A partial discharge online monitoring method and storage medium for switch cabinet

Abstract

The present invention discloses an online monitoring method and storage medium for partial discharge of a switch cabinet, and relates to the technical field of power equipment monitoring. The method constructs a three-level processing architecture of local, edge, and cloud to obtain multi-dimensional state parameters such as vibration amplitude and pulse density in real time, and realizes adaptive allocation of computing tasks based on a dynamic unloading scoring formula. The local node performs time domain feature extraction, the edge node performs frequency domain enhancement and confidence assessment, and the cloud completes spatiotemporal depth analysis. An innovative multi-domain feature weighted fusion mechanism is adopted, combined with a wavelet dual-threshold noise reduction algorithm to improve signal quality, and the quantization error is corrected by a residual compensation module. When the edge node load is >80%, dynamic resource scheduling is triggered to optimize computing resource allocation.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and in particular to an online partial discharge monitoring method and a storage medium for a switch cabinet. Background Art

[0002] In the field of power equipment monitoring, online monitoring of partial discharge in switchgear is of great significance for preventing equipment failures and ensuring the safe operation of the power grid. Existing technologies mostly use fixed computing architectures for discharge signal analysis, which has the following shortcomings:

[0003] When a single node processes complex signals, the computational latency is high, making it difficult to meet real-time requirements. The static resource allocation model cannot dynamically adjust the computational load based on signal characteristics, resulting in overloaded edge nodes and idle cloud resources.

[0004] Traditional feature extraction methods are limited to single-domain analysis and lack multi-dimensional feature fusion, which affects the accuracy of discharge pattern classification. Accumulated quantization errors in cross-node transmission lead to feature information loss, affecting the inference accuracy of deep learning models. Furthermore, existing confidence assessment methods lack dynamic adjustment mechanisms, making them prone to misjudgment when node resources fluctuate. Summary of the Invention

[0005] (1) Technical issues to be resolved

[0006] To solve the above problems, the present invention proposes an online monitoring method and storage medium for partial discharge of switchgear, aiming to solve the problems in the prior art such as high computational delay when a single node processes complex signals, traditional feature extraction methods being limited to single domain analysis, and lacking multi-dimensional feature fusion, which affects the accuracy of discharge pattern classification.

[0007] (2) Technical solution

[0008] A method for online monitoring of partial discharge of a switch cabinet according to the present invention comprises:

[0009] Acquiring status monitoring parameters of the switchgear, wherein the status monitoring parameters include vibration amplitude, pulse density, signal-to-noise ratio, and edge load;

[0010] Constructing a three-level processing architecture for extracting characteristic data of the condition 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 computing tasks to optimal nodes in the three-level processing architecture based on the offloading score to obtain a multi-level processing result;

[0011] After fusing the multi-level processing results, the discharge mode classification is output, and a switch cabinet partial discharge identification report is generated and sent to the terminal device.

[0012] In the present invention, the three-level processing architecture includes local nodes, edge nodes and cloud server nodes. The local nodes are used to perform time domain feature extraction to generate time domain statistics, the edge nodes are used to perform frequency domain feature enhancement to generate frequency domain energy spectrum, and the cloud server nodes are used to perform spatiotemporal depth analysis to generate spatiotemporal depth features.

[0013] In the present invention, allocating computing tasks to optimal nodes in the three-level processing architecture to obtain multi-level processing results includes:

[0014] A dynamic offloading score is generated based on real-time signal parameters and resource status. The calculation formula of the dynamic offloading score is: ,in is the original value of the condition monitoring parameter, is the weight coefficient for dynamic optimization, represents the dynamic offloading score, reflecting the task complexity and resource requirements, is the normalization function;

[0015] Assign tasks to corresponding nodes based on the preset threshold range:

[0016] like , the tasks are assigned to the local node;

[0017] like , tasks are assigned to edge nodes;

[0018] like , the tasks are assigned to the cloud server nodes.

[0019] In the present invention, the generation of the confidence level includes:

[0020] A confidence threshold judgment module is set up at the edge node, and the early exit mechanism is triggered when the confidence exceeds the preset threshold;

[0021] The triggering conditions of the early exit mechanism are:

[0022] and ,in, is the classification probability distribution, is the classification confidence, is the entropy value of the classification probability distribution, and the calculation formula is ,in is the number of classification categories, For the model category If the predicted probability does not reach the threshold, the full model calculation in the cloud will be triggered.

[0023] In the present invention, the model segmentation of the three-level processing architecture specifically includes:

[0024] The local node runs the first three layers of the convolutional network and implements 8-bit integer quantization;

[0025] The edge nodes run the fourth to fifth layers of the convolutional network and use FP16 mixed precision calculations;

[0026] The cloud server node runs the fully connected layer and LSTM module to output spatiotemporal fusion features.

[0027] The present invention also includes an anti-interference processing step:

[0028] The noise reduction mode is switched according to the real-time signal-to-noise ratio. When the signal-to-noise ratio is less than 10dB, the wavelet double-threshold noise reduction algorithm is enabled. The calculation formula is: ,in represents the wavelet coefficients after denoising, are the original wavelet coefficients, is the positive threshold, and its value is , is the negative threshold, and its value is , is the noise standard deviation, estimated by the signal background noise.

[0029] In the present invention, the edge node is configured with a residual compensation module to correct the quantization error of cross-node transmission. The compensation formula is:

[0030] in, is the compensation coefficient, is the quantization function, Represents the inverse quantization function, which restores the quantized feature map. represents the edge node feature map after compensation, Represents the original feature map of the edge node.

[0031] The present invention also includes a dynamic resource scheduling strategy:

[0032] When the CPU load rate of the edge node is greater than 80%, the confidence threshold is forced to be lowered to 0.85, and the computing tasks of the fifth-layer convolutional network are preferentially migrated to the cloud node.

[0033] In the present invention, the fusion of the multi-domain feature set includes:

[0034] The time domain statistics, frequency domain energy spectrum and time-space depth features are weightedly fused, and the weights are dynamically adjusted according to the signal complexity. The fusion formula is:

[0035] ,in, represents the final eigenvector after fusion, represents the time domain statistical characteristics, Frequency domain energy spectrum characteristics, Represents the spatiotemporal depth characteristics, is a dynamic weight coefficient and satisfies , Adaptive adjustment based on signal-to-noise ratio, Increases with the increase of pulse density.

[0036] Another computer-readable storage medium of the present invention stores a computer program, which, when executed by a processor, implements the method for online monitoring of partial discharge of a switch cabinet as described in any one of the above technical solutions.

[0037] (3) Beneficial effects

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] The present invention achieves optimal allocation of computing tasks through collaborative processing of local, edge, and cloud nodes, combined with a dynamic offloading scoring mechanism, greatly reducing system latency; and integrates time domain statistics, frequency domain energy spectrum, and time-space depth features to greatly improve the accuracy of discharge mode classification.

[0040] In this invention, the quantization error is corrected through the residual compensation module, combined with dynamic weight adjustment to ensure feature integrity; when the CPU load of the edge node is greater than 80%, the computing task is automatically migrated and the confidence threshold is adjusted to improve system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 Schematic diagram of the logical framework structure of the online monitoring method;

[0043] Figure 2 Schematic diagram of the logic flow of the online monitoring method;

[0044] Figure 3 Schematic diagram of the curve showing the change of feature weight with SNR;

[0045] Figure 4 The following is a curve diagram showing the change of edge node CPU load over time. DETAILED DESCRIPTION

[0046] Example

[0047] like Figures 1-4The method for online monitoring of partial discharge of a switch cabinet shown includes the following steps:

[0048] S100: Acquire status monitoring parameters of a switch cabinet, where the status monitoring parameters include vibration amplitude, pulse density, signal-to-noise ratio, and edge load.

[0049] Typically, the online monitoring method for partial discharge of switchgear is usually deployed in the form of a system carrier in a substation. The present invention realizes the architecture construction of online monitoring by deploying embedded sensor nodes, edge computing gateways and cloud GPU servers on the switchgear.

[0050] The vibration amplitude is extracted from the time-domain waveform using a peak detection algorithm. The pulse density is calculated by counting the number of discharge pulses per unit time. The signal-to-noise ratio is determined by the ratio of background noise power to signal power. Edge load is represented by the weighted average of CPU utilization and memory usage.

[0051] Specifically, the original signal is subjected to a sliding average filter to suppress high-frequency noise. The window length is usually set to 1 / 4 of the signal period. For example, for a 50Hz power frequency signal, the window width is 5ms. After filtering, the derivative method is used to locate the extreme point. When the first-order derivative of the signal changes from positive to negative, it is determined to be 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 continuous occurrence of multi-peak phenomena, a minimum time interval constraint is adopted, such as the interval between adjacent peaks is greater than 10μs to avoid repeated counting. The final extracted vibration amplitude is the median of the absolute amplitude of the effective peak, and the calculation formula is:

[0052] ,in is the vibration amplitude, which represents the median amplitude of the effective discharge pulse, Indicates the The voltage value of each peak point is extracted through the peak detection algorithm. Represents the median function, which is used to eliminate the interference of outliers.

[0053] The calculation of pulse density is based on the statistics of the number of pulses in a discretized time window. Taking 1 second as the basic time unit, the discharge pulses in each window are counted. Pulse determination must meet both amplitude conditions and timing conditions: the amplitude must 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 0.1μs to 10ms. In order to avoid short-term noise interference, a double verification mechanism is adopted - adjacent pulses with an interval less than the minimum discharge period, for example 1ms, are considered to be the same event. Pulse density The expression is: ,in Indicates the number of effective pulses, Indicates the length of the time window.

[0054] The determination of the signal-to-noise ratio (SNR) requires separating the signal from the background noise components. The system collects pure noise signals during the period of no discharge activity and calculates their power spectral density. When the discharge signal is detected, the total power spectrum Calculated by FFT, the signal power is the difference between the total power and the noise power: , SNR is expressed in decibels as: ,in is the noise power across the entire frequency band. To improve real-time performance, the system uses a sliding window to iteratively update the noise power estimate, with a window length typically set to 30 seconds.

[0055] The comprehensive evaluation of edge load combines CPU and memory utilization to reflect the real-time resource pressure of computing nodes. Obtain memory usage by reading operating system process status information Calculated as the ratio of used memory to total memory. Since CPU computing tasks are more sensitive to latency, the load weight distribution is 70% for CPU and 30% for memory. The calculation formula is: This weight configuration has been verified experimentally. When the CPU load exceeds 80%, task processing latency increases significantly, while memory pressure has a relatively small impact on performance. Dynamic load data is sampled every 200ms and used for resource scheduling decisions after exponential smoothing filtering. The smoothing formula is:

[0056] In order to avoid misjudgment caused by instantaneous fluctuations, is the current sampling, the original load value measured by the edge node in the latest sampling period, is the filter value of the previous cycle, that is, the previous sampling moment The calculated smoothed load value is used to preserve historical status and avoid misjudging sudden load changes, such as short-term peaks. The system introduces a hysteresis threshold in the load balancing algorithm (for example, local calculations are resumed only when the load drops to 75%), further improving stability.

[0057] S200. Construct a three-level processing architecture for extracting feature data of the state monitoring parameters to obtain a multi-domain feature set, dynamically generate an offloading score and a confidence level based on the multi-domain feature set, and allocate computing tasks to the optimal node in the three-level processing architecture based on the offloading score to obtain a multi-level processing result.

[0058] Specifically, the three-level processing architecture includes local nodes, edge nodes and cloud server nodes. The local nodes are used to perform time domain feature extraction to generate time domain statistics, the edge nodes are used to perform frequency domain feature enhancement to generate frequency domain energy spectrum, and the cloud server nodes are used to perform spatiotemporal depth analysis to generate spatiotemporal depth features.

[0059] In this three-level processing architecture, multi-domain feature extraction of condition monitoring parameters is achieved through collaborative processing among local, edge, and cloud nodes. The local node first receives the raw signal, including time-domain data such as vibration amplitude and pulse density, and performs preprocessing and normalization. The resulting multi-domain feature set includes time-domain data such as vibration amplitude and pulse density, as well as feature maps.

[0060] The preprocessed signal is converted into a 224×224×3 input tensor and fed into the first three convolutional layers (Conv1–Conv3) of the local node. The convolution kernels are 7×7 (Conv1) and 3×3 (Conv2–Conv3), respectively. The convolution kernels are gradually downsampled through stride and pooling operations, ultimately outputting a 56×56×64 feature map. This feature map contains local temporal patterns, such as the steepness of the pulse's rising edge and the periodicity of the waveform. The calculation formula is:

[0061] ,in is the normalized signal input, Represents the first to fifth layers of the convolutional network.

[0062] After receiving the original signal, such as vibration amplitude and pulse density, the local node performs time-domain feature extraction, which includes convolution operations on the signal and calculation of time-domain statistics. Specifically, the first three convolutional networks (Conv1-Conv3) extract basic features from the preprocessed signal, generating a 56×56×64 feature map, which is compressed and transmitted using 8-bit integer quantization (INT8). The quantization formula is: The quantized feature map is transmitted to the edge node, which performs inverse quantization and introduces a residual compensation module to correct the error. The compensation coefficient , the correction formula is: in, is the compensation coefficient, is the quantization function, Represents the inverse quantization function, which restores the quantized feature map. represents the edge node feature map after compensation, Represents the original feature map of the edge node.

[0063] The edge node further runs the 4th and 5th convolutional networks (Conv4-Conv5), using FP16 mixed precision calculation to generate a 14×14×256 intermediate feature map. At the same time, it uses fast Fourier transform (FFT) to extract the frequency domain energy spectrum and calculate the energy proportion of each frequency band: ,in ,in represents the wavelet coefficients after denoising, are the original wavelet coefficients, is the positive threshold, and its value is , is the negative threshold, and its value is , is the noise standard deviation, estimated by the signal background noise.

[0064] After receiving the features transmitted by the edge node, the cloud server runs the fully connected layer and LSTM module to analyze the feature sequence of the continuous time window and output the spatiotemporal fusion feature vector , capturing the periodicity of the discharge signal.

[0065] The dynamic unloading scoring system integrates parameters such as vibration amplitude, pulse density, signal-to-noise ratio and edge load in real time, normalizes them and weights them to generate a score. The normalization function uses linear scaling, for example, the vibration amplitude The normalized interval of , the corresponding formula is: , the weight coefficient is dynamically optimized through reinforcement learning, and the initial setting (vibration amplitude), (Pulse density), (Signal-to-noise ratio), (edge ​​load), the comprehensive scoring formula is:

[0066] in, Dynamic offloading score, a comprehensive evaluation of task complexity and resource requirements, ranging from 0 to 1, is used to decide whether to allocate tasks to local nodes, edge nodes, or cloud server nodes. is the normalized value of the vibration amplitude, that is, the vibration amplitude The normalized version scales the original value to the interval [0,1] to eliminate the dimension effect 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 rule is as follows: , the task is directly processed by the local node and triggers the early exit mechanism; 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 to perform in-depth analysis of the full model.

[0067] In the optimization of the weight coefficient of dynamic offloading score, the system uses the reinforcement learning algorithm based on Q-learning to adjust the weight in real time. Specifically, the local node collects task processing delay, classification accuracy and resource utilization data every 5 minutes, constructs the state space such as current SNR, pulse density, edge load, and the action space is the weight coefficient. Fine-tuning, the reward function is defined as:

[0068] ,in Indicates the total reward value, which is used to quantify the current system strategy, such as the comprehensive effect of weight adjustment. The higher the value, the better the strategy. Indicates the classification accuracy, which refers to the accuracy of partial discharge pattern classification, such as the proportion of correctly identified discharge types, with a value range of 0~1 or 0%~100%. Task processing latency refers to the total time from data collection to generating classification results, including the calculation and transmission time of local, edge or cloud nodes. Indicates the real-time CPU usage of edge nodes, expressed as a percentage, reflecting the resource pressure of computing nodes. The optimal weight combination is explored 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 prioritize the timeliness of task offloading.

[0069] When the early exit mechanism is triggered, the edge node directly outputs the current classification result to the terminal device and generates a lightweight report containing confidence, entropy, and feature summary. It also sends a termination command to the cloud to avoid redundant computation. The system records historical data on early exit decisions. If the classification of three consecutive early exit results is consistent with the full cloud model, the confidence threshold is dynamically increased to 0.96. Otherwise, it is lowered to 0.93, forming an adaptive feedback loop.

[0070] The confidence generation module evaluates the reliability of the classification results at the edge node. The confidence is determined by the maximum probability value output by Softmax, and the calculation formula is:

[0071] ,in For the model category The predicted probability of . At the same time, the entropy of the classification probability distribution is calculated to evaluate the uncertainty. The calculation formula is ,in 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. If it does not reach the threshold, the task continues to be uploaded to the cloud to complete the full model inference.

[0072] When the task is migrated to the cloud, the edge node initiates a data synchronization protocol: the 14×14×256 feature map and quantization parameters output by the Conv4 layer are packaged into data blocks, timestamped and CRC checksums are added, and the data is uploaded piecemeal via a resumable transmission mechanism or the UDP protocol. Upon receipt by the cloud, the computational state is restored using the context cache of the LSTM module to ensure continuity of spatiotemporal analysis. The cloud records the index of received data blocks. If the network is interrupted, the edge node only retransmits the missing pieces.

[0073] If the network interruption times out, the edge node automatically falls back to local simplified model processing, marks the result confidence as "low reliability", and caches the unuploaded data.

[0074] 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 computing tasks to the cloud to ensure that the edge node load rate is less than 30%. In addition, when the signal-to-noise ratio is lower than 10dB, the wavelet dual-threshold noise reduction algorithm is enabled, and the noise standard deviation is 0. Estimated by background signal.

[0075] S300: Fusing the multi-level processing results, outputting a discharge mode classification, generating a switch cabinet partial discharge identification report, and sending the report to a terminal device.

[0076] The multi-domain feature set generates the final classification feature through weighted fusion, and the weight is dynamically adjusted according to the signal complexity. The fusion formula is: ,in, represents the final eigenvector after fusion, represents the time domain statistical characteristics, Frequency domain energy spectrum characteristics, Represents the spatiotemporal depth characteristics, is a dynamic weight coefficient and satisfies , Adaptive adjustment based on signal-to-noise ratio, As the pulse density increases, it increases. For example, when the pulse density Time, spatiotemporal feature weight Increased to 0.6.

[0077] The multi-domain feature fusion weight adopts piecewise linear interpolation rule: when hour, , ;when hour , . Spatial-temporal feature weights and pulse density The relationship is ,when hour, When executing the above piecewise linear interpolation rule, the weight ratio is checked every 200ms, that is, If the limit is exceeded, it will be scaled proportionally. Example

[0078] An embodiment of the present invention provides a computer-readable storage medium.

[0079] The computer-readable storage medium provided in the embodiment of the present invention stores a computer program, and when the computer program is executed by a processor, the steps of any of the above-mentioned methods for online monitoring of partial discharge of a switch cabinet can be implemented.

[0080] The computer-readable storage medium may include: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc., which can store program codes.

[0081] For an introduction to the computer-readable storage medium provided in an embodiment of the present invention, please refer to the above method embodiment, and the present invention will not elaborate on it here.

[0082] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0083] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the above description has generally described the components and steps of each example according to their functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0084] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein may be implemented directly using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0085] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the concept and scope of the present invention. Any modifications and improvements made to the technical solution of the present invention by a person of ordinary skill in the art without departing from the design concept of the present invention shall fall within the scope of protection of the present invention. The technical content for which protection is sought in the present invention is fully set forth in the claims.

Claims

1. A method for online monitoring of partial discharge of a switch cabinet, characterized in that: include: Acquiring status monitoring parameters of the switchgear, wherein the status monitoring parameters include vibration amplitude, pulse density, signal-to-noise ratio, and edge load; Constructing a three-level processing architecture for extracting characteristic data of the condition 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 computing tasks to optimal nodes in the three-level processing architecture based on the offloading score to obtain a multi-level processing result; After fusing the multi-level processing results, output the discharge mode classification, generate the switch cabinet partial discharge identification report and send it to the terminal device; The three-level processing architecture includes local nodes, edge nodes, and cloud server nodes. The local nodes are used to perform time domain feature extraction to generate time domain statistics, the edge nodes are used to perform frequency domain feature enhancement to generate frequency domain energy spectrum, and the cloud server nodes are used to perform spatiotemporal depth analysis to generate spatiotemporal depth features. A dynamic offloading score is generated based on real-time signal parameters and resource status. The calculation formula of the dynamic offloading score is: ,in is the original value of the condition monitoring parameter, is the weight coefficient for dynamic optimization, represents the dynamic offloading score, reflecting the task complexity and resource requirements, is the normalization function; Assign tasks to corresponding nodes based on the preset threshold range: like , the tasks are assigned to the local node; like , tasks are assigned to edge nodes; like , the tasks are assigned to the cloud server nodes.

2. The method for online monitoring of partial discharge of a switch cabinet according to claim 1, characterized in that: The generation of the confidence level includes: A confidence threshold judgment module is set up at the edge node, and the early exit mechanism is triggered when the confidence exceeds the preset threshold; The triggering conditions of the early exit mechanism are: and ,in, is the classification probability distribution, is the classification confidence, is the entropy value of the classification probability distribution, and the calculation formula is ,in is the number of classification categories, For the model category If the predicted probability does not reach the threshold, the full model calculation in the cloud will be triggered.

3. The method for online monitoring of partial discharge of a switch cabinet according to claim 2, 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 nodes run the fourth to fifth layers of the convolutional network and use FP16 mixed precision calculations; The cloud server node runs the fully connected layer and LSTM module to output spatiotemporal fusion features.

4. The method for online monitoring of partial discharge of a switch cabinet according to claim 3, characterized in that: It also includes anti-interference processing steps: The noise reduction mode is switched according to the real-time signal-to-noise ratio. When the signal-to-noise ratio is less than 10dB, the wavelet double-threshold noise reduction algorithm is enabled. The calculation formula is: ,in represents the wavelet coefficients after denoising, are the original wavelet coefficients, is the positive threshold, and its value is , is the negative threshold, and its value is , is the noise standard deviation, estimated by the signal background noise.

5. The method for online monitoring of partial discharge of a switch cabinet according to claim 4, characterized in that: The edge node is configured with a residual compensation module to correct the quantization error of cross-node transmission. The compensation formula is: in, is the compensation coefficient, is the quantization function, Represents the inverse quantization function, which restores the quantized feature map. represents the edge node feature map after compensation, Represents the original feature map of the edge node.

6. The method for online monitoring of partial discharge of a switch cabinet according to claim 1, characterized in that: It also includes dynamic resource scheduling strategies: When the CPU load rate of the edge node is greater than 80%, the confidence threshold is forced to be lowered to 0.85, and the computing tasks of the fifth-layer convolutional network are preferentially migrated to the cloud node.

7. The method for online monitoring of partial discharge of a switch cabinet according to claim 1 or 6, characterized in that: The fusion of the multi-level processing results includes: The time domain statistics, frequency domain energy spectrum and time-space depth features are weightedly fused, and the weights are dynamically adjusted according to the signal complexity. The fusion formula is: ,in, represents the final eigenvector after fusion, represents the time domain statistical characteristics, Frequency domain energy spectrum characteristics, Represents the spatiotemporal depth characteristics, is a dynamic weight coefficient and satisfies , Adaptive adjustment based on signal-to-noise ratio, Increases with the increase of pulse density.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method for online monitoring of partial discharge of a switch cabinet according to any one of claims 1 to 7 is implemented.

Images