Switch cabinet partial discharge on-line monitoring method and storage medium

Abstract

The invention discloses a partial discharge on-line monitoring method of a switch cabinet and a storage medium, and relates to the technical field of power equipment monitoring. According to the method, multi-dimensional state parameters such as vibration amplitude and pulse density are obtained in real time by constructing a local, edge and cloud three-level processing architecture, and self-adaptive distribution of calculation tasks is achieved based on a dynamic unloading scoring formula. The local nodes execute time domain feature extraction, the edge nodes perform frequency domain enhancement and confidence evaluation, and the cloud completes time-space depth analysis. A multi-domain feature weighted fusion mechanism is innovatively adopted, the signal quality is improved in combination with a wavelet dual-threshold noise reduction algorithm, and quantization errors are corrected through a residual compensation module. When the edge node loads gt; and when 80%, dynamic resource scheduling is triggered, and resource allocation is optimized and calculated.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and particularly relates to a method for on-line partial discharge monitoring of a switch cabinet and a storage medium. Background Art

[0002] In the field of power equipment monitoring, on-line partial discharge monitoring of switch cabinets is of great significance for preventing equipment failures and ensuring the safe operation of the power grid. Most of the existing technologies use a fixed computing architecture for discharge signal analysis, and there are the following deficiencies: When a single node processes complex signals, the calculation delay is high, and it is difficult to meet the real-time requirements. The static resource allocation mode cannot dynamically adjust the calculation load according to the signal characteristics, resulting in easy overload of edge nodes or idle cloud resources; Traditional feature extraction methods are limited to single-domain analysis, lacking multi-dimensional feature fusion, which affects the classification accuracy of discharge patterns. The cumulative quantization error during cross-node transmission leads to loss of feature information, affecting the inference accuracy of deep learning models. In addition, the existing confidence evaluation methods lack a dynamic adjustment mechanism and are prone to misjudgment when node resources fluctuate. Summary of the Invention

[0003] (I) Technical Problems to be Solved To solve the above problems, the present invention proposes a method for on-line partial discharge monitoring of a switch cabinet and a storage medium, aiming to solve the problems that when a single node processes complex signals, the calculation delay is high, and traditional feature extraction methods are limited to single-domain analysis and lack multi-dimensional feature fusion, which affects the classification accuracy of discharge patterns.

[0004] (II) Technical Solutions A method for on-line partial discharge monitoring of a switch cabinet according to the present invention includes: Obtaining state monitoring parameters of the switch cabinet, where the state monitoring parameters include vibration amplitude, pulse density, signal-to-noise ratio, and edge load; Constructing a three-level processing architecture for extracting feature data of the state monitoring parameters to obtain a multi-domain feature set, dynamically generating an offloading score and a confidence level based on the multi-domain feature set, and allocating calculation tasks to the optimal node in the three-level processing architecture according to the offloading score to obtain multi-level processing results; Fusing the multi-level processing results and outputting a discharge pattern classification, generating a local discharge identification report of the switch cabinet and sending it to a terminal device.

[0005] In the present invention, the three-level processing architecture includes a local node, an edge node, and a cloud server node. The local node is used to perform time-domain feature extraction to generate time-domain statistics, the edge node is used to perform frequency-domain feature enhancement to generate a frequency-domain energy spectrum, and the cloud server node is used to perform spatio-temporal depth analysis to generate spatio-temporal depth features.

[0006] In the present invention, the obtaining of multi-level processing results by allocating computing tasks to the optimal nodes in the three-level processing architecture includes: Generating a dynamic offloading score based on real-time signal parameters and resource status, and the calculation formula of the dynamic offloading score is where is the normalized value of the state monitoring parameter, is the weight coefficient of dynamic optimization, represents the dynamic offloading score, reflecting task complexity and resource requirements, is the normalization function; Allocating tasks to corresponding nodes according to a preset threshold interval: If , the task is allocated to the local node; If , the task is allocated to the edge node; If , the task is allocated to the cloud server node.

[0007] In the present invention, the generation of the confidence includes: Setting a confidence threshold judgment module at the edge node to trigger an early exit mechanism when the confidence exceeds the preset threshold; The triggering condition of the early exit mechanism is: and , where is the classification probability distribution, is the classification confidence, is the entropy value of the classification probability distribution, and the calculation formula is where is the number of classification categories, is the predicted probability of the model for the category . When the threshold is not reached, the cloud full model calculation is triggered.

[0008] In the present invention, the model segmentation of the three-level processing architecture specifically includes: The local node runs the first three convolutional networks and implements 8-bit integer quantization; The edge node runs the fourth to fifth convolutional networks and uses FP16 mixed precision calculation; The cloud server node runs the fully connected layer and the LSTM module to output spatio-temporal fusion features.

[0009] In the present invention, it further includes an anti-interference processing step: Switching the noise reduction mode according to the real-time signal-to-noise ratio. When the signal-to-noise ratio is less than 10 dB, the wavelet double-threshold noise reduction algorithm is enabled, and the calculation formula is: where Denotes the wavelet coefficients after noise reduction, is the original wavelet coefficient, is the positive threshold, and its value is , is the negative threshold, and its value is , is the standard deviation of the noise, estimated from the background noise of the signal.

[0010] In the present invention, the edge node is configured with a residual compensation module for correcting the quantization error in cross-node transmission. The compensation formula is: where, is the compensation coefficient, is the quantization function, represents the inverse quantization function for restoring the quantized feature map, represents the feature map of the edge node after compensation, represents the original feature map of the edge node.

[0011] In the present invention, a dynamic resource scheduling strategy is further included: When the CPU load rate of the edge node is greater than 80%, the confidence threshold is forced to be reduced to 0.85, and the computing tasks of the fifth-layer convolutional network are preferentially migrated to the cloud node.

[0012] In the present invention, the fusion of the multi-domain feature sets includes: Performing weighted fusion on the time-domain statistics, frequency-domain energy spectrum, and spatio-temporal depth features. The weights are dynamically adjusted according to the signal complexity. The fusion formula is: , where, represents the final feature vector after fusion, represents the time-domain statistical features, frequency-domain energy spectrum features, represents the spatio-temporal depth features, is the dynamic weight coefficient and satisfies , is adaptively adjusted according to the signal-to-noise ratio, and increases with the increase of the pulse density.

[0013] Another computer-readable storage medium of the present invention stores a computer program, which when executed by a processor implements the on-line partial discharge monitoring method for the switchgear described in any one of the above technical solutions.

[0014] (III) Beneficial effects Compared with the prior art, the beneficial effects of the present invention are: In the present invention, through the collaborative processing of local, edge, and cloud nodes and in combination with a dynamic offloading scoring mechanism, optimal allocation of computing tasks is achieved, greatly reducing system latency; by fusing time-domain statistics, frequency-domain energy spectra, and spatio-temporal depth features, the classification accuracy of discharge patterns is greatly improved.

[0015] In the present invention, a quantization error is corrected through a residual compensation module, and in combination with dynamic weight adjustment, feature integrity is ensured; when the CPU load of the edge node > 80%, computing tasks are automatically migrated and the confidence threshold is adjusted to enhance system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following-described drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0017] Figure 1 It is a schematic diagram of the logical framework structure of the online monitoring method; Figure 2 It is a schematic diagram of the logical flow of the online monitoring method; Figure 3 It is a schematic diagram of the curve of the feature weight changing with SNR; Figure 4 It is a schematic diagram of the curve of the CPU load of the edge node changing with time. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] Embodiment 1

[0019] As Figures 1 - 4 shown, a partial discharge online monitoring method for a switch cabinet includes the following method steps: S100. Obtain the status monitoring parameters of the switch cabinet, where the status monitoring parameters include vibration amplitude, pulse density, signal-to-noise ratio, and edge load.

[0020] Generally, the partial discharge online monitoring method for a switch cabinet is usually deployed in a substation in the form of a system carrier. The present disclosure realizes the construction of an online monitoring architecture by deploying an embedded sensor node, an edge computing gateway, and a cloud GPU server on the switch cabinet.

[0021] The vibration amplitude is extracted from the time-domain waveform through a peak detection algorithm, the pulse density is calculated by counting the number of discharge pulses within a unit time, and the signal-to-noise ratio is determined by the ratio of the background noise power to the signal power. The edge load is represented by the weighted average of the CPU occupancy rate and the memory usage rate.

[0022] Specifically, a moving average filter is applied to the original signal to suppress high-frequency noise. The window length is usually set to 1 / 4 of the signal period. For example, for a 50 Hz power frequency signal, the window width is 5 ms. After filtering, the derivative method is used to locate the extreme points. When the first derivative of the signal changes from positive to negative, it is determined as the peak point. At the same time, an amplitude threshold is set, such as exceeding 3 times the standard deviation of the baseline noise, to exclude interference pulses. For the phenomenon of multiple consecutive peaks, a minimum time interval constraint is adopted, such as the adjacent peak interval being greater than 10 μs, to avoid repeated counting. The finally extracted vibration amplitude is the median absolute value of the effective peak values, and the calculation formula is: , where is the vibration amplitude, representing the median amplitude of the effective discharge pulses, represents the th peak voltage value, which is extracted by the peak detection algorithm, represents the median function, which is used to eliminate the interference of outliers.

[0023] The calculation of the pulse density is based on the statistics of the number of pulses within a discretized time window. Taking 1 second as the basic time unit, the discharge pulses within each window are counted. The pulse determination needs to meet both the amplitude condition and the timing condition: the amplitude needs to exceed the dynamic threshold, that is, the threshold is 2.5 times the peak-to-peak value of the background noise, and the pulse width is in the range of 0.1 μs to 10 ms. To avoid short-term noise interference, a double verification mechanism is adopted - if the adjacent pulse interval is less than the minimum discharge period, such as 1 ms, it is regarded as the same event. The pulse density is expressed as: , where represents the number of effective pulses, represents the time window length.

[0024] To determine the signal-to-noise ratio (SNR), the signal and background noise components need to be separated. The system collects pure noise signals during the period without discharge activity and calculates its power spectral density . When a discharge signal is detected, the total power spectrum is calculated by FFT. The signal power is the difference between the total power and the noise power: , and the SNR is expressed in decibels as: , where is the full-band noise power. To improve the real-time performance, the system uses a sliding window to iteratively update the noise power estimate, and the window length is usually 30 seconds.

[0025] The comprehensive evaluation of the edge load combines the CPU and memory utilization rates, reflecting the real-time resource pressure of the computing node. The CPU occupancy rate is obtained by reading the operating system process status information, and the memory usage rate Calculated as the ratio of used memory to total memory. Since CPU computing tasks are more sensitive to latency, the load weight is allocated with 70% for CPU and 30% for memory, and the calculation formula is: . This weight configuration is verified through experiments. When the CPU load exceeds 80%, the task processing latency increases significantly, while the memory pressure has relatively little impact on performance. The dynamic load data is sampled every 200 ms and used for resource scheduling decisions after exponential smoothing filtering. The smoothing formula is: To avoid misjudgment caused by instantaneous fluctuations, where is the current sample, the original load value actually measured by the edge node in the latest sampling period, is the filtered value of the previous period, that is, the smoothed load value calculated at the previous sampling moment , which is used to retain the historical state and avoid misjudgment of load mutations, such as short-term peaks. The system introduces a hysteresis threshold in the load balancing algorithm (such as restoring local computing only when the load drops to 75%), further improving the stability.

[0026] S200. Construct a three-level processing architecture to extract the feature data of the state monitoring parameters to obtain a multi-domain feature set, dynamically generate an offloading score and confidence based on the multi-domain feature set, and allocate the computing task to the optimal node in the three-level processing architecture based on the offloading score to obtain a multi-level processing result.

[0027] Specifically, the three-level processing architecture includes a local node, an edge node, and a cloud server node. The local node is used to perform time-domain feature extraction to generate time-domain statistics, the edge node is used to perform frequency-domain feature enhancement to generate a frequency-domain energy spectrum, and the cloud server node is used to perform spatio-temporal depth analysis to generate spatio-temporal depth features.

[0028] In the three-level processing architecture, the multi-domain feature extraction of the state monitoring parameters is achieved through the collaborative processing of local, edge, and cloud nodes. The local node first receives the original signal, including time-domain data such as vibration amplitude and pulse density, and performs preprocessing and normalization operations. That is, the multi-domain feature set includes time-domain data such as vibration amplitude and pulse density, as well as feature maps.

[0029] The preprocessed signal is converted into an input tensor of 224×224×3 and input into the first 3 convolutional layers (Conv1~Conv3) of the local node. The convolutional kernel sizes are 7×7 (Conv1) and 3×3 (Conv2~Conv3) respectively, and through stride and pooling operations, the sampling is gradually reduced, and finally a feature map of 56×56×64 is output. This feature map contains time-domain local patterns, such as the steepness of the pulse rising edge and waveform periodicity. The calculation formula is: , where is the normalized signal input, Indicates the first to fifth layers of the convolutional network hierarchy.

[0030] After the local node receives the original signals such as vibration amplitude and pulse density, it performs time-domain feature extraction, including convolution operation on the signal and calculation of time-domain statistics. Specifically, the first 3 layers of the convolutional network (Conv1~Conv3) extract basic features from the preprocessed signal, generate a feature map of 56×56×64, and compress and transmit it through 8-bit integer quantization (INT8). The quantization formula is: The quantized feature map is transmitted to the edge node. The edge node performs dequantization and introduces a residual compensation module to correct the error. The compensation coefficient , and the correction formula is: Where, is the compensation coefficient, is the quantization function, represents the dequantization function, which restores the quantized feature map, represents the feature map of the edge node after compensation, represents the original feature map of the edge node.

[0031] The edge node further runs the 4th to 5th layers of the convolutional network (Conv4~Conv5), and uses FP16 mixed-precision calculation to generate an intermediate feature map of 14×14×256. At the same time, it extracts the frequency-domain energy spectrum through the fast Fourier transform (FFT) and calculates the energy ratio of each frequency band: , where , where represents the wavelet coefficient after noise reduction, is the original wavelet coefficient, is the positive threshold, and the value is , is the negative threshold, and the value is , is the noise standard deviation, which is estimated through the signal background noise.

[0032] After the cloud server receives the features transmitted by the edge node, it runs the fully connected layer and the LSTM module, analyzes the feature sequence of consecutive time windows, and outputs the spatio-temporal fusion feature vector , and captures the periodic law of the discharge signal.

[0033] The dynamic offloading scoring system comprehensively combines parameters such as vibration amplitude, pulse density, signal-to-noise ratio, and edge load in real time, and generates a score after normalization and weighting. The normalization function uses linear scaling. For example, the vibration amplitude The normalization interval is , and the corresponding formula is: , and the weight coefficient is dynamically optimized through reinforcement learning. The initial setting is (vibration amplitude), (pulse density), (Signal-to-Noise Ratio), (Edge Load), and the comprehensive scoring formula is: Wherein, is the dynamic offloading score, a scoring value comprehensively evaluating the task complexity and resource requirements, ranging from 0 to 1, used for decision-making on task allocation to local nodes, edge nodes or cloud server nodes, is the normalized value of the vibration amplitude, that is, the normalized version of the vibration amplitude scaled the original value to the interval [0, 1] to eliminate the influence of dimensions and make different parameters comparable, is the normalized value of the pulse density, is the normalized value of the signal-to-noise ratio, is the normalized value of the edge load. According to the scoring threshold interval, the task allocation rules are as follows: If , the task is directly processed by the local node and the early exit mechanism is triggered; if , the task is forwarded to the edge node for frequency domain feature enhancement and confidence calculation; if , the task is uploaded to the cloud for full model depth analysis.

[0034] In the optimization of the weight coefficient of the dynamic offloading score, the system uses a reinforcement learning algorithm based on Q-learning to adjust the weight in real time. Specifically, the local node collects data on task processing delay, classification accuracy and resource utilization every 5 minutes, constructs a state space such as the current SNR, pulse density, edge load, and the action space is the fine-tuning of the weight coefficient The reward function is defined as: Where represents the total reward value, used to quantify the current system policy, such as the comprehensive effect of weight adjustment, and the higher the value, the better the policy, represents the classification accuracy, referring to the correct rate of partial discharge pattern classification, such as the proportion of correctly identifying the discharge type, with a value range of 0~1 or 0%~100%, Task processing delay, representing the total time from data acquisition to generating classification results, including the calculation and transmission time of local, edge or cloud nodes, represents the real-time CPU usage rate of the edge node, expressed as a percentage, reflecting the resource pressure of the computing node. Explore the optimal weight combination through the ε-greedy strategy. For example, when the pulse density suddenly increases, the system automatically increases the pulse density weight to 0.35 to give priority to ensuring the timeliness of task offloading.

[0035] After the early exit mechanism is triggered, the edge node directly outputs the current classification result to the terminal device and generates a lightweight report, including confidence, entropy value, and feature summary. At the same time, it sends a termination instruction to the cloud to avoid redundant calculations. The system records the historical data of the early exit decisions. If the classifications of the early exit results are consistent with the results of the full model in the cloud for three consecutive times, the confidence threshold is dynamically increased to 0.96; otherwise, it is decreased to 0.93, forming an adaptive feedback loop.

[0036] The confidence generation module evaluates the reliability of the classification result at the edge node. The confidence is determined by the maximum probability value output by Softmax, and the calculation formula is: , where is the prediction probability of the model for class . At the same time, the entropy value of the classification probability distribution is calculated to evaluate the uncertainty, and the calculation formula is , where is the number of classification categories. The conditions for triggering the early exit mechanism are: and , is the classification probability distribution, is the classification confidence, is the entropy value of the classification probability distribution. When the threshold is not reached, the task continues to be uploaded to the cloud to complete the full model inference.

[0037] When the task is migrated to the cloud, the edge node starts the data synchronization protocol: packs the 14×14×256 feature map output by the Conv4 layer and the quantization parameters into data blocks, attaches a timestamp and a CRC check code, and uploads them in fragments through the breakpoint resumption mechanism or the UDP protocol. After receiving the data, the cloud restores the calculation state based on the context cache of the LSTM module to ensure the continuity of spatio-temporal analysis. The cloud records the received data block index. After a network interruption, the edge node only retransmits the missing fragments.

[0038] If the network interruption times out, the edge node automatically reverts to local simplified model processing, marks the result confidence as "low reliability", and caches the data that has not been uploaded.

[0039] In dynamic resource scheduling, if the CPU load rate of the edge node exceeds 80%, the system automatically lowers the confidence threshold from 0.95 to 0.85 and migrates the Conv5 layer calculation task to the cloud to ensure that the edge node load rate is lower than 30%. In addition, when the signal-to-noise ratio is lower than 10 dB, the wavelet double-threshold denoising algorithm is enabled, and the noise standard deviation

[0040] S300. After fusing the multi-level processing results, output the discharge mode classification, generate a switchgear partial discharge identification report, and send it to the terminal device.

[0041] The multi-domain feature set generates the final classification feature through weighted fusion, and the weights are dynamically adjusted according to the signal complexity. The fusion formula is , where represents the final feature vector after fusion, represents the time-domain statistical feature, represents the frequency-domain energy spectrum feature, represents the spatio-temporal depth feature, is the dynamic weight coefficient and satisfies , is adaptively adjusted according to the signal-to-noise ratio, increases with the increase of pulse density. For example, when the pulse density is , the spatio-temporal feature weight

[0042] The multi-domain feature fusion weight adopts a piecewise linear interpolation rule: when is , ; when is , . The spatio-temporal feature weight and the pulse density are related as . When is , it is locked at 0.6. When performing the above piecewise linear interpolation rule, the weight ratio is verified every 200 ms, that is . If it exceeds the limit, it is scaled proportionally.

[0043] Example 2

[0044] The embodiment of the present invention provides a computer-readable storage medium.

[0045] The computer program stored on the computer-readable storage medium provided by the embodiment of the present invention can implement the steps of any one of the above-mentioned on-line partial discharge monitoring methods for switchgear when executed by a processor.

[0046] The computer-readable storage medium may include: various media such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks, or optical discs that can store program codes.

[0047] For the introduction of the computer-readable storage medium provided by the embodiment of the present invention, please refer to the above method embodiment, and the present invention will not be elaborated here.

[0048] The various embodiments in the specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple. For the relevant parts, reference can be made to the description in the method section.

[0049] Those skilled in the art can further realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described according to functions in the above description. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention.

[0050] The steps of the methods or algorithms described in combination with the embodiments disclosed herein can be directly implemented by hardware, software modules executed by a processor, or a combination of both. The software modules can be placed in a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0051] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made by those of ordinary skill in the art to the technical solution of the present invention shall fall within the protection scope of the present invention. The technical content claimed by the present invention has been fully recorded in the claims.

Claims

1. An on-line partial discharge monitoring method for a switchgear, characterized in that, Including: Obtain the status monitoring parameters of the switchgear, where the status monitoring parameters include vibration amplitude, pulse density, signal-to-noise ratio, and edge load; Construct a three-level processing architecture to extract the feature data of the status monitoring parameters to obtain a multi-domain feature status monitoring parameter set, dynamically generate an offloading score and confidence based on the multi-domain feature set, and allocate the computing task to the optimal node in the three-level processing architecture based on the offloading score to obtain multi-level processing results; Fuse the multi-level processing results and output a discharge mode classification, generate a local discharge identification report for the switchgear and send it to the terminal device.

2. The on-line partial discharge monitoring method for the switchgear according to claim 1, characterized in that, The three-level processing architecture includes a local node, an edge node, and a cloud server node. The local node is used to perform time-domain feature extraction to generate time-domain statistics, the edge node is used to perform frequency-domain feature enhancement to generate a frequency-domain energy spectrum, and the cloud server node is used to perform spatio-temporal depth analysis to generate spatio-temporal depth features.

3. The on-line partial discharge monitoring method for the switchgear according to claim 2, wherein, The allocation of the computing task to the optimal node in the three-level processing architecture to obtain multi-level processing results includes: Generate a dynamic offloading score based on real-time signal parameters and resource status. The calculation formula of the dynamic offloading score is , where is the normalized value of the status monitoring parameter, is the weight coefficient of dynamic optimization, represents the dynamic offloading score, reflecting the task complexity and resource requirements, is the normalization function; Allocate tasks to the corresponding nodes according to a preset threshold interval: If , the task is assigned to the local node; If , the task is assigned to the edge node; If , the task is assigned to the cloud server node.

4. The on-line partial discharge monitoring method for the switchgear according to claim 1, characterized in that, The generation of the confidence includes: Set a confidence threshold judgment module at the edge node, and trigger an early exit mechanism when the confidence exceeds the preset threshold; The triggering conditions of the early exit mechanism are as follows: and , where is the classification probability distribution, is the classification confidence, is the entropy value of the classification probability distribution, and the calculation formula is , where is the number of classification categories, is the predicted probability of the model for category . When the threshold is not reached, the cloud full model calculation is triggered.

5. The on-line partial discharge monitoring method for the switchgear according to claim 4, characterized in that, The model segmentation of the three-level processing architecture specifically includes: The local node runs the first three layers of the convolutional network and implements 8-bit integer quantization; The edge node runs the fourth to fifth layers of the convolutional network and uses FP16 mixed-precision calculation; The cloud server node runs the fully connected layer and the LSTM module, and outputs spatio-temporal fusion features.

6. The on-line partial discharge monitoring method for the switchgear according to claim 5, characterized in that, It also includes an anti-interference processing step: Switch the noise reduction mode according to the real-time signal-to-noise ratio. When the signal-to-noise ratio is less than 10 dB, enable the wavelet double-threshold noise reduction algorithm. The calculation formula is: , where represents the wavelet coefficient after noise reduction, is the original wavelet coefficient, is the positive threshold, and its value is , is the negative threshold, and its value is , is the standard deviation of the noise, which is estimated through the background noise of the signal.

7. The on-line partial discharge monitoring method for the switchgear according to claim 6, characterized in that The edge node is configured with a residual compensation module for correcting the quantization error in cross-node transmission. The compensation formula is as follows: Among them, is the compensation coefficient, is the quantization function, represents the inverse quantization function, which restores the quantized feature map, represents the feature map of the edge node after compensation, represents the original feature map of the edge node.

8. The on-line partial discharge monitoring method for the switchgear according to claim 1, characterized in that It also includes a dynamic resource scheduling strategy: When the CPU load rate of the edge node is greater than 80%, forcibly reduce the confidence threshold to 0.85, and preferentially migrate the computing task of the fifth layer of the convolutional network to the cloud node.

9. The on-line partial discharge monitoring method for the switchgear according to claim 1 or 8, characterized in that The fusion of the multi-domain feature set includes: Perform weighted fusion on time-domain statistics, frequency-domain energy spectrum, and spatio-temporal depth features. The weights are dynamically adjusted according to the signal complexity. The fusion formula is: , where represents the final feature vector after fusion, represents the time-domain statistical features, frequency-domain energy spectrum features, represents the spatio-temporal depth features, is the dynamic weight coefficient and satisfies , is adaptively adjusted according to the signal-to-noise ratio, increases with the increase of pulse density.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the on-line partial discharge monitoring method for the switchgear according to any one of claims 1-9.

Images