Insulation defect characteristic detection method and system based on switch cabinet

Abstract

The invention discloses an insulation defect feature detection method and system based on a switch cabinet, and relates to the field of insulation defect feature detection, and the method comprises the steps: installing a sensor array in the switch cabinet, collecting an original data stream of a multi-modal space-time physical field, and generating a reconstructed four-dimensional space-time field tensor through a multi-modal space-time cooperative sensing network; based on the four-dimensional space-time field tensor, generating an enhanced sample set conforming to a physical rule through physical regularization; generating an optimized detection parameter combination through gradient-evolution collaboration based on the enhanced sample set conforming to the physical law; digital twin enhanced closed-loop adaptive optimization is adopted, and the optimized detection parameter combination is converted into verified stable parameter configuration; according to the verified stable parameter configuration, real-time optimization parameters are obtained through embedded real-time acceleration optimization and solidification; the detection parameters are effectively adjusted and optimized by adopting a multi-target covariance evolutionary strategy to optimize a parameter combination process based on a space-time gradient characteristic spectrum.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of insulation defect feature detection, in particular to switch cabinet-based insulation defect feature detection and system. BACKGROUND

[0002] In recent years, with the rapid development of smart grid and power equipment state monitoring technology, the detection method for switch cabinet insulation defects gradually evolves from single parameter threshold judgment to multi-modal data fusion analysis. Traditional detection technology mainly relies on single physical quantity detection means such as local discharge signal acquisition, temperature monitoring and gas composition analysis, for example, capturing local discharge pulses through high-frequency current transformers, or monitoring temperature rise anomalies using infrared thermal imagers. With the development of sensor miniaturization and embedded technology, distributed optical fiber sensing, miniaturized gas sensors and microwave interferometry technology have been gradually applied to internal state perception of switch cabinets, realizing multi-dimensional data acquisition of local discharge, SF6 decomposition gas concentration and mechanical deformation. In the data processing layer, machine learning-based fault diagnosis models have attempted to combine multi-source data for feature extraction, and some studies have used spatial interpolation algorithms to reconstruct the spatial field of discrete sensor data to improve defect positioning accuracy. In addition, the application of digital twin technology in equipment state simulation and predictive maintenance provides a new idea for parameter optimization, but existing methods still have significant technical bottlenecks in physical law embedding, dynamic parameter adaptation and real-time performance.

[0003] Although existing technologies have made some progress in multi-modal perception and data analysis, their core defects are reflected in the following three aspects: first, the lack of spatio-temporal coordination of multi-modal data. Traditional sensor networks are mostly deployed in regular grids, which is difficult to achieve full coverage perception of complex cavity structures, resulting in limited spatial resolution, and no strict spatio-temporal alignment mechanism is established, with timestamp deviations of different modal data often exceeding 1ms, which seriously affects the field reconstruction accuracy. Second, the separation of physical laws and data-driven methods. Existing machine learning methods rely on historical data to train models, but the generated abnormal samples often violate physical laws such as conservation of mass and energy conduction, for example, the lack of spatio-temporal correlation between temperature rise and gas diffusion in local discharge simulation, resulting in a significant decrease in model generalization ability under extreme conditions. Third, the contradiction between parameter optimization and real-time performance. Traditional parameter optimization methods do not consider the influence of spatio-temporal gradient features on detection sensitivity, and lack hardware-level acceleration design, resulting in a delay of more than 100ms from data acquisition to parameter output, which is difficult to meet the millisecond-level fault response requirements of smart grids. SUMMARY

[0004] In view of the above existing problems, the present application is proposed.

[0005] Therefore, the application provides a switch cabinet-based insulation defect feature detection to solve the problems of insufficient spatiotemporal coordination perception accuracy of multi-modal data, sample distortion caused by the disconnection of physical laws and data-driven models, poor real-time performance of dynamic parameter optimization, and decreased detection stability under equipment aging.

[0006] To solve the above technical problems, the application provides the following technical solutions.

[0007] In a first aspect, the application provides a switch cabinet-based insulation defect feature detection, comprising:

[0008] A sensor array is installed to collect multi-modal spatiotemporal physical field original data streams, and a multi-modal spatiotemporal coordination perception network is used to generate reconstructed four-dimensional spatiotemporal field tensors;

[0009] Based on the four-dimensional spatiotemporal field tensors, physical regularization is used to generate an enhanced sample set conforming to physical laws;

[0010] Based on the enhanced sample set conforming to physical laws, gradient-evolution coordination is used to generate an optimized detection parameter combination;

[0011] A digital twin enhanced closed-loop adaptive optimization is used to convert the optimized detection parameter combination into a verified stable parameter configuration;

[0012] According to the verified stable parameter configuration, embedded real-time acceleration optimization solidification is used to obtain real-time optimization parameters.

[0013] As a preferred solution of the switch cabinet-based insulation defect feature detection, the multi-modal spatiotemporal coordination perception network comprises an artificial intelligence model that integrates multi-source heterogeneous sensor data, deeply fuses spatiotemporal features, and realizes cross-modal collaborative reasoning.

[0014] As a preferred solution of the switch cabinet-based insulation defect feature detection, the generation of the reconstructed four-dimensional spatiotemporal field tensors by the multi-modal spatiotemporal coordination perception network comprises the following steps: time synchronization and space calibration of the collected original sensor data streams to obtain spatiotemporally synchronized multi-modal data streams; weight optimization of temperature, deformation, and gas in the spatiotemporally synchronized multi-modal data streams to fuse the modal data of temperature, deformation, and gas and obtain three physical field data of temperature, deformation, and gas concentration; and fusion processing of spatial discretization, time serialization, and blind area compensation on the three physical field data to generate the reconstructed four-dimensional spatiotemporal field tensors.

[0015] As a preferred scheme of the switch cabinet-based insulation defect feature detection provided in the application, the generating of the enhanced sample set conforming to the physical law through physical regularization comprises embedding a four-dimensional space-time field tensor through a physical law and setting a physical prior parameter to obtain a physically constrained embedded generative adversarial model; the physically constrained embedded generative adversarial model is subjected to adversarial training through a data distribution matching stage and a physical constraint strengthening stage to obtain an extreme working condition sample set; and the enhanced sample set conforming to the physical law is obtained through multi-scale frequency domain inspection and conservation law joint verification based on the extreme working condition sample set.

[0016] As a preferred scheme of the switch cabinet-based insulation defect feature detection provided in the application, the generating of the optimized detection parameter combination through gradient-evolution coordination comprises obtaining a space-time gradient feature map through four-dimensional gradient calculation and feature enhancement based on the enhanced sample set conforming to the physical law; and the optimized detection parameter combination is generated by dynamically searching for a Pareto optimal solution under physical constraints through a multi-objective covariance evolution strategy based on the space-time gradient feature map.

[0017] As a preferred scheme of the switch cabinet-based insulation defect feature detection provided in the application, the digital twin enhanced closed-loop adaptive optimization comprises obtaining a verification parameter configuration through multi-fault mode injection and virtual-real data fusion verification based on the optimized parameter combination; and the stable parameter configuration is obtained through performance degradation monitoring and dynamic parameter re-optimization based on the verification physical prior parameter configuration.

[0018] As a preferred scheme of the switch cabinet-based insulation defect feature detection provided in the application, the obtaining of the real-time optimization parameter through embedded real-time acceleration optimization solidification comprises obtaining a hardware acceleration module through a four-dimensional convolution pulsating array accelerator and memory computation optimization based on the stable parameter configuration; and the real-time optimization parameter is obtained through real-time task scheduling and a time-sensitive network protocol based on the hardware acceleration module.

[0019] In a second aspect, the application provides a switch cabinet insulation defect feature detection system, comprising:

[0020] The sensor array is installed in the switch cabinet to collect multi-modal space-time physical field original data streams, and a multi-modal space-time collaborative perception network is used to generate a reconstructed four-dimensional space-time field tensor.

[0021] The physical law enhancement module generates an enhanced sample set conforming to the physical law through physical regularization based on the four-dimensional space-time field tensor.

[0022] The gradient-optimized parameter determination module generates an optimized detection parameter combination through gradient-evolution coordination based on the enhanced sample set conforming to the physical law.

[0023] The number of twin closed loop modules adopts digital twin enhanced closed loop adaptive optimization, and converts the optimized detection parameter combination into verified stable parameter configuration.

[0024] The hard plus argument module obtains real-time optimization parameters through embedded real-time acceleration optimization solidification according to the verified stable parameter configuration.

[0025] In a third aspect, the present application provides a computer device, comprising a memory and a processor, and the memory stores a computer program, wherein the computer program is executed by the processor to implement any step of the switch cabinet-based insulation defect feature detection according to the first aspect of the present application.

[0026] In a fourth aspect, the present application provides a computer readable storage medium, which stores a computer program, wherein the computer program is executed by the processor to implement any step of the switch cabinet-based insulation defect feature detection according to the first aspect of the present application.

[0027] The present application has the beneficial effects that: through the process of optimizing parameter combination based on the spatiotemporal gradient feature map using the multi-objective covariance evolution strategy, the effective tuning of the detection parameters is realized. The role and purpose are to overcome the problems that the traditional parameter adjustment method relies on experience and trial and error, and is difficult to balance the multi-objective conflict. By calculating the four-dimensional gradient quantization parameter sensitivity, an accurate parameter-field strength mapping relationship is established; the search direction is adjusted by the adaptive covariance matrix, the Pareto front solution is preferentially explored; the convergence to the high value area is accelerated along the gradient direction variation, and the Pareto optimal balance of sensitivity and false alarm rate in arc detection is realized. BRIEF DESCRIPTION OF DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below, and obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0029] Figure 1 The flowchart of the switch cabinet-based insulation defect feature detection method in the embodiment.

[0030] Figure 2 The temperature gradient graph in the spatiotemporal gradient feature map of the switch cabinet-based insulation defect feature detection method in the embodiment.

[0031] Figure 3 The sound pressure level waveform graph of the switch cabinet-based insulation defect feature detection method in the embodiment.

[0032] Figure 4This is a simulated spectrum diagram of the insulation defect feature detection method based on the switch cabinet in the embodiment.

[0033] Figure 5 Graph showing simulated sensor data for the switchgear-based insulation defect feature detection method in an embodiment. DETAILED DESCRIPTION

[0034] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0035] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0036] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0037] Example 1, with reference to Figures 1-4 , which is the first embodiment of the present invention, provides a switch cabinet-based insulation defect feature detection method, comprising the following steps:

[0038] S1: Install a sensor array to collect the original data stream of the multimodal space-time physical field, and generate a reconstructed four-dimensional space-time field tensor through a multimodal space-time collaborative perception network.

[0039] It should be noted that the multimodal spatiotemporal collaborative perception network includes an artificial intelligence model that integrates multi-source heterogeneous sensor data, deeply integrates spatiotemporal features, and realizes cross-modal collaborative reasoning; the technical architecture is: a high-precision four-dimensional perception system builds a comprehensive monitoring framework for power equipment by integrating three-dimensional space and time dimensions. Data comes from multimodal collaborative collection: Figure 4 As shown, a laser rangefinder or a 3D laser scanner is used to scan the surface of the equipment to generate a point cloud, and the reverse engineering software is used to fit the geometric dimensions and collect the dimensions of the power equipment; Figure 4In the example, Simulated Spectrum (Frequency vs. Intensity) represents simulated spectrum (frequency and intensity); Intensity (au) represents intensity (au); Spectralintensity represents spectral intensity; and Frequency (Hz) represents frequency. The original CAD design file of the power equipment is obtained, and the equipment is reverse-modeled using a handheld 3D scanner. The coordinates of key components such as bolts and insulators are extracted to construct a CAD model of the electronic equipment. A FLIRT865 infrared thermal imager is used to receive infrared radiation in the 8-14 μm band, and a temperature field is generated using an uncooled microbolometer array. The temperature matrix is ​​output through the built-in software of the thermal imager, and the timestamp is synchronized with the GPS. A Velodyne VLP-16 3D lidar is used to trigger the piezoelectric ceramic to generate a charge signal through acoustic vibration, and a preamplifier is used to filter and reduce noise to obtain a sound pressure level waveform in dB, as shown in the following example. Figure 3 As shown, the HIOKICT6860 current and voltage transformer uses a Rogowski coil and a capacitor voltage divider to sense current changes and convert high voltage into low voltage signals. The core function of this system is to generate a dynamic digital twin model of power equipment in real time. Figure 3 Here, Simulated Sound Pressure Level Waveform in dB represents the simulated sound pressure waveform (unit: dB); Sound Pressure Level (dB) represents the sound pressure level (dB).

[0040] Furthermore, a reconstructed four-dimensional space-time field tensor is generated through a multimodal space-time collaborative perception network. The steps are as follows: the collected original sensor data stream is time synchronized and spatially calibrated to obtain a space-time synchronized multimodal data stream; the temperature, deformation and gas in the space-time synchronized multimodal data stream are weighted optimized, and the modal data of temperature, deformation and gas are fused to obtain three physical field data of temperature, deformation and gas concentration; the three physical field data are fused through spatial discretization, time serialization and blind spot compensation to generate a reconstructed four-dimensional space-time field tensor.

[0041] The local sending time of the Sync message sent by the main controller, the local arrival time of the Sync message received by the sensor, the local sending time of the Delay_Req (delay request) message sent by the sensor, and the local arrival time of the Delay_Req message received by the main controller are extracted from the original sensor data packet. The clock deviation is calculated using the multi-sensor temperature drift resistance weighted clock synchronization deviation formula, which is expressed as:

[0042]

[0043] Where t1 is the local sending time of the Sync message sent by the master controller, t2 is the local arrival time of the Sync message received by the sensor, t3 is the local sending time of the Delay_Req message sent by the sensor, and t4 is the local arrival time of the Delay_Req message received by the master controller; is the historical drift variance of the kth sensor clock; is the crystal oscillator temperature drift compensation obtained by fitting the temperature-frequency curve; Δt is the clock offset, N is the total number of sensors involved in clock synchronization, and k is the index in a multi-sensor system or multiple measurement scenarios.

[0044] According to the calculated clock deviation, the timestamps of all sensor data packets are corrected to obtain the time-synchronized data stream.

[0045] The time-synchronized data stream is passed through the lidar as the main sensor, with a sampling interval of 10ms as the reference window. For the non-uniformly sampled gas sensor data, the four data points closest to the target time point are selected within the window. A continuous time function is constructed based on the cubic Lagrange interpolation method, and the equivalent value of the target time is calculated. At each sampling moment of the lidar, the data of other sensors are forced to be mapped to that moment through interpolation, resulting in a multi-sensor data matrix with uniform distribution on the time axis, where each row includes the spatial coordinates, alignment timestamp and sensor reading.

[0046] Feature points on the calibration plate are selected in the multi-sensor data matrix, and the three-dimensional coordinates of these points in the lidar coordinate system and the camera coordinate system are recorded respectively to establish the correspondence between the two sets of coordinates; a nonlinear optimization algorithm is used to calculate the optimal rotation matrix and translation vector to minimize the coordinate mapping error between the lidar and the camera; and the gas sensor data in the camera coordinate system is uniformly converted to the lidar coordinate system through the calculated rotation matrix and translation vector to obtain a multimodal data matrix with consistent spatial coordinates.

[0047] The multimodal data matrix with consistent spatial coordinates is used to establish an adaptive octree grid according to the spatial range, and the maximum resolution is set to 5 cm, supporting dynamic refinement or coarsening; the data point of each sensor is mapped to the corresponding network voxel, and the same voxel is fused according to the rules of weight calculation and weighted averaging. At the same time, the timestamp and multimodal data of each voxel are recorded in the grid to form a spatiotemporal integrated data cube, resulting in a gridded four-dimensional data cube.

[0048] The data distribution of three modes of temperature, deformation and gas concentration in the gridded four-dimensional data cube is analyzed, the information entropy of each type of data is calculated, and the initial weight is dynamically assigned based on the information entropy results of each type of data.

[0049] Furthermore, through the Kalman filter algorithm, combined with the difference between the real-time sensor data and the predicted value, the weight of each modality is dynamically corrected, and the weight of the high-noise sensor will be automatically suppressed, while the weight of the high-precision modality will be increased. The sensor noise variance will be introduced as an adjustment parameter to ensure the stability of the weight update process and avoid drastic fluctuations in weight due to single data anomalies.

[0050] A coupling equation between temperature, deformation, and gas concentration is established, and an iterative algorithm is used to discretize the coupling equation into a finite difference form to obtain the fused physical field tensor.

[0051] The fused physical field tensor is used to calculate the deformation gradient at each position, and the area where the deformation rate exceeds the threshold is determined as a "dramatic change zone". New grid nodes are inserted in the drastic change zone to generate an adaptive grid to balance computational efficiency and accuracy.

[0052] The spatial average data spacing and the temporal average sampling interval are calculated based on the adaptive grid and coupled physical field, and the spatiotemporal covariance function is defined to quantify the spatiotemporal correlation strength of adjacent data. For each blind spot position, the interpolation result is generated by weighted fusion based on the spatiotemporal correlation weights of the adjacent known data to obtain a complete four-dimensional tensor.

[0053] Based on the global data of the complete four-dimensional tensor at each moment, the maximum and minimum values ​​are calculated, and the data is linearly mapped to the interval [0,1] to obtain a normalized tensor, which is expressed as:

[0054]

[0055] in, It is expressed as the value of the normalized four-dimensional space-time field tensor at the spatial position (x, y, z) and time t; T (x,y,z,t) Represents the value of the original four-dimensional space-time field tensor at the spatial position (x, y, z) and time t; min(T :,:,:,t ) represents the minimum value at all spatial positions at time t; max(T :,:,:,t ) represents the maximum value at all spatial locations at time t.

[0056] S2: Based on the four-dimensional space-time field tensor, through physical regularization, an enhanced sample set that conforms to physical laws is generated.

[0057] It should be noted that generating an enhanced sample set that conforms to physical laws through physical regularization involves embedding the four-dimensional space-time field tensor through physical laws and setting physical prior parameters to obtain a generative adversarial model with physical constraint embedding. Based on the generative adversarial model with physical constraint embedding, adversarial training is performed through the data distribution matching stage and the physical constraint reinforcement stage to obtain an extreme working condition sample set containing arc faults and mechanical deformations. Based on the extreme working condition sample set, an enhanced sample set that conforms to physical laws is obtained through multi-scale frequency domain testing and conservation law joint verification.

[0058] The orthogonal basis functions of the original four-dimensional space-time field tensor are pre-calculated according to the mathematical characteristics of the physical equations. The time dimension basis functions are generated through harmonic analysis to construct the physical orthogonal basis of the original space-time field data. The original space-time long data is expanded according to the orthogonal basis to extract the core tensor. The core tensor retains the dominant physical mode, filters out noise and non-physical high-frequency components, fixes the orthogonal basis functions, optimizes the core tensor parameters, and obtains a low-dimensional core tensor.

[0059] Furthermore, based on the low-dimensional core tensor, the generator output receives random noise and the core tensor, outputs the initial space-time field data, and constructs an orthogonal basis based on the physical equations. The Tucker tensor decomposition algorithm is used to remove noise interference and capture the physical dominant modes, forming a low-rank core tensor feature library that contains the core laws of the field distribution. The core features are then dynamically controlled through a differentiable physical residual layer. A wave equation residual correction term is superimposed on the generator output, and the physical constraint weight coefficients are adaptively adjusted based on the field evolution trend, achieving a deep coupling between the data generation process and the governing laws of the partial differential equations.

[0060] In the adversarial training framework, the statistical authenticity of generated samples is first improved through a distribution matching mechanism. The discriminator generates a tensor based on Wasserstein distance quantization (Wasserstein distance quantization) and the probability distribution difference between the real field data. A generator-discriminator network alternating training strategy is used to optimize feature space alignment and obtain preliminary generated samples that conform to macroscopic statistical properties. A physical enhancement mechanism is then introduced, using conservation law projection operators to impose manifold constraints on the generated samples. Deviations from conserved quantities such as mass and momentum are embedded as regularization terms in the residual backpropagation process, forcing the generator to automatically compensate for conservation law errors during parameter updates.

[0061] Further, by splicing the random fluctuation characteristics of the random noise simulated breakdown resistance in the space-time field, while embedding the plasma diffusion-recombination dynamics constraints described by the arc current differential equation, the cross-scale coupling of the arc temperature field and the electromagnetic field is generated. For the mechanical deformation reinforcement task, the J2 plasticity constitutive model is used to establish the nonlinear mapping relationship between the strain tensor and the stress tensor, and the plastic strain accumulation is used as the feedback signal to dynamically adjust the material parameter space of the generator, effectively capturing the nonlinear deformation characteristics such as material yield and creep; In the frequency domain characteristic test, wavelet multi-resolution analysis is used to decompose the mechanical deformation field into frequency bands, and by calculating the KL divergence value of the energy distribution of each frequency band and the real data, the abnormal frequency band is located and the abnormal region in the space-time domain is marked. For the marked region, further detect the energy density gradient, momentum flux divergence and other conservation quantity residuals, and based on the iterative correction algorithm of local field quantity, the abnormal points are compensated for physical compatibility, and finally the high-fidelity field data that meet the time-frequency domain statistical characteristics, global conservation law and local dynamics constraints are generated.

[0062] S3: Based on the enhanced sample set conforming to the physical law, the optimized detection parameter combination is generated through gradient-evolution cooperation.

[0063] It should be noted that the generation of the optimized detection parameter combination through gradient-evolution cooperation includes obtaining a space-time gradient feature map through four-dimensional gradient calculation and feature enhancement based on the enhanced sample set conforming to the physical law; as shown in Figure 2 Based on the space-time gradient feature map, the optimized detection parameter combination is generated through a multi-objective covariance evolution strategy under physical constraints; Figure 2 Time Gradient: Temperature (First Time Step) represents the time gradient: temperature (first time step); Time represents time; Time Gradient: Temperature (Last Time Step) represents the time gradient: temperature (last time step); Gradient represents gradient.

[0064] The three-dimensional space in the enhanced sample set conforming to the physical law is divided into a uniform grid, the momentum equation and the pressure Poisson equation are discretized using the finite volume method, the domain boundary is set to no-slip condition, and the velocity field and pressure field are solved step by step by projection method, and the basic space-time velocity field dataset is obtained based on time, height, width, depth and channel; The velocity field tensor is substituted into the encoder, the 5-layer 3D convolution network compresses the velocity field to 32-dimensional hidden variables, and spherical linear interpolation is performed in the hidden variable space to generate a new velocity field that satisfies the conservation of mass. The 32-dimensional hidden variables are substituted into the new velocity field that satisfies the conservation of mass to obtain the enhanced sample set that conforms to the physical law.

[0065] Furthermore, the derivatives of the velocity field in the x, y, and z directions are calculated using a fifth-order central difference based on a physically compliant enhanced sample set. The eigenvalue correction method is used at the mesh boundary to suppress numerical oscillations, and the time derivative is solved using the Adams-Bashforth method to obtain a four-dimensional gradient tensor.

[0066] A physical mask is constructed based on the pressure gradient and vorticity distribution, and the attention weight is learned to dynamically fuse the gradients in all directions to obtain the spatiotemporal gradient feature map; based on the sensitivity target, false alarm suppression target and resource constraint target in the spatiotemporal gradient feature map, the sensitivity target is used to extract the peak intensity of the characteristic response through a sliding time window, the false alarm suppression target imposes L1 regularization constraint in the low vorticity region, and the resource constraint target normalizes the parameter range. At the same time, three mutually constrained objective functions are constructed, and then the improved CMA-ES algorithm (covariance matrix adaptive evolutionary strategy algorithm) is used to update the covariance matrix by fusing the evolutionary path and the elite solution distribution, and the Das-Dennis reference point method is combined to maintain the diversity of the Pareto frontier to obtain the optimized parameter combination.

[0067] Furthermore, based on the optimized parameter combinations, comprehensive performance indicators are calculated online, monitoring changes in the solution set entropy. When the entropy value drops below 0.1, a population restart mechanism is triggered, resetting the covariance matrix and retaining the historically optimal individual. The learning rate is dynamically adjusted based on the rate of change of the quality evaluation indicator of the Pareto solution set to achieve a stable parameter configuration.

[0068] S4: Digital twin-enhanced closed-loop adaptive optimization is used to convert the optimized detection parameter combination into a verified stable parameter configuration.

[0069] It should be noted that the use of digital twin enhanced closed-loop adaptive optimization includes obtaining verified parameter configuration based on optimized parameter combinations through multi-fault mode injection and virtual-real data fusion verification; and obtaining stable parameter configuration based on verified physical prior parameter configuration through performance degradation monitoring and dynamic parameter reoptimization.

[0070] The optimization parameter combination is based on the crack propagation model and the wear model, and the crack growth and surface wear processes are dynamically simulated in the twin model. Through the thermal-mechanical-fluid multi-physics field coupling analysis, the interaction between the temperature field and the stress field is calculated, and oscillating pressure boundary conditions are applied in the fluid domain to simulate the real working condition disturbance, obtaining a multi-fault mode data set including displacement field, stress field and temperature field.

[0071] Further, based on the multi-failure mode data set, the optimal transmission algorithm is used to align the space-time distribution of simulation and measured data, and the virtual vibration frequency domain features and the actual monitoring signals are dynamically weighted and fused through information entropy; the peak signal-to-noise ratio consistency index is defined to verify the effectiveness of the parameter combination in the hybrid feature space, and when it does not meet the standard, it is automatically rolled back to the historical optimal configuration, and the parameter configuration verified by physical consistency and the virtual-real feature mapping relationship are obtained; based on the verified parameters and real-time running data flow, the manifold dimension reduction technology is used to construct the Riemann space benchmark model of the equipment health state, and the time-varying manifold offset is calculated through the sliding window; the adaptive threshold updating strategy is used to evaluate the degradation degree in real time, trigger multi-level alarm, and mark the key period that needs to be re-optimized, and obtain the degradation index time sequence curve and graded alarm signal; based on the verified parameters, alarm signals and real-time features, a Gaussian process proxy model of a composite kernel function is constructed, and an improved entropy search strategy is used to optimize the local parameter space; through safety constraint verification, it is ensured that the new parameters meet the mechanical and thermal constraints, and based on the industrial communication protocol, the parameter update to the physical device controller is realized in seconds, and the online optimized stable parameter configuration and the synchronously updated digital twin model are obtained.

[0072] S5: According to the verified stable parameter configuration, the real-time optimization parameter is obtained through embedded real-time acceleration optimization solidification.

[0073] It should be noted that the real-time optimization parameter obtained through embedded real-time acceleration optimization solidification includes obtaining a hardware acceleration module based on stable parameter configuration through four-dimensional convolution pulsatile array accelerator and memory calculation optimization; obtaining real-time optimization parameter based on hardware acceleration module through real-time task scheduling and time-sensitive network protocol.

[0074] The stable four-dimensional convolution kernel parameter configuration is compressed to 8-bit fixed point through non-uniform segmented quantization of floating point weight, combined with four-dimensional space-time correlation analysis for structured pruning, which significantly reduces the calculation redundancy; design a customized pulsatile array architecture to block process input data and optimize data flow path, realize parallel computing in space-time dimension; adopt 3D stacked memory and Z-type address mapping strategy to improve data locality and reduce access delay, and obtain high-performance hardware acceleration module and supporting compilation tool chain.

[0075] The real-time sensor data stream is substituted into the hardware acceleration module, the priority is dynamically divided based on the task critical path and resource utilization, the time-sensitive network (TSN) protocol is fused to realize microsecond-level time synchronization, and the control instruction, data input, calculation inference and result return are divided into exclusive time windows through the deterministic scheduling table; design an elastic resource management mechanism to dynamically adjust the computing resources according to the queue depth and hardware temperature, support FPGA (field programmable gate array) reconfiguration to quickly respond to high-priority tasks, and obtain real-time optimization parameters.

[0076] Embodiment 2 provides an insulation defect feature detection system of a switch cabinet as an embodiment of the application, comprising:

[0077] a sensor array is installed inside the switch cabinet to collect multi-modal spatio-temporal physical field original data streams, and a multi-modal spatio-temporal collaborative perception network is used to generate a reconstructed four-dimensional spatio-temporal field tensor;

[0078] a physical law enhancement module generates an enhanced sample set conforming to physical laws based on the four-dimensional spatio-temporal field tensor through physical regularization;

[0079] a gradient-optimization parameter calibration module generates an optimized detection parameter combination through gradient-evolution collaboration based on the enhanced sample set conforming to physical laws;

[0080] a digital twin closed-loop module converts the optimized detection parameter combination into a verified stable parameter configuration through digital twin enhanced closed-loop adaptive optimization;

[0081] a hard parameter addition module obtains real-time optimization parameters through embedded real-time acceleration optimization solidification according to the verified stable parameter configuration.

[0082] The embodiment also provides a computer device suitable for insulation defect feature detection of a switch cabinet, comprising a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute the computer executable instructions to realize the insulation defect feature detection of the switch cabinet as proposed in the above embodiment.

[0083] The computer device can be a terminal, and the computer device comprises a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The communication interface of the computer device is used to communicate with external terminals in a wired or wireless manner. The wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device. In addition, the input device can be an external keyboard, touchpad or mouse, etc.

[0084] The embodiment also provides a storage medium on which a computer program is stored, the program being executed by a processor to implement the switch cabinet-based insulation defect feature detection provided in the above embodiment; and the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk, or an optical disk.

[0085] Embodiment 3, refer to Figure 5 As an embodiment of the present application, the embodiment provides a switch cabinet-based insulation defect feature detection and system, in order to verify the beneficial effects of the present application, the scientific demonstration is carried out through economic benefit calculation and simulation time.

[0086] In order to verify the effectiveness of the present application, a two-month simulation experiment was carried out in an industrial park. The experimental object was a high-voltage switch cabinet of the same type, which integrated four types of heterogeneous sensors, including temperature, local, optical and electromagnetic sensors. A total of 48 sensors were installed, the collection frequency was 50 Hz, and the data was input into the model for fault detection and evaluation after preprocessing by the edge computing unit. The sensor data is as shown in Table 1. Figure 5 Table 1: Sensor data

[0087] In the artificial experience parameter tuning stage, the artificial parameter tuning method commonly used in the traditional operation and maintenance scene is simulated. An experienced engineering and technical personnel sets initial detection parameters according to past cases in the field environment, and then repeatedly tests and optimizes by trial and error. The whole process lasts about 215 minutes, and a total of 20 rounds of parameter combinations are adjusted. After adjusting a single parameter, the complete detection process is performed to observe the performance change. The final detection sensitivity is 82.4%, the false positive rate is 7.8%, the field strength fitting error (RMSE) is 4.2%, and the average fault response time is 1.72 seconds. Although this method can achieve preliminary usable detection performance, it highly depends on the experience of the debugging personnel, and has problems of long time consumption, unstable tuning accuracy and poor adaptability, especially the recognition accuracy significantly decreases in extreme working conditions.

[0088] In the rule search optimization stage, a three-dimensional parameter space is constructed according to temperature, partial discharge frequency band window and time sequence sampling window using the network search method in the traditional algorithm. A total of 150 combinations are formed, and a unified data set is used to run the complete detection task for each parameter combination, and the detection sensitivity, false positive rate, fitting error and response time are recorded. The total experiment time is about 147 minutes, which is significantly more efficient than the artificial parameter tuning. The final selected optimal parameter combination achieves a sensitivity of 91.1%, a false positive rate of 5.6%, a fitting error of 3.1%, and a fault response time of 1.45 seconds. It is better than the artificial parameter tuning stage.

[0089] In the final co-evolution optimization stage, first, 100 groups of multi-modal sensor arrays are installed inside the switch cabinet to collect physical quantities such as temperature, electric field strength and partial discharge signal. After time and space synchronization and modal fusion processing, a four-dimensional space-time field tensor is constructed, which includes space, time and modal dimensions. A space-time gradient feature map is generated to depict the gradient direction and distribution trend of abnormal changes. On this basis, the system introduces a multi-objective covariance evolution algorithm, constructs a Pareto optimization model according to the two objective functions of comprehensive sensitivity and false positive rate, and guides the search direction through an adaptive covariance matrix to quickly converge to a high-quality parameter solution set. Finally, combined with the digital twin simulation environment, virtual defects are injected for verification, stability selection and deployment. The whole process takes 58 minutes, the final detection sensitivity is improved to 97.3%, the false positive rate is reduced to 2.4%, the field strength fitting error is only 1.2%, and the average response time is shortened to 1.06 seconds. This stage fully demonstrates the advantages of the technical solution of the present application in high-dimensional modeling, multi-objective optimization and adaptive regulation, and has stronger engineering adaptability and better practical performance.

[0090] The specific values are shown in Table 1 as follows:

[0091] Table 1 Comparison of performance of the present application and prior art

[0092]

[0093]

[0094] From the comparison data in Table 1, it can be seen that scheme C breaks through the limitations of low trial and error efficiency and difficulty in multi-objective trade-off of traditional methods through the innovative integration of multi-modal data fusion, physical constraint reinforcement and co-evolution strategy, embodies better robustness, adaptability and practical value, and fully proves the technical advancement and engineering generalizability of the present application.

[0095] It should be noted that the above examples are only used to illustrate the technical solutions of the present application and are not limiting. Although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A switchgear-based insulation defect feature detection method, characterized by: include, Install a sensor array to collect raw data streams of multimodal space-time physical fields, and generate a reconstructed four-dimensional space-time field tensor through a multimodal space-time collaborative perception network; Based on the four-dimensional space-time field tensor, through physical regularization, an enhanced sample set that conforms to physical laws is generated; Based on the enhanced sample set that conforms to physical laws, the optimized detection parameter combination is generated through gradient-evolution collaboration; Adopt digital twin-enhanced closed-loop adaptive optimization to convert the optimized detection parameter combination into a verified stable parameter configuration; According to the verified stable parameter configuration, the real-time optimized parameters are obtained through embedded real-time acceleration optimization and solidification.

2. The switchgear-based insulation defect feature detection method according to claim 1, wherein: The multimodal spatiotemporal collaborative perception network includes an artificial intelligence model that integrates multi-source heterogeneous sensor data, deeply fuses spatiotemporal features, and realizes cross-modal collaborative reasoning.

3. The switchgear-based insulation defect feature detection method according to claim 1, wherein: The generation of a reconstructed four-dimensional space-time field tensor through a multimodal space-time collaborative perception network includes time synchronization and spatial calibration of the collected original sensor data stream to obtain a space-time synchronized multimodal data stream; weight optimization of temperature, deformation and gas in the space-time synchronized multimodal data stream, and fusion of the modal data of temperature, deformation and gas to obtain three physical field data of temperature, deformation and gas concentration; and fusion processing of the three physical field data through spatial discretization, time serialization and blind spot compensation to generate a reconstructed four-dimensional space-time field tensor.

4. The switchgear-based insulation defect feature detection method according to claim 1, wherein: The method of generating an enhanced sample set that conforms to physical laws through physical regularization includes embedding the four-dimensional space-time field tensor through physical laws and setting physical prior parameters to obtain a generative adversarial model with physical constraints embedded; based on the generative adversarial model with physical constraints embedded, adversarial training is performed through a data distribution matching stage and a physical constraint reinforcement stage to obtain an extreme working condition sample set; based on the extreme working condition sample set, an enhanced sample set that conforms to physical laws is obtained through multi-scale frequency domain testing and conservation law joint verification.

5. The switchgear-based insulation defect feature detection method according to claim 1, wherein: The generation of optimized detection parameter combinations through gradient-evolution collaboration includes obtaining a spatiotemporal gradient feature map based on an enhanced sample set that conforms to physical laws through four-dimensional gradient calculation and feature enhancement; based on the spatiotemporal gradient feature map, dynamically searching for Pareto optimal solutions under physical constraints through a multi-objective covariance evolution strategy to generate an optimized detection parameter combination.

6. The switchgear-based insulation defect feature detection method according to claim 1, wherein: The digital twin enhanced closed-loop adaptive optimization includes obtaining a verified parameter configuration based on an optimized parameter combination through multi-fault mode injection and virtual-real data fusion verification; and obtaining a stable parameter configuration based on the verified physical prior parameter configuration through performance degradation monitoring and dynamic parameter reoptimization.

7. The switchgear-based insulation defect feature detection method according to claim 1, wherein: The real-time optimization parameters obtained by the embedded real-time acceleration optimization solidification include obtaining a hardware acceleration module based on stable parameter configuration through a four-dimensional convolution systolic array accelerator and memory calculation optimization; Based on the hardware acceleration module, real-time optimization parameters are obtained through real-time task scheduling and time-sensitive network protocols.

8. A switchgear insulation defect feature detection system, based on the switchgear insulation defect feature detection system according to any one of claims 1 to 7, characterized in that: include, The sensor fusion and reconstruction module installs a sensor array inside the switch cabinet to collect the original data stream of the multimodal space-time physical field and generate a reconstructed four-dimensional space-time field tensor through a multimodal space-time collaborative perception network; The physical regulation enhancement module generates an enhanced sample set that conforms to physical laws through physical regularization based on the four-dimensional space-time field tensor; The gradient parameter determination module generates optimized detection parameter combinations through gradient-evolution collaboration based on an enhanced sample set that conforms to physical laws; The digital twin closed-loop module uses digital twin enhanced closed-loop adaptive optimization to convert the optimized detection parameter combination into a verified stable parameter configuration; The hard-added parameter module is configured according to the verified stable parameters and solidified through embedded real-time acceleration optimization to obtain real-time optimized parameters.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of detecting insulation defect characteristics based on the switchgear according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of detecting insulation defect characteristics based on a switch cabinet according to any one of claims 1 to 7 are implemented.

Images