical record is a 16-bit integer variable, with a value range of 0-1000 corresponding to an actual discharge of 0-1000 pC; the comprehensive risk confidence is marked as a single-precision floating-point number, with a numerical range of 0.0-1.0. The data is stored in a sparse matrix format, with only non-zero data units retaining physical memory space.

[0125] The monitoring strategy parameter space contains three adjustable dimensions: the acquisition frequency parameter supports gradient adjustment, with 10 adjustment levels set: 1 Hz, 5 Hz, 10 Hz, 50 Hz, 100 Hz, 200 Hz, 500 Hz, 700 Hz, 800 Hz, and 1 kHz; the sensor configuration parameter includes sensitivity selection (low / medium / high three levels), working mode (single-point trigger / continuous scanning / event-driven), and channel enable state (independent switch); the diagnostic trigger parameter includes risk confidence threshold (0.3-0.9 variable) and area range threshold (minimum detection area size can be set to 1-20 units). The historical contribution of each parameter is calculated through the historical database: the activation times of each parameter, the number of defects detected during parameter setting, and the system resource consumption under the parameter combination are weighted to calculate the contribution index, which is converted to a percentage and stored in the parameter attribute table.

[0126] The optimization target system defines three core indicators: the risk suppression target requires calculating the relative reduction of the maximum discharge intensity in all grids, with the benchmark being the average peak value of the previous 30-minute monitoring period. The state awareness target calculates the diversity of grid cell confidence distribution: all grid cells with a confidence greater than 0.2 are traversed, and the proportion of cells in different confidence intervals (0.2-0.4, 0.4-0.6, 0.6-0.8, 0.8-1.0) is calculated, and the information entropy value is calculated based on this distribution. The resource economy target summarizes four indicators: the ratio of the total number of enabled sensors to the maximum available number, the proportion of high sampling rate channels, the memory consumption increment percentage, and the network bandwidth occupancy growth.

[0127] The constraint conditions are divided into two categories. The physical immutable constraints include: the minimum sensor start-stop interval period (≥1 second), the maximum scanning area coverage (≤95% of the cabinet volume), and the upper limit of the number of channels (fixed at 48). The strategy adjustable constraints include: the maximum number of high-sensitivity mode enabled channels is 12, the area proportion of high-frequency sampling regions is ≤25%, and the maximum scanning period of low-risk regions is ≤10 seconds. These constraints are written in the optimization strategy library in the form of logical rules.

[0128] The multi-criteria optimization engine performs a parallel computing mechanism. The first round generates 500 sets of random strategy solutions, each set containing a combination of parameters in three dimensions. The parameter solution generation adopts the Latin hypercube sampling method to ensure uniform coverage of the parameter space. The three-objective evaluation process of each solution: the risk suppression target calls the peak discharge analysis algorithm to calculate the percentage reduction of the maximum discharge intensity under the current solution relative to the historical benchmark; the state-aware target runs the confidence partition statistics module to output the information entropy calculation result; the resource economy target estimates the total CPU cycle consumption, memory usage increment, and network bandwidth demand through the resource simulator.

[0129] The target achievement score conversion rule: the risk suppression value (0-100% interval) is linearly mapped to 0-100 points; the state-aware entropy value is logarithmically converted, with a value range of 1.5-4.0 mapped to 0-100 points; the total resource consumption (0-100 resource points) is inversely mapped to the score (100-0 points). The weight distribution is a fixed proportion: risk suppression 40%, state awareness 35%, and resource economy 25%. The weighted sum of the three scores of each set of solutions gives the comprehensive evaluation value, which is stored in the solution evaluation record table.

[0130] The optimal strategy selection adopts a two-round screening: the first round selects the top 50 solutions in terms of comprehensive evaluation value; the second round tests the stability of the solutions by injecting ±5% parameter perturbations and recalculating the score fluctuation value, and selects the solution with the smallest fluctuation as the final strategy. The parameter combination of this strategy is sent to the instruction compilation module.

[0131] The executable monitoring instruction set includes three types of core instructions: acquisition frequency instructions are encoded in hexadecimal format, with an example instruction 0xA103 indicating that channel 1 is set to an acquisition frequency of 700Hz (the frequency gear corresponding to the code 03); sensor configuration instructions contain a 32-bit word length: the first 8 bits are the device address code (such as 0x1F), the middle 8 bits are the operation code (such as 0x02 to start the high sensitivity mode), and the last 16 bits are the parameter value (such as 0x0003 for continuous scanning mode); the diagnostic trigger instruction structure is divided into three parts: the first part defines the range of the cube region (6 bytes store the starting coordinates and size), the second part sets the risk threshold (1 byte), and the third part specifies the response action (1 byte identifies the sound and light alarm or data recording). All instructions are added with an instruction header (containing a timestamp and a check code) before sending, and are delivered through the industrial Ethernet frame format. The edge computing unit receives and decodes the instructions for execution: first, verify the integrity of the instructions; then lock the corresponding hardware interface; finally, update the device control register. The minimum interval of instruction update is 1 second, ensuring the real-time response requirements of the system. The instruction execution status code is fed back to the central processing unit through the reverse channel, forming a closed loop of strategy execution.

[0132] It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting; it is not intended to exclude myriad other embodiments of the present application that other inventors can develop based on the same general inventive concepts embodied by the described embodiments. That is, although the present application is described in terms of particular embodiments and illustrative figures, it should be apparent that the scope of the present application is not limited to these specific embodiments.

[0133] While the embodiments of the application have been shown and described herein, it will be understood by those skilled in the art that many changes, modifications, substitutions and alterations to these embodiments can be made without departing from the principles and spirits of the application, and it is intended that the scope of the application be limited solely by the scope of the appended claims and the equivalents thereof.

Claims

1. A switchgear partial discharge intelligent monitoring system for fault identification, characterized in that, Comprise: A signal dynamic capture layer obtains multiple-source partial discharge signals of a switch cabinet based on a spatial perception mechanism, applies a space-time coding network to the multiple-source partial discharge signals, generates a full-dimensional discharge intensity distribution spectrum, and identifies a high-risk area for the full-dimensional discharge intensity distribution spectrum; A feature reconstruction mapping layer deploys a distributed electromagnetic sensing array according to the identified high-risk area, transmits a directional probe pulse to penetrate the internal structure of the switch cabinet, analyzes internal reflection signals by a phase analysis algorithm, and generates a three-dimensional positioning map of insulation defects of the switch cabinet; A heterogeneous data collaboration layer establishes a space-time association between the multiple-source partial discharge signals and the three-dimensional positioning map of insulation defects, aligns a time reference by using a precise time synchronization protocol, integrates the full-dimensional discharge intensity distribution spectrum and the three-dimensional positioning map of insulation defects by a multi-channel fusion network, and generates a comprehensive risk confidence; An adaptive diagnostic decision layer converts the comprehensive risk confidence into an executable monitoring instruction set based on a multi-criteria optimization engine, and real-time issues the executable monitoring instruction set to an edge computing unit through an industrial Internet of Things protocol.

2. The fault identification oriented intelligent monitoring system of partial discharge of switchgear according to claim 1, characterized in that, Further comprise: A closed-loop optimization layer continuously tracks the state evolution of the switch cabinet after the edge computing unit executes the executable monitoring instruction set, calculates the deviation degree of the state evolution data from a preset state evolution threshold, generates a monitoring strategy effectiveness index, and dynamically iterates the switch cabinet partial discharge monitoring strategy until the monitoring strategy effectiveness index converges to a stable interval.

3. The fault identification oriented intelligent monitoring system of partial discharge of switchgear according to claim 1, characterized in that, The method for obtaining multiple-source partial discharge signals of a switch cabinet based on a spatial perception mechanism comprises: Deploying three types of sensing devices, i.e., ultra-high frequency sensors, transient ground voltage sensors, and radio frequency current transformers, at key monitoring points of the switch cabinet, and collecting multiple-source partial discharge signals of the switch cabinet body; the multiple-source partial discharge signals include ultra-high frequency electromagnetic radiation waveforms, transient ground voltage pulse sequences, and radio frequency current spectrum characteristics; Simulating the non-uniform response mechanism of a biological perception system, dynamically configuring the sensitivity and sampling frequency of the sensors; based on a switch cabinet structure model, marking the cable joints, insulator surfaces, and busbar connection points of the switch cabinet as high-sensitivity focus areas; marking the remaining structure areas of the switch cabinet as low-sensitivity background areas; dynamically constructing a sensitivity function to divide the monitoring space of the switch cabinet into high-sensitivity focus areas and low-sensitivity background areas; Real-time analyzing transient characteristics in the multiple-source partial discharge signals through an abnormal response network; the input of the abnormal response network is the multiple-source partial discharge signals, and the output is a transient characteristic active area; dynamically updating the range of the focus area according to the transient characteristic active area; if a new transient characteristic active area is detected, upgrading the new transient characteristic active area to a high-sensitivity focus area; if no transient characteristic activity is detected in the original high-sensitivity focus area, downgrading the original high-sensitivity focus area to a low-sensitivity background area; the high-sensitivity focus area adopts millisecond-level continuous sampling, and the low-sensitivity background area adopts second-level interval sampling; dynamically regulating the working mode of the sensors through configuration instructions to obtain optimized multiple-source partial discharge signals.

4. The fault identification oriented intelligent monitoring system of partial discharge of switchgear according to claim 3, characterized in that, The method for applying a space-time coding network to the multiple-source partial discharge signals to generate a full-dimensional discharge intensity distribution spectrum comprises: Perform kurtosis normalization processing on the multi-source partial discharge signals, integrate the normalized multi-source partial discharge signals into a five-dimensional tensor according to the signal source dimension, wherein the dimensions of the five-dimensional tensor include a time axis, a frequency axis, an amplitude axis, a phase axis and a sensor channel; construct a space-time coding network architecture, which includes a forward feature extraction path, a backward feature reconstruction path and a cross-scale connection; Input the multi-source partial discharge signals into the space-time coding network architecture, in the forward feature extraction path, use a deep convolutional network to compress the features of the input signals, generate a multi-scale feature mapping set, and extract high-order discharge mode representations layer by layer, the multi-scale feature mapping set covers microscopic transient features, mesoscopic pattern features and macroscopic trend features; in the backward feature reconstruction path, the macroscopic trend features are recovered step by step through interpolation operation; a cross-scale connection is established between the forward feature extraction path and the backward feature reconstruction path, and the same scale feature mapping is associated and fused; Apply a convolution kernel to reduce the feature dimension at each output node of the space-time coding network architecture, superimpose and reconstruct the multi-scale feature mapping set output by the space-time coding network architecture, and generate a full-dimensional discharge intensity distribution spectrum.

5. The fault identification oriented intelligent monitoring system of partial discharge of switchgear according to claim 4, characterized in that, The method for identifying high-risk areas in the full-dimensional discharge intensity distribution spectrum includes: Calculate the discharge risk score of each monitoring unit in the full-dimensional discharge intensity distribution spectrum, and obtain the discharge risk score by nonlinearly weighting and aggregating the ultra-high frequency electromagnetic radiation waveform, transient voltage pulse sequence and radio frequency current spectrum features contained in the multi-source partial discharge signals; Pre-set a primary discharge risk score threshold and a senior discharge risk score threshold, compare the discharge risk score with the primary discharge risk score threshold and the senior discharge risk score threshold respectively, if the discharge risk score is lower than the primary discharge risk score threshold, mark the unit corresponding to the score as a safe domain; if the discharge risk score is higher than the primary discharge risk score threshold and lower than the senior discharge risk score threshold, mark the unit corresponding to the score as a warning domain; if the discharge risk score is higher than the senior discharge risk score threshold, mark the unit corresponding to the score as a high-risk domain; Identify the high-risk areas of the full-dimensional discharge intensity distribution spectrum using different color coding, different colors correspond to different risk levels, define the safe domain as a risk-free area, the warning domain as a potential risk area, and the high-risk domain as a high-probability fault area.

6. The fault identification oriented intelligent monitoring system of partial discharge of switchgear according to claim 5, characterized in that, The method for generating a three-dimensional positioning map of switch cabinet insulation defects includes: According to the identified high-risk area, a gradient density of electromagnetic sensors is arranged at a selected position of the switch cabinet shell; the electromagnetic sensor is a distributed electromagnetic sensor array, a differentiated detection strategy is implemented for different risk areas, a directional detection pulse is emitted by the electromagnetic sensor to the inside of the switch cabinet, an initial global scan is performed, and the detection beam angle is dynamically adjusted according to the high-risk area of the full-dimensional discharge intensity distribution spectrum, wherein the detection beam angle interval of the high-risk area is less than that of the early warning area, and the detection beam angle interval of the early warning area is less than that of the safe area; a beam forming algorithm is applied to real-time regulate and control the detection beam angle, a wave speed correction model of multi-physical field coupling is introduced, the propagation parameters in the wave speed correction model are continuously optimized by an iterative phase analysis algorithm, and the iteration is terminated when the convergence condition is met; an electromagnetic wave propagation path matrix is constructed based on a ray path tracking algorithm, a defect space distribution is calculated by a regularization inverse problem solving algorithm, and finally a switch cabinet insulation defect three-dimensional positioning map with a risk level identifier is output.

7. The fault identification oriented intelligent monitoring system of switchgear partial discharge according to claim 6, characterized in that, The method for establishing the space-time correlation between the multi-source partial discharge signal and the insulation defect three-dimensional positioning map comprises: A global space coordinate system is established with the center point of the bottom surface of the switch cabinet as the coordinate origin, the X-axis and the Y-axis being parallel to the bottom surface of the switch cabinet, and the Z-axis being perpendicular to the bottom surface of the switch cabinet; a reference signal source is used to calibrate the ultra-high frequency sensor to obtain the position parameters and direction parameters of the ultra-high frequency sensor; the sensor coordinates of the full-dimensional discharge intensity distribution spectrum are converted to the device coordinate system through the position parameters of the ultra-high frequency sensor, and then the device coordinate system is converted to the global coordinate system through the direction parameters, thereby obtaining the discharge intensity distribution data in the global coordinate system; the origin of the global coordinate system is used as the reference point to calibrate the position of the distributed electromagnetic sensor array to obtain the spatial parameters of the distributed electromagnetic sensor array; The switch cabinet insulation defect three-dimensional positioning map is converted to the global coordinate system through the spatial parameters of the distributed electromagnetic sensor array, thereby obtaining the defect positioning data in the global coordinate system; in the global coordinate system, the discharge intensity distribution data and the defect positioning data are spatially registered to establish the spatial correlation between the multi-source partial discharge signal and the insulation defect three-dimensional positioning map; the clock source connected to the ultra-high frequency sensor is set as the master clock, the clock source connected to the distributed electromagnetic sensor array is set as the slave clock, nanosecond-level time alignment is performed through the precise time synchronization protocol, and the time correlation between the multi-source partial discharge signal and the insulation defect three-dimensional positioning map is established.

8. The fault identification oriented intelligent monitoring system of switchgear partial discharge according to claim 7, characterized in that, The method for integrating the full-dimensional discharge intensity distribution spectrum and the insulation defect three-dimensional positioning map through a multi-channel fusion network comprises: According to the discharge intensity distribution data and the defect positioning data in the global coordinate system, a heterogeneous correlation graph is constructed; a multi-head graph neural network based on a feature cross mechanism is used to build a multi-channel fusion network to perform feature-level fusion on the discharge intensity distribution data and the defect positioning data in the global coordinate system; the multi-channel fusion network comprises a data input layer, a feature conversion layer, a graph attention operation layer, a cross-modal interaction layer and a confidence output layer; the heterogeneous correlation graph is taken as the input of the data input layer of the multi-channel fusion network, and the comprehensive risk confidence is generated through the confidence output layer.

9. The fault identification oriented intelligent monitoring system of switchgear partial discharge according to claim 8, characterized in that, The method for constructing the heterogeneous correlation graph comprises: Each high-risk area in the discharge intensity distribution data in the global coordinate system is taken as a discharge node, and a multi-dimensional feature vector of each high-risk area is extracted as a discharge node feature; each defect cluster in the defect positioning data in the global coordinate system is taken as a positioning node, and a spatial distribution feature vector of each defect cluster is extracted as a positioning node feature; all discharge nodes and positioning nodes are collected to form a node set; All discharge nodes are traversed, and the spatial distance between any two discharge nodes in the global coordinate system is calculated; a discharge node distance threshold is preset, if the spatial distance between any two discharge nodes is less than the discharge node distance threshold, a bidirectional connection edge is established between the corresponding discharge nodes; if the spatial distance between any two discharge nodes is greater than or equal to the discharge node distance threshold, no connection is established; All positioning nodes are traversed, and the spatial distance between any two positioning nodes in the global coordinate system is calculated; a positioning node distance threshold is preset, if the spatial distance between any two positioning nodes is less than the positioning node distance threshold, a bidirectional connection edge is established between the corresponding positioning nodes; if the spatial distance between any two positioning nodes is greater than or equal to the positioning node distance threshold, no connection is established; through spatial nearest neighbor search, the nearest positioning node is matched for each discharge node, and a bidirectional connection edge between the discharge node and the corresponding positioning node is established; all bidirectional connection edges are collected to form an edge set; the heterogeneous correlation graph is constructed based on the node set and the edge set.

10. The fault identification oriented intelligent monitoring system of partial discharge of switchgear according to claim 9, characterized in that, The method for converting the comprehensive risk confidence into an executable monitoring instruction set based on the multi-criteria optimization engine comprises: The internal space of the switch cabinet is discretized into micro-element grids, and each micro-element grid records the current discharge intensity value and the comprehensive risk confidence; meanwhile, all configurable monitoring strategy parameters are enumerated, the monitoring strategy parameters including the adjustment range of each parameter, the calculation resource consumption and the contribution to defect positioning; Three optimization objectives are set, the optimization objectives including a risk suppression objective, a state awareness objective and a resource economy objective; two types of constraint conditions are configured, the constraint conditions including a physical invariable constraint and a strategy adjustable constraint; The multi-criteria optimization engine is called to randomly generate N sets of monitoring strategy parameter optimization schemes, to evaluate the achievement degrees of the three optimization objectives for each set of schemes, to obtain achievement degree scores of the three optimization objectives, to perform weighted aggregation on the three achievement degree scores, to obtain a comprehensive evaluation value of the schemes, and to select a scheme with the highest comprehensive evaluation value from the N sets of schemes as an optimal monitoring strategy scheme; the selected optimal monitoring strategy scheme is converted into an actual executable monitoring instruction set; the executable monitoring instruction set includes a collection frequency instruction, a sensor configuration instruction, and a diagnosis trigger instruction. The achievement degree score of the three optimization objectives is obtained by: The achievement degree score of the risk suppression objective is obtained by counting the maximum discharge intensity drop in the switch cabinet; the achievement degree score of the state perception objective is obtained by calculating the information entropy of the comprehensive risk confidence of different micro-element grids; and the achievement degree score of the resource economy objective is obtained by aggregating the total amount of calculation resources caused by adjustment of all monitoring strategy parameters.

Images