A method, apparatus, device, and storage medium for identifying partial discharge types

Abstract

This invention relates to the field of discharge detection technology, and more particularly to a method, apparatus, device, and storage medium for identifying partial discharge types. The method involves optimizing multi-channel data to obtain an optimized binary signal; constructing a standardized dataset based on the optimized binary signal; optimizing a preset FP32 model based on a pre-defined short-time Fourier transform algorithm and the standardized dataset to obtain an optimized FP32 model; detecting surface acoustic emission signals and background noise of the device based on the optimized FP32 model to obtain detection results; analyzing the detected discharge event to determine the partial discharge type when the detection result indicates a discharge event; ensuring data integrity through multi-channel synchronous data acquisition; and combining the optimized FP32 model with short-time Fourier transform, resulting in a fast inference, accurate identification, and ability to analyze discharge types, adapting to complex power scenarios and meeting detection requirements.

Description

Technical Field

[0001] This invention relates to the field of discharge detection technology, and in particular to a method, apparatus, device, and storage medium for identifying partial discharge types. Background Technology

[0002] Partial discharge detection is a key method for evaluating the insulation of power equipment. Traditional detection methods each have their limitations: pulse current methods are susceptible to electromagnetic interference, ultrasonic methods have low signal-to-noise ratios, and ultra-high frequency methods are costly. Acoustic detection methods, while resistant to interference and low-cost, are difficult to meet engineering requirements due to overlapping discharge frequency bands and environmental noise interference. Machine learning methods rely on manual feature extraction and have poor generalization. Deep learning requires time-frequency conversion of acoustic signals, which can easily lead to the loss of temporal information. Furthermore, the models are complex, slow inference, and large in size, making them difficult to deploy on edge devices. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the present invention aims to provide a method, apparatus, device and storage medium for identifying partial discharge types.

[0004] The first aspect of this invention provides a method for identifying partial discharge types, comprising: acquiring multi-channel data according to a preset multi-channel parallel architecture; optimizing the multi-channel data to obtain an optimized binary signal; constructing a standardized dataset based on the optimized binary signal; optimizing a preset FP32 model based on a preset short-time Fourier transform algorithm and the standardized dataset to obtain an optimized FP32 model; acquiring surface acoustic emission signals of a device according to a preset sampling rate; acquiring background noise according to a preset data acquisition time; detecting the surface acoustic emission signals of the device and the background noise according to the optimized FP32 model to obtain a detection result; when the detection result indicates that a discharge event has been detected, analyzing the discharge event to obtain the partial discharge type; and generating a diagnostic report based on the discharge event and the partial discharge type.

[0005] Furthermore, the optimization processing of multi-channel data to obtain an optimized binary signal includes: preprocessing the multi-channel data to obtain the original binary signal; normalizing the original binary signal to obtain a normalized signal; updating a preset MobileNet network architecture according to a preset attention mechanism and a preset global average pooling mechanism to obtain an updated MobileNet network architecture; and optimizing the normalized signal based on the updated MobileNet network architecture to obtain an optimized binary signal.

[0006] Furthermore, the step of updating the preset MobileNet network architecture according to a preset attention mechanism and a preset global average pooling mechanism to obtain an updated MobileNet network architecture includes: obtaining the input layer structure from the MobileNet network architecture; reconstructing the input layer structure to obtain a reconstructed input layer; obtaining the stacking structure from the MobileNet network architecture; configuring the stacking structure according to the attention mechanism to obtain a configured stacking structure; generating channel weights according to the global average pooling mechanism and preset fully connected layers; and updating the MobileNet network architecture according to the reconstructed input layer, configured stacking structure, and channel weights to obtain the updated MobileNet network architecture.

[0007] Furthermore, the step of constructing a standardized dataset based on the optimized binary signal includes: analyzing the optimized binary signal according to a preset sliding window peak detection algorithm to obtain the dynamic range; performing a linear mapping on the optimized binary signal based on the dynamic range and a preset standardized interval to obtain a tip discharge type signal, a floating discharge type signal, a surface discharge type signal, and a non-discharge type signal; and constructing a standardized dataset based on a preset interference scenario, the tip discharge type signal, the floating discharge type signal, the surface discharge type signal, and the non-discharge type signal.

[0008] Furthermore, the optimization of the preset FP32 model based on the preset short-time Fourier transform algorithm and standardized dataset to obtain an optimized FP32 model includes: dividing the standardized dataset according to a preset partitioning ratio to obtain a training set and a validation set; performing transform processing on the optimized binary signal based on the short-time Fourier transform algorithm to obtain a time-spectrum graph; training the FP32 model according to the training set, the time-spectrum graph, a preset loss function, and preset class weights to obtain a primary FP32 model; validating the primary FP32 model according to the validation set to obtain a loss result; updating the preset learning rate decay factor according to the loss result to obtain an updated learning rate decay factor; and optimizing the primary FP32 model according to the updated learning rate decay factor to obtain the optimized FP32 model.

[0009] Furthermore, the step of detecting the surface acoustic emission signal of the device based on the optimized FP32 model to obtain the detection result includes: converting the optimized FP32 model according to a preset quantization format to obtain a compressed FP32 model; acquiring background noise according to a preset data acquisition time, and constructing an environmental noise baseline based on the background noise; extracting features from the surface acoustic emission signal of the device to obtain discharge features; and detecting the discharge features based on the compressed FP32 model, a preset fixed period, and the environmental noise baseline to obtain the detection result.

[0010] Furthermore, the step of detecting discharge characteristics based on the compressed FP32 model, a preset fixed period, and an environmental noise baseline to obtain detection results includes: comparing and verifying the discharge characteristics based on the environmental noise baseline to obtain amplitude verification results, frequency domain verification results, and trend verification results; analyzing the discharge characteristics based on the compressed FP32 model and the fixed period to obtain discharge type probability and signal confidence level; and detecting the amplitude verification results, frequency domain verification results, and trend verification results based on the discharge type probability and signal confidence level to obtain the detection results.

[0011] Furthermore, a partial discharge type identification device includes: a first data acquisition module for acquiring multi-channel data according to a preset multi-channel parallel architecture; an optimization processing module for optimizing the multi-channel data to obtain an optimized binary signal; a dataset construction module for constructing a standardized dataset based on the optimized binary signal; a model optimization module for optimizing a preset FP32 model based on a preset short-time Fourier transform algorithm and the standardized dataset to obtain an optimized FP32 model; a signal acquisition module for acquiring surface acoustic emission signals of the device according to a preset sampling rate; a background noise acquisition module for acquiring background noise according to a preset data acquisition time; a detection module for detecting the surface acoustic emission signals of the device and the background noise according to the optimized FP32 model to obtain a detection result; an analysis module for analyzing the discharge event when the detection result indicates that a discharge event has been detected to obtain the partial discharge type; and a diagnostic report generation module for generating a diagnostic report based on the discharge event and the partial discharge type.

[0012] Furthermore, a partial discharge type identification device includes: a memory and at least one processor, wherein the memory stores instructions; at least one processor invokes the instructions in the memory to cause the computer device to perform the steps of a partial discharge type identification method as described in any one of the above descriptions.

[0013] Furthermore, a computer-readable storage medium stores instructions that, when executed by a processor, implement the steps of a partial discharge type identification method as described in any of the preceding claims.

[0014] In the technical solution of this invention, a multi-channel parallel architecture synchronously receives multi-channel data, ensuring timing alignment and complete capture of signal features. In the model optimization stage, a standardized dataset is constructed by processing and optimizing binary signals, and time-frequency features are extracted by combining short-time Fourier transform, thereby improving the feature recognition capability of the optimized FP32 model. The optimized FP32 model has a fast inference speed and can quickly and accurately identify discharge events, reducing false positives and false negatives. After a discharge is detected, the type can be analyzed and a diagnostic report can be generated, forming a visualized retrospective diagnostic report. The overall solution takes into account data integrity, model accuracy, and practicality, adapts to complex power scenarios, and effectively meets the engineering requirements for efficient, accurate, and convenient partial discharge detection. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a first flowchart of a partial discharge type identification method provided by an embodiment of the present invention; Figure 2 This is a second flowchart of a partial discharge type identification method provided in an embodiment of the present invention; Figure 3 This is a third flowchart of a partial discharge type identification method provided in an embodiment of the present invention; Figure 4 This is a fourth flowchart of a partial discharge type identification method provided in an embodiment of the present invention; Figure 5 The fifth flowchart of a partial discharge type identification method provided in an embodiment of the present invention; Figure 6 The sixth flowchart of a partial discharge type identification method provided in an embodiment of the present invention; Figure 7 The seventh flowchart of a partial discharge type identification method provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the structure of a partial discharge type identification device provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of a partial discharge type identification device provided in an embodiment of the present invention. Detailed Implementation

[0016] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0017] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of a partial discharge type identification method according to the present invention includes: 101. Obtain multi-channel data according to the preset multi-channel parallel architecture; In this embodiment, the data input layer adopts a multi-channel parallel architecture, which can simultaneously receive raw data from different sensors such as the built-in ultrasonic sensor and the transient ground voltage (TEV) probe, and ensure the timing alignment of signals in each channel through a hardware synchronization mechanism. 102. Optimize the multi-channel data to obtain an optimized binary signal; 103. A standardized dataset is constructed based on the optimized binary signal; 104. Optimize the preset FP32 model based on the preset short-time Fourier transform algorithm and standardized dataset to obtain an optimized FP32 model; In this embodiment, the short-time Fourier transform can extract the time-frequency features of the signal, the standardized dataset ensures data quality and generalization, and the optimized FP32 model can more accurately identify discharge features, reduce interference, improve detection accuracy, and adapt to the partial discharge detection needs of power equipment. 105. Acquire the surface acoustic emission signal of the device according to the preset sampling rate; In this embodiment, during the signal acquisition stage, a wideband ultrasonic sensor array is used to capture the surface acoustic emission signal of the device in real time at a sampling rate of 200kHz. The sensor frequency response range covers the key frequency band of 20kHz-200kHz, ensuring complete recording of various discharge characteristics of the surface acoustic emission signal of the device. 106. Background noise is collected according to the preset data collection time; 107. Detect the surface acoustic emission signal and background noise of the equipment based on the optimized FP32 model to obtain the detection results; In this embodiment, the optimized FP32 model is used to detect the surface acoustic emission signal and background noise of the equipment. It can accurately distinguish between effective discharge signals and interference. The model balances accuracy and practicality, can efficiently identify the discharge state, reduce false detections and missed detections caused by background noise, provide reliable results for partial discharge detection of equipment, and adapt to the needs of complex power scenarios. 108. When the detection result indicates that a discharge event has been detected, the discharge event is analyzed to determine the type of partial discharge. 109. Generate diagnostic reports based on discharge events and partial discharge types; In this embodiment, the system output module provides multi-level alarm functions. When a discharge event is detected, in addition to local audible and visual alarms, key waveform features, discharge events, and partial discharge types are uploaded to the cloud monitoring platform via the 4G module. When operators confirm false alarms or missed alarms, relevant samples are automatically added to the incremental training dataset. The FP32 model is periodically fine-tuned to adapt to changes in equipment status. The operation and maintenance interface provides visualization analysis tools, which can trace back historical waveforms, feature maps, etc., forming a complete partial discharge monitoring solution. In this embodiment, a multi-channel parallel architecture synchronously receives data from multiple channels, ensuring timing alignment and complete capture of signal features. During the model optimization stage, a standardized dataset is constructed by processing and optimizing binary signals. Time-frequency features are extracted using short-time Fourier transform, improving the feature recognition capability of the optimized FP32 model. The optimized FP32 model has a fast inference speed and can quickly and accurately identify discharge events, reducing false positives and false negatives. After a discharge is detected, the type can be analyzed and a diagnostic report can be generated, forming a visualized retrospective diagnostic report. The overall solution takes into account data integrity, model accuracy, and practicality, adapts to complex power scenarios, and effectively meets the engineering requirements for efficient, accurate, and convenient partial discharge detection.

[0018] Please see Figure 2 A second embodiment of a partial discharge type identification method according to the present invention includes: 201. Preprocess the multi-channel data to obtain the original binary signal; In this embodiment, the original binary signal is directly input to preserve the time and frequency domain details of the partial discharge signal to the greatest extent. 202. Normalize the original binary signal to obtain a normalized signal; In this embodiment, the amplitude is normalized on the original binary signal in binary format; 203. Update the preset MobileNet network architecture according to the preset attention mechanism and the preset global average pooling mechanism to obtain the updated MobileNet network architecture. In this embodiment, the MobileNet network architecture is a lightweight convolutional neural network. Using this architecture reduces parameters and computational cost, balancing efficiency and accuracy, and is suitable for tasks such as feature optimization in mobile and embedded devices. The updated MobileNet backbone network contains nine depthwise separable convolutional layers, employing a progressive kernel size design. The first layer uses a large 7x1 kernel to extract macroscopic features, the next two layers use 5x1 kernels to capture detailed patterns, and subsequent layers use 3x1 convolutions for feature compression. Each convolutional layer is followed by batch normalization and a ReLU6 activation function, with an SE attention module inserted between every two convolutional layers. The input layer structure of the MobileNet backbone network has been reconstructed, enabling it to directly accept 1D raw waveform data, eliminating the time-frequency conversion step required in traditional methods. Regarding network depth, experiments determined the optimal layer configuration for the MobileNet network architecture: shallow layers use larger kernels (7x1) to capture broadband features, while deeper layers use smaller kernels (3x1) to extract fine structures. This design achieves the best balance between computational efficiency and feature extraction capability. 204. Based on the updated MobileNet network architecture, the normalized signal is optimized to obtain an optimized binary signal; In this embodiment, the normalized signal is optimized using the updated MobileNet network architecture to obtain an optimized binary signal, thereby improving the signal optimization accuracy and providing a high-quality signal for subsequent detection, thus meeting the requirements of discharge detection. In this embodiment, the multi-channel data is first preprocessed to obtain the original binary signal, which is then directly input to preserve the time-frequency details of the partial discharge signal to the greatest extent. Amplitude normalization is then applied to lay a unified data foundation for subsequent network optimization. Simultaneously, the MobileNet network architecture is updated with a lightweight design to reduce parameters and computational load, optimizing the data and adapting it to edge devices. The depthwise separable convolutional layer uses progressive convolution kernels, combined with batch normalization, ReLU6 activation function, and SE attention module. The input layer is also reconstructed to support 1D original waveform input, eliminating the time-frequency conversion step. Finally, this architecture is used to optimize the normalized signal to obtain an optimized binary signal, ensuring both computational efficiency and improved signal optimization accuracy, providing a high-quality signal for subsequent discharge detection and effectively meeting the efficiency and accuracy requirements of partial discharge detection in power equipment.

[0019] Please see Figure 3 A third embodiment of a partial discharge type identification method according to the present invention includes: 301. Obtain the input layer structure from the MobileNet network architecture; 302. Reconstruct the input layer structure to obtain the reconstructed input layer; In this embodiment, the input layer structure is reconstructed, such as adjusting the number of input channels from 3 to 2, and adapting the resolution of the spectrogram when the input size is matched (e.g., 224×224 pixels) to ensure that the network can fully receive and initially process the core features of the discharge signal, and avoid feature loss due to input mismatch. 303. Obtain the stacking structure from the MobileNet network architecture; 304. Configure the stacking structure according to the attention mechanism to obtain the configured stacking structure; In this embodiment, a stacked structure of depthwise separable convolutions is used in the main part of the MobileNet backbone network to extract multi-scale features. Through nine layers of convolution operations, different levels of features from micro pulses to macro discharge modes are captured step by step. In particular, an SE (Squeeze-and-Excitation) attention mechanism is inserted between every two convolutional layers of the stacked structure. This stacked structure can adaptively adjust the weights of the features of each channel, thereby improving the ability to identify key discharge features. 305. Generate channel weights based on the global average pooling mechanism and the preset fully connected layer; 306. Update the MobileNet network architecture based on the reconstructed input layer, configured stacking structure, and channel weights to obtain the updated MobileNet network architecture; In this embodiment, based on the global average pooling mechanism, the multi-channel feature map output by the configuration stack structure is compressed, and the spatial dimension of each channel is transformed into a 1×1 value to eliminate spatial redundancy. Then, the compressed value is mapped through a fully connected layer (such as 256-dimensional, 64-dimensional, and 4-dimensional, corresponding to four discharge states) to generate the weight coefficient of each channel. The weight coefficient of each channel is the channel weight, which can accurately reflect the contribution of different channels to the discharge type classification. Combined with dropout regularization (ratio 0.5), it prevents overfitting of the updated MobileNet network architecture. In this embodiment, the input layer of the reconstructed MobileNet network architecture has its number of channels adjusted to 2 to adapt to the spectrogram resolution and avoid feature loss. A stacked structure is configured to extract multi-scale features with 9 layers of depthwise separable convolutions, and an SE attention mechanism is inserted to enhance the recognition of key discharge features. Global average pooling is used to compress features, and fully connected layers generate channel weights. Dropout regularization is used to prevent overfitting of the updated network architecture. The updated network balances lightweight design with high accuracy, fewer parameters, and faster inference. It is adapted to edge devices, more accurate in recognizing the four types of discharge states, and has excellent anti-interference ability and generalization performance, effectively meeting the complex scenario requirements of partial discharge detection in power equipment.

[0020] Please see Figure 4A fourth embodiment of a partial discharge type identification method according to the present invention includes: 401. Analyze the optimized binary signal according to the preset sliding window peak detection algorithm to obtain the dynamic range; In this embodiment, the analysis logic of the sliding window peak detection algorithm is to slide the optimized binary signal into segments with a fixed window length (e.g., 100ms), detect the peak and valley values ​​of the signal within each window, calculate the difference between the peak and valley values, and finally statistically analyze the distribution range of all window differences to determine the dynamic range of the signal (i.e., the effective fluctuation range of the signal energy). The dynamic range can eliminate extreme outliers in the signal (e.g., instantaneous pulse interference), clarify the energy boundary of the effective signal, provide a benchmark range for subsequent signal classification mapping, and avoid classification deviations caused by outliers. 402. Based on the dynamic range and the preset standardized interval, the optimized binary signal is linearly mapped to obtain the tip discharge type signal, the floating discharge type signal, the surface discharge type signal, and the undischarged type signal. In this embodiment, the signal amplitude within the dynamic range is linearly mapped to a standardized interval (e.g., [0, 0.25) corresponds to no discharge, [0.25, 0.5) corresponds to surface discharge, [0.5, 0.75) corresponds to floating discharge, and [0.75, 1] ​​corresponds to tip discharge), and finally outputs four types of signals with clear labels (tip discharge type signal, floating discharge type signal, surface discharge type signal, and no discharge type signal). 403. A standardized dataset is constructed based on the preset interference scenarios, tip discharge type signals, floating discharge type signals, surface discharge type signals, and no-discharge type signals; In this embodiment, the data enhancement module simulates various interference scenarios in real time, including adding Gaussian white noise to simulate electromagnetic interference, superimposing sine waves to simulate mechanical vibration, and introducing amplitude attenuation to simulate propagation loss, and finally constructs a standardized dataset containing four types of discharge states (point discharge type signal, floating discharge type signal, surface discharge type signal and no discharge type signal). In this embodiment, the binary signal is first analyzed and optimized using a sliding window peak detection algorithm. After segmentation by a fixed window, the peak-to-valley difference is calculated to determine the dynamic range and eliminate extreme outliers, providing a precise benchmark for signal classification. Then, by combining the dynamic range with a standardized interval linear mapping, four types of discharge signal labels are identified, transforming unlabeled signals into structured data. Finally, the data augmentation module simulates real-world scenarios such as electromagnetic interference and mechanical vibration to enrich the types of interference in the dataset. The overall solution improves data validity and label accuracy, enhances the generalization ability of the dataset, adapts to the complex detection environment of power equipment, reduces on-site data collection costs, and provides high-quality, multi-scenario data support for subsequent model training, helping to improve the model's discharge type recognition accuracy and anti-interference capability.

[0021] Please see Figure 5 A fifth embodiment of a partial discharge type identification method according to the present invention includes: 501. Divide the standardized dataset according to the preset partitioning ratio to obtain the training set and the validation set; In this embodiment, the standardized dataset is first split into a training set and a validation set according to a ratio (such as 7:3 or 8:2). The training set is used for learning model parameters, and the validation set is used to evaluate the model's generalization ability and avoid overfitting. 502. Based on the short-time Fourier transform algorithm, the optimized binary signal is transformed to obtain the time-spectrum diagram; In this embodiment, the short-time Fourier transform algorithm optimizes the binary signal by using a sliding window segmentation, which not only preserves the frequency characteristics at each time point, but also presents the frequency change over time, adapting to the pulse-like and time-varying characteristics of the discharge signal; the time-spectrum diagram can simultaneously present the time and frequency characteristics of the signal, and can more intuitively reflect the dynamic change law of the partial discharge signal. Compared with single time domain or frequency domain features, it provides richer feature information for the model and improves the model's ability to distinguish discharge types. 503. Train the FP32 model based on the training set, time-spectrum graph, preset loss function, and preset class weights to obtain a primary FP32 model; In this embodiment, to address the sample imbalance problem, an improved FocalLoss function is used as the loss function, with class weights [1.0, 1.5, 1.5, 0.8] corresponding to the four discharge states. Model training employs an end-to-end optimization strategy. The FP32 model's network input layer is designed with a dual-channel structure. The training process uses a phased optimization method, initially employing transfer learning to initialize the network's bottom-layer weights with pre-trained audio classification model parameters. The backbone network is trained using the Nadam optimizer, with an initial learning rate set to 0.001. Cosine annealing scheduling is dynamically adjusted to address the sample imbalance problem. An improved FocalLoss loss function is used, with class weights [1.0, 1.5, 1.5, 0.8] corresponding to the four discharge states. The core of using the FocalLoss function is to solve the sample imbalance problem of the four discharge states. The loss penalty is increased for the discharge type with fewer samples (weight 1.5), avoiding the model being dominated by majority class samples (weight 0.8) during training. This allows the model to pay more attention to the difficult-to-distinguish and minority class samples, thereby improving the balance and accuracy of discharge state classification. 504. Validate the primary FP32 model using the validation set to obtain the loss results; 505. Update the preset learning rate decay factor based on the loss result to obtain the updated learning rate decay factor; 506. Optimize the primary FP32 model based on the updated learning rate decay factor to obtain an optimized FP32 model; In this embodiment, the validation set accuracy after each training round determines when the model is saved. When the validation loss does not decrease for 5 consecutive rounds, the learning rate decay (factor 0.1) is automatically triggered for updating. The primary FP32 model is optimized based on the updated learning rate decay factor to obtain the optimized FP32 model. In this embodiment, the dataset is first divided into reasonable proportions. The training set is used to learn parameters, and the validation set is used to evaluate generalization ability, avoiding overfitting. Short-time Fourier transform generates time-frequency spectrograms, providing rich time-frequency features to help the model distinguish discharge types. During training, the improved FocalLoss is combined with class weights to solve the sample imbalance problem. At the same time, end-to-end strategies, dual-channel input, transfer learning, and dynamic learning rate scheduling are used to balance training efficiency and accuracy. The model is optimized by updating the learning rate decay factor based on the validation loss. The final optimized FP32 model has balanced and accurate classification and strong generalization ability, providing a reliable model foundation for power equipment discharge detection and adapting to subsequent compression and edge deployment requirements.

[0022] Please see Figure 6 The sixth embodiment of a partial discharge type identification method in this invention includes: 601. Convert the optimized FP32 model according to the preset quantization format to obtain a compressed FP32 model; In this embodiment, the quantization format is INT8 quantization (INT8 quantization is a model optimization technique that compresses high-precision data such as 32-bit floating-point numbers (FP32) in a neural network model into 8-bit integers. Its core purpose is to reduce storage usage and improve inference speed while ensuring acceptable model accuracy, especially for edge devices). The quantized model size is reduced to 4.3MB, and optimized inference is achieved on an embedded platform in conjunction with the TensorRT acceleration engine. 602. An environmental noise baseline is constructed based on the background noise; In this embodiment, background noise is collected for 30 minutes under normal device conditions to establish an environmental noise baseline; 603. Extract features from the surface acoustic emission signal of the equipment to obtain discharge characteristics; In this embodiment, the discharge characteristics include amplitude characteristics, pulse characteristics, statistical characteristics, frequency peak characteristics, and frequency band characteristics; 604. The discharge characteristics are detected based on the compressed FP32 model, the preset fixed period, and the environmental noise baseline to obtain the detection results; In this embodiment, the compressed FP32 model balances accuracy and efficiency, enabling precise analysis of discharge characteristics; fixed-period detection assesses the stability of discharge signals, ensuring reliable model analysis results; and the environmental noise baseline filters out interference and reduces the impact of invalid signals. The combination of these three features not only improves the accuracy of discharge detection and reduces the false detection and missed detection rates, but also adapts to edge deployment, meets the real-time detection needs on site, and provides a clear basis for subsequent operation and maintenance, effectively adapting to complex partial discharge detection scenarios for power equipment. In this embodiment, the optimized FP32 model is converted into a compressed FP32 model using the INT8 quantization format, reducing the model size. This, combined with the TensorRT acceleration engine, enables optimized inference on embedded platforms, balancing accuracy and efficiency, and adapting to edge deployments. A baseline is constructed by collecting 30 minutes of background noise under normal device conditions, effectively filtering interference and reducing the impact of invalid signals. Multi-dimensional discharge characteristics are extracted, providing a comprehensive basis for detection. Simultaneously, fixed-period detection is used to evaluate signal stability, ensuring the reliability of model analysis results, meeting real-time on-site detection needs, and providing a clear basis for subsequent operation and maintenance, adapting to complex power scenarios.

[0023] Please see Figure 7 The seventh embodiment of a partial discharge type identification method in this invention includes: 701. Compare and verify the discharge characteristics based on the environmental noise baseline to obtain amplitude verification results, frequency domain verification results, and trend verification results; In this embodiment, the amplitude parameters (such as peak value and RMS value) of the discharge characteristics are extracted and compared with the amplitude range (including the average value and a threshold of 2-3 times the standard deviation) of the environmental noise baseline. If the discharge characteristic amplitude is higher than the baseline threshold (e.g., exceeding 3 times the standard deviation), the amplitude verification result is valid, and it is preliminarily determined that there may be a real discharge. If the discharge characteristic amplitude is lower than the baseline threshold, the amplitude verification result is determined to be invalid, and it is highly likely to be environmental noise. By analyzing the frequency domain distribution of the discharge characteristics (such as main frequency components and bandwidth) and comparing it with the frequency domain characteristics of the environmental noise baseline, for example, the partial discharge sound signal is mostly concentrated in the range of 2... The frequency range is 0-200kHz, while environmental noise (such as fan noise and electromagnetic noise) may be distributed below 50kHz. If the frequency domain of the discharge characteristics has low overlap with the baseline and matches the typical discharge frequency band, the frequency domain verification result is valid; otherwise, the frequency domain verification result is invalid. By analyzing the changing trend of discharge characteristics over time (such as pulse occurrence frequency and amplitude stability) and comparing it with the random fluctuation trend of the environmental noise baseline, the real discharge will show a periodic pulse or a gradually increasing amplitude pattern, while the environmental noise has no fixed trend. If the difference between the two is obvious, the trend verification result is valid; otherwise, the trend verification result is invalid. 702. Analyze the discharge characteristics based on the compressed FP32 model and fixed period to obtain the discharge type probability and signal confidence. In this embodiment, after receiving discharge characteristics, the compressed FP32 model (balancing accuracy and computational efficiency) outputs probability values ​​for various types of discharges (point discharge, hover discharge, surface discharge, etc.) through built-in classification logic (such as time-spectrum feature matching and historical data mapping); for example, it outputs "point discharge probability 88%, hover discharge probability 10%, no discharge probability 2%", clearly identifying the most likely discharge type; the signal confidence level is assessed by repeatedly detecting the signal at a fixed period to evaluate its stability: a fixed period is set (such as 10 seconds / time), and the analysis results of the compressed FP32 model are detected according to the fixed period. If the probability distribution of discharge types of multiple discharge characteristics is stable (such as the probability of point discharge ≥85% for 3 consecutive times), it indicates that the signal is not accidental interference, and the confidence level is ≥90%; if the probability fluctuates greatly (such as the probability of point discharge 88% in the first time and suddenly drops to 40% in the second time), the confidence level is ≤60%, reflecting low signal reliability. 703. The amplitude verification result, frequency domain verification result, and trend verification result are tested based on the discharge type probability and signal confidence level to obtain the test results; In this embodiment, the partial discharge detection process first prepares data by obtaining three-dimensional verification results (valid / invalid results of amplitude, frequency domain, and trend verification, and counting the number of valid items N) and model quantization results (maximum probability P_max of discharge type output by compressed FP32 model, and signal confidence C). Then, it determines four scenarios according to the priority of P_max, C, and N: high probability and high confidence (P_max≥70% and C≥80%), medium probability and medium confidence (60%≤P_max<70% and 60%≤C<80%), high probability and low confidence (P_max≥70% and C<60%), and low probability (P_max<60%). The result of detecting, not detecting, or suspected discharge event is obtained according to the different N values. In this embodiment, the solution constructs a partial discharge detection system through three-dimensional verification, model quantization, and multi-scenario judgment, possessing multi-dimensional core advantages. Three-dimensional verification of amplitude, frequency domain, and trend based on an environmental noise baseline can accurately filter environmental interference such as fans and electromagnetic interference, reducing invalid signal interference. The compressed FP32 model balances accuracy and efficiency, accurately outputting discharge type probabilities. Fixed-period detection can also assess confidence through signal stability, solving the problems of difficult identification and low reliability in traditional detection methods. Simultaneously, scenario-based judgment based on discharge type probability, signal confidence, and the number of valid verification items avoids single-dimensional misjudgments and flexibly responds to different detection situations, improving detection accuracy and shortening fault location time. The lightweight design adapts to edge deployment, and the confidence level and data feedback mechanism support operation and maintenance decisions, meeting the detection needs of complex power scenarios.

[0024] The above describes a partial discharge type identification method according to an embodiment of the present invention. The following describes a partial discharge type identification device according to an embodiment of the present invention. Please refer to [link / reference]. Figure 8 One embodiment of the partial discharge type identification device of the present invention includes: The first data acquisition module 1 is used to acquire multi-channel data according to a preset multi-channel parallel architecture; Optimization processing module 2 is used to optimize multi-channel data to obtain optimized binary signals; Dataset construction module 3 is used to construct a standardized dataset based on the optimized binary signal; Model optimization module 4 is used to optimize the preset FP32 model based on the preset short-time Fourier transform algorithm and standardized dataset to obtain an optimized FP32 model. Signal acquisition module 5 is used to acquire the surface acoustic emission signal of the device according to a preset sampling rate; Background noise acquisition module 6 is used to acquire background noise according to a preset data acquisition time. The detection module 7 is used to detect the surface acoustic emission signal and background noise of the equipment based on the optimized FP32 model to obtain the detection results; Analysis module 8 is used to analyze the discharge event when the detection result indicates that a discharge event has been detected, in order to obtain the type of partial discharge. Diagnostic report generation module 9 is used to generate diagnostic reports based on discharge events and partial discharge types. In this embodiment, a multi-channel parallel architecture synchronously receives data from multiple channels, ensuring timing alignment and complete capture of signal features. During the model optimization stage, a standardized dataset is constructed by processing and optimizing binary signals. Time-frequency features are extracted using short-time Fourier transform, improving the feature recognition capability of the optimized FP32 model. The optimized FP32 model has a fast inference speed and can quickly and accurately identify discharge events, reducing false positives and false negatives. After a discharge is detected, the type can be analyzed and a diagnostic report can be generated, forming a visualized retrospective diagnostic report. The overall solution takes into account data integrity, model accuracy, and practicality, adapts to complex power scenarios, and effectively meets the engineering requirements for efficient, accurate, and convenient partial discharge detection.

[0025] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0026] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0027] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying partial discharge types, characterized in that, include: Multi-channel data is obtained based on a pre-defined multi-channel parallel architecture; The multi-channel data is optimized to obtain an optimized binary signal; A standardized dataset is constructed based on optimized binary signals; The preset FP32 model is optimized based on the preset short-time Fourier transform algorithm and standardized dataset to obtain an optimized FP32 model. The surface acoustic emission signal of the device is acquired according to the preset sampling rate; Background noise was collected according to the preset data collection time. The surface acoustic emission signal and background noise of the equipment are detected based on the optimized FP32 model to obtain the detection results; If the detection result indicates that a discharge event has been detected, the discharge event is analyzed to determine the type of partial discharge. A diagnostic report is generated based on the discharge event and the type of partial discharge.

2. The partial discharge type identification method as described in claim 1, characterized in that, The optimization processing of multi-channel data to obtain an optimized binary signal includes: The multi-channel data is preprocessed to obtain the original binary signal; The original binary signal is normalized to obtain a normalized signal; The preset MobileNet network architecture is updated based on the preset attention mechanism and the preset global average pooling mechanism to obtain the updated MobileNet network architecture. The normalized signal is optimized based on the updated MobileNet network architecture to obtain an optimized binary signal.

3. The partial discharge type identification method as described in claim 2, characterized in that, The step of updating the preset MobileNet network architecture according to a preset attention mechanism and a preset global average pooling mechanism to obtain an updated MobileNet network architecture includes: The input layer structure is obtained from the MobileNet network architecture; The input layer structure is reconstructed to obtain the reconstructed input layer; The stacking structure is obtained from the MobileNet network architecture; The stacking structure is configured based on the attention mechanism to obtain the configured stacking structure; Channel weights are generated based on the global average pooling mechanism and the preset fully connected layer. The MobileNet network architecture is updated by reconstructing the input layer, configuring the stacking structure, and adjusting the channel weights to obtain the updated MobileNet network architecture.

4. The partial discharge type identification method as described in claim 1, characterized in that, The standardized dataset constructed based on the optimized binary signal includes: The optimized binary signal is analyzed according to the preset sliding window peak detection algorithm to obtain the dynamic range; The optimized binary signal is linearly mapped according to the dynamic range and the preset standardized interval to obtain the tip discharge type signal, the floating discharge type signal, the surface discharge type signal and the no-discharge type signal; A standardized dataset is constructed based on the preset interference scenarios, tip discharge type signals, floating discharge type signals, surface discharge type signals, and no discharge type signals.

5. The partial discharge type identification method as described in claim 1, characterized in that, The optimization of the preset FP32 model based on the preset short-time Fourier transform algorithm and standardized dataset to obtain an optimized FP32 model includes: The standardized dataset is divided according to a preset partitioning ratio to obtain a training set and a validation set; The optimized binary signal is transformed using the short-time Fourier transform algorithm to obtain the time-spectrum diagram; The FP32 model is trained based on the training set, time-spectrum graph, preset loss function, and preset class weights to obtain a primary FP32 model. The primary FP32 model was validated using the validation set to obtain the loss results; The preset learning rate decay factor is updated based on the loss result to obtain the updated learning rate decay factor. The primary FP32 model is optimized by updating the learning rate decay factor to obtain an optimized FP32 model.

6. The partial discharge type identification method as described in claim 1, characterized in that, The detection of surface acoustic emission signals and background noise of the device based on the optimized FP32 model to obtain detection results includes: The optimized FP32 model is converted according to the preset quantization format to obtain a compressed FP32 model; An environmental noise baseline is constructed based on background noise. Feature extraction is performed on the surface acoustic emission signal of the equipment to obtain discharge characteristics; The discharge characteristics were detected based on the compressed FP32 model, a preset fixed period, and an ambient noise baseline to obtain the detection results.

7. The partial discharge type identification method as described in claim 6, characterized in that, The detection of discharge characteristics based on the compressed FP32 model, a preset fixed period, and an environmental noise baseline to obtain detection results includes: The discharge characteristics are compared and verified based on the environmental noise baseline to obtain amplitude verification results, frequency domain verification results, and trend verification results. The discharge characteristics are analyzed based on the compressed FP32 model and a fixed period to obtain the discharge type probability and signal confidence. The amplitude verification result, frequency domain verification result, and trend verification result are tested based on the discharge type probability and signal confidence level to obtain the detection result.

8. A partial discharge type identification device, characterized in that, include: The first data acquisition module is used to acquire multi-channel data according to a preset multi-channel parallel architecture; The optimization processing module is used to optimize multi-channel data to obtain optimized binary signals; The dataset construction module is used to construct a standardized dataset based on the optimized binary signal. The model optimization module is used to optimize the preset FP32 model based on the preset short-time Fourier transform algorithm and the standardized dataset to obtain the optimized FP32 model. The signal acquisition module is used to acquire the surface acoustic emission signal of the device according to a preset sampling rate; The background noise acquisition module is used to acquire background noise according to a preset data acquisition time. The detection module is used to detect the surface acoustic emission signal and background noise of the equipment based on the optimized FP32 model to obtain the detection results; The analysis module is used to analyze the discharge event when the detection result indicates that a discharge event has been detected, in order to determine the type of partial discharge. The diagnostic report generation module is used to generate diagnostic reports based on discharge events and partial discharge types.

9. A partial discharge type identification device, characterized in that, The partial discharge type identification device includes: a memory and at least one processor, wherein the memory stores instructions; At least one of the processors invokes the instructions in the memory to cause the partial discharge type identification device to perform the steps of the partial discharge type identification method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the partial discharge type identification method as described in any one of claims 1-7.

Images