Lightweight YOLOv8n model partial discharge type detection and classification method based on PRPD spectrogram

Abstract

The invention discloses a light-weight YOLOv8n model partial discharge type detection and classification method based on a PRPD spectrogram, and aims to solve the problems of low partial discharge type detection precision, high model complexity, large calculation amount, poor real-time performance and the like in the prior art. According to the method, PRPD spectrograms of four typical partial discharge types of tip, air gap, suspension and surface are collected, normalization, gray processing, data enhancement and expansion and other preprocessing are carried out, a LabelImg platform is utilized to carry out labeling, then a YOLOv8n model is subjected to lightweight improvement, ShuffleNet-V2 is adopted as a trunk feature extraction network, a CoordAttention mechanism is introduced to enhance the feature extraction capability, and the feature extraction efficiency is improved. An EIOU loss function is applied to optimize target frame regression precision; the partial discharge type detection and classification precision is improved, the model calculation complexity and parameter quantity are reduced, real-time detection is realized, and the method has important practical application value.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment monitoring, and particularly to a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model using PRPD spectrograms. Background Art

[0002] In the field of power equipment operation and maintenance, partial discharge (PD) is an important monitoring indicator for the state of the power equipment insulation system, and plays a crucial role in ensuring the safe and stable operation of the power system. Partial discharge refers to an electrical discharge phenomenon caused by excessive electric field strength in a local area of the power equipment insulation system, and its occurrence often indicates the deterioration of the equipment insulation performance or the existence of potential defects. If not detected and processed in time, partial discharge may trigger major power accidents, resulting in equipment damage, large-scale power outages, and even threatening personal safety.

[0003] Currently, the methods for detecting and classifying partial discharge in power equipment are mainly divided into the following categories, and each has its own technical defects and deficiencies: 1. Methods based on traditional signal analysis: Core architecture and basic principle: Such methods mainly rely on signal feature extraction and rule classification. By detecting parameters such as the amplitude of discharge pulses, the phase distribution of discharge signals, and the discharge repetition rate, and combining expert experience or standards to identify and classify discharge types. Methods such as short-time Fourier transform, wavelet transform, and Hilbert-Huang transform are used to extract time-frequency features, and pattern recognition methods such as KNN, support vector machine (SVM), and decision tree are combined for classification.

[0004] Existing problems: These methods have high computational complexity, low sensitivity, are easily affected by environmental interference, and the input cost is relatively high.

[0005] 2. Methods based on machine learning: Core architecture and basic principle: Principal component analysis (PCA) is used to reduce the dimension of high-dimensional partial discharge signal data and extract the main features, and then support vector machine (SVM) is used for classification; or ensemble learning methods such as random forest (RF) and extreme gradient boosting (XGBoost) are used, and multiple decision trees are used for partial discharge pattern classification.

[0006] Existing problems: These methods are limited to small-scale, low-dimensional feature data, and manual feature extraction is required, which is time-consuming and laborious, and it is difficult to meet the requirements of real-time monitoring.

[0007] 3. Methods based on deep learning: Core architecture and basic principle: Utilize the Convolutional Neural Network (CNN) to convert one-dimensional partial discharge signals into two-dimensional time-frequency diagrams or PRPD (Phase Resolved Partial Discharge) spectrograms, and automatically extract local features in the images; or directly process the original time-series data using the Recurrent Neural Network (RNN) and its variants (such as LSTM and GRU) to capture dynamic changes and long-term dependence characteristics.

[0008] Existing problems: Although the deep learning-based methods have significantly improved the detection accuracy, the model structure is complex, the training time and computational volume have also increased substantially, the operating cost is high and the real-time performance is poor, which limits their popularization in practical applications.

[0009] In recent years, with the continuous development of computer vision and deep learning technologies, many scholars have begun to attempt to combine the image detection method based on the PRPD spectrogram of partial discharge with deep learning algorithms to improve the accuracy and efficiency of partial discharge type detection and classification. For example, denoising training and Bayesian optimization algorithms are introduced into the Deformable DETR model to optimize the partial discharge object detection model; multi-algorithm fusion based on ResNet-50 is used to select multiple sub-algorithms for defect type recognition; the MobileNetV1 model and transfer learning are combined to test and process the collected PRPD spectrograms. These methods have improved the detection accuracy to a certain extent, but there are still problems such as complex model structure, long training time, and large computational volume.

[0010] For example, CN115327304A proposes a method and system for identifying partial discharge types in transformers based on deep learning. This method uses the ResNet50V2 model as the basic model for transfer learning, and identifies partial discharge types by training PRPD color image samples. However, this method still needs to be improved in terms of model complexity, training efficiency, and real-time performance, especially in dealing with large-scale data sets and real-time monitoring requirements.

[0011] For example, CN119515816A discloses a method and system for detecting substation insulators and their defects with an improved YOLOv9. This method improves the YOLOv9 model for the substation insulator and its defect detection task, and improves the detection accuracy and efficiency by embedding a Diversity Branch Block (DBB) and using Haar wavelet downsampling and other technologies. However, this method mainly focuses on insulator defect detection and does not involve the identification of partial discharge types, and the applicability and generalization ability of the model in the field of partial discharge detection need to be verified.

[0012] In view of the problems and deficiencies of the above-mentioned existing technologies, the present invention proposes a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model of PRPD spectrograms. This method realizes the efficient and accurate detection and classification of partial discharge PRPD spectrograms by lightweight improvement of the YOLOv8n model and combination with advanced technologies such as the ShuffleNet-V2 lightweight network structure, the CoordAttention mechanism, and the EIOU loss function. At the same time, it reduces the complexity and computational amount of the model, improves the real-time performance, and provides more reliable technical support for the safe operation of the power system. Summary of the Invention

[0013] The technical problem to be solved by the present invention is to provide a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model of PRPD spectrograms, and solve the problems of low accuracy of partial discharge type detection and classification, high model complexity, large computational amount, and poor real-time performance in the technical field of power equipment monitoring.

[0014] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model of PRPD spectrograms, including the following steps: Step1: Collect PRPD spectrogram data of partial discharge types; Step2: Preprocess the PRPD spectrograms; Step3: Divide the data set into a training set, a validation set, and a test set; Step4: Lightweight improvement of the YOLOv8n model; Step5: Use the annotation platform to perform target annotation on the PRPD spectrograms in the training set, and input the annotated training set into the improved YOLOv8n model for training; Step6: Use the validation set data to evaluate the model performance and adjust the hyperparameters; Step7: Use the test set to evaluate the performance of the improved model; Step8: Deploy the trained improved YOLOv8n model to the power system site for real-time detection and identification of partial discharge types in power equipment.

[0015] In a preferred solution, the partial discharge types in Step1 include tip discharge, floating discharge, air gap discharge, and surface discharge.

[0016] In a preferred solution, the preprocessing in Step2 includes normalization, grayscale processing, and data augmentation. Among them, the normalization process limits the discharge signal amplitude of the PRPD spectrogram between 0 and 1 through linear transformation.

[0017] In a preferred solution, the data augmentation step includes random translation, scaling, resampling, and rotation operations to increase the diversity of PRPD spectrogram data.

[0018] In a preferred solution, the division ratio of the training set, validation set, and test set in Step 3 is 8:1:1.

[0019] In a preferred solution, the lightweight improvement of the YOLOv8n model in Step 4 is to use the ShuffleNet-V2 network structure to replace the original backbone feature extraction network of YOLOv8n, and add the CoordAttention mechanism and EIOU loss function to the network.

[0020] In a preferred solution, the lightweight improvement in Step 4 includes using ShuffleNet-V2 to replace the original backbone feature extraction network of YOLOv8n and optimizing other parts of the network to reduce the amount of computation and the number of parameters; the ShuffleNet-V2 network structure significantly reduces the amount of computation through the combination of grouped channel convolution and point convolution, and improves the network's ability to express features through the channel shuffle mechanism.

[0021] In a preferred solution, the CoordAttention mechanism is introduced into the Neck of the lightweight YOLOv8n model after improvement in Step 4. The CoordAttention mechanism is used to enhance the ability to focus on important features in the improved YOLOv8n model, improve the feature extraction ability and classification performance in complex backgrounds. This mechanism performs one-dimensional global average pooling on the input feature map in the horizontal and vertical directions respectively to obtain position information, generates attention weights for height and width, and finally performs weighted correction on the input feature map.

[0022] In a preferred solution, the EIOU loss function is also introduced into the model after the lightweight improvement of YOLOv8n in Step 4. The EIOU loss function is used to accelerate the training convergence speed of the network and optimize the accuracy of target box regression in the improved YOLOv8n model. This function introduces a center distance term and width-height penalty terms on the basis of the traditional IoU loss to more comprehensively measure the difference between the predicted box and the ground truth box.

[0023] In a preferred solution, the target annotation in Step 5 uses the LabelImg annotation platform to generate an annotation file containing the coordinates of the discharge feature bounding box and the corresponding discharge type label.

[0024] In a preferred solution, the training in Step5 is to iteratively train the improved YOLOv8n model using optimization algorithms such as gradient descent, continuously adjust the network weights through the optimization algorithm in the model until the loss function value of the model on the training data converges below a predetermined threshold, and the optimization algorithm is the gradient descent algorithm or its variant.

[0025] In a preferred solution, the performance evaluation in Step7 is to comprehensively evaluate the performance of the improved YOLOv8n model in the partial discharge type detection and classification task by calculating performance indicators such as precision, mean average precision, computational complexity, and the number of model parameters.

[0026] The method for detecting and classifying partial discharge types based on the PRPD spectrogram of the lightweight YOLOv8n model provided by the present invention has the following beneficial effects: 1. Through targeted data preprocessing, lightweight network improvement (such as using ShuffleNet-V2 as the backbone feature extraction network), attention mechanism enhancement (introducing the CoordAttention module), and optimized loss function (EIOU loss function), the present invention realizes high-precision detection and classification of various partial discharge types such as tip, air gap, suspension, and surface discharge in complex environments. Experimental results show that the present invention shows obvious advantages in detection accuracy compared with the prior art.

[0027] 2. The present invention adopts the ShuffleNet-V2 network structure, significantly reducing the computational complexity and the number of parameters of the model, improving the operation efficiency, making the method suitable for resource-constrained embedded devices and edge computing environments, and meeting the requirements of real-time monitoring and fault prevention of power equipment.

[0028] 3. Due to the significant reduction in the computational complexity of the model, the method of the present invention can achieve real-time detection on edge computing devices, meeting the real-time monitoring requirements of the power system, and greatly improving the efficiency and reliability of equipment maintenance.

[0029] 4. By introducing the CoordAttention mechanism and the EIOU loss function, the present invention significantly improves the feature extraction ability and target localization accuracy of the model in the presence of noise interference and complex backgrounds; at the same time, data augmentation processing (such as translation, scaling, resampling, etc.) enhances the adaptability of the model to different discharge types and spectrogram changes, ensuring the stable operation of the system in power equipment monitoring.

[0030] 5. On the premise of ensuring high detection accuracy, the overall system of the present invention significantly reduces the operation cost and response time; this provides reliable technical support for the real-time monitoring and fault prevention of power equipment, further ensuring the safe and stable operation of the power system.

[0031] 6. The method of the present invention not only solves the problems existing in the prior art (such as low detection accuracy, high model complexity, large computational amount, poor real-time performance, etc.), but also provides more reliable technical support for the safe operation of power equipment.

[0032] 7. Through detailed data preprocessing (such as normalization, grayscale processing) and data augmentation steps (such as translation, scaling, resampling, etc.), the present invention further improves the generalization ability and robustness of the model. This provides a strong guarantee for the stable operation of the model in different environments and conditions.

[0033] 8. The present invention shows remarkable beneficial effects in the field of partial discharge detection. It not only improves the detection accuracy and real-time performance, but also reduces the computational complexity and operation cost of the model, providing an important technical guarantee for the safe operation of power equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The following further describes the present invention with reference to the drawings and embodiments: Figure 1 is the technical flow chart of Embodiment 1 of the present invention; Figure 2 is the diagram of simulating four typical partial discharge devices in Embodiments 1 and 3 of the present invention; Figure 3 is the experimental circuit wiring diagram of Embodiments 1 and 3 of the present invention; Figure 4 is part of the PRPD spectrograms collected in Embodiments 1 and 3 of the present invention; Figure 5 is part of the preprocessed PRPD spectrograms in Embodiments 1 and 3 of the present invention; Figure 6 is the flow chart of the improved YOLOv8n algorithm in Embodiment 3 of the present invention; Figure 7 is the network structure diagram of ShuffleNet-V2 (stride = 1) in Embodiment 3 of the present invention; Figure 8 is the structure diagram of CoordAttention in Embodiment 3 of the present invention; Figure 9 is the model detection result diagram in Embodiment 3 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0035] The following further describes the technical solutions in the present invention with reference to the drawings and embodiments: Embodiment 1 As Figures 1 to 6 shown, this embodiment details the specific implementation process of the method for detecting and classifying partial discharge types based on the PRPD spectrogram of the lightweight YOLOv8n model of the present invention.

[0036] 1. Data collection of the PRPD spectrogram of partial discharge Four typical partial discharge type experiments, namely tip discharge, floating discharge, air gap discharge and surface discharge, were simulated in the high-voltage hall, as Figure 2 and Figure 3 shown. The PRPD (Phase Resolved Partial Discharge) spectrograms of these discharge types were collected using high-precision sensors and acquisition equipment, as Figure 4 shown. A sufficient number of samples were collected for each discharge type to ensure the diversity and representativeness of the data.

[0037] 2. Preprocessing such as data normalization, grayscale processing, and data augmentation 2.1 Normalization processing Due to the different magnitudes of discharge amplitudes, there will be some differences in the spectrograms of the same discharge model, which may cause errors in subsequent discharge type recognition. Therefore, the collected PRPD spectrograms were normalized. The specific method of normalization is to limit the amplitudes of all discharge signals between 0 and 1 through linear transformation. In this way, the spectrograms of different discharge types have a unified standard in terms of amplitude, facilitating subsequent processing and analysis.

[0038] 2.2 Grayscale processing To avoid the influence of color on recognition and classification, the normalized PRPD spectrograms were subjected to grayscale processing. Grayscale processing converts each pixel point in the image into a grayscale value, which reflects the brightness information of the pixel point. Through grayscale processing, a PRPD spectrogram containing only brightness information was obtained, providing a more concise data basis for subsequent feature extraction and classification.

[0039] 2.3 Data augmentation To improve the generalization ability of the model, data augmentation operations were performed on the preprocessed PRPD spectrograms, as Figure 5 shown. The specific methods include translation, scaling, rotation, and resampling, etc. Through these operations, more training samples were generated, increasing the diversity of the data.

[0040] 3. Division of the training set, validation set, and test set The enhanced dataset was divided into a training set, a validation set, and a test set, usually in the ratio of 80%, 10%, and 10%. In this way, the training set can be used to train the model, the validation set can be used to evaluate the performance of the model, and the test set can be used to finally verify the performance of the model.

[0041] 4. Improvement of the YOLOv8n model 4.1 ShuffleNet-V2 lightweight module Although the traditional YOLOv8 model has good detection performance, its backbone network is relatively complex, with a large amount of computation, and is not suitable for running on resource-constrained edge devices. Therefore, the YOLOv8n model is improved for lightweight. First, the ShuffleNet-V2 structure is used to replace the backbone feature extraction network of the original model. ShuffleNet-V2 significantly reduces the amount of computation through a combination of grouped channel convolution and point convolution, as well as a channel shuffle mechanism. Specifically, ShuffleNet-V2 evenly divides the input feature map into two parts along the channel dimension. One part is directly passed as the shortcut branch, and the other part enters the main branch for a series of transformations. The processing flow of the main branch includes 1×1 convolution (point convolution), 3×3 depth convolution, and another 1×1 convolution (channel recombination). After completing the transformation of the main branch, the shortcut branch and the output of the main branch are concatenated along the channel dimension, and the channel shuffle operation is used to promote the full interaction of information between channels.

[0042] 4.2, CoordAttention Mechanism To further improve the feature extraction ability and classification performance of the model in complex backgrounds, the CoordAttention mechanism is introduced in the Neck part of the improved YOLOv8n model. The CoordAttention mechanism is a self-attention mechanism that can capture global channel dependencies while embedding long-range spatial location information. Specifically, the CoordAttention mechanism performs one-dimensional global average pooling on the input feature map in the horizontal and vertical directions respectively to obtain two sub-features describing location information. Then, through shared 1×1 convolution and branch processing, attention weights for height and width are generated. Finally, the generated attention weights are applied to the corresponding spatial dimensions of the original input respectively to achieve a more refined feature representation.

[0043] 4.3, EIOU Loss Function In object detection tasks, the choice of loss function has a crucial impact on the performance of the model. The traditional IoU (Intersection over Union) loss function mainly focuses on the overlapping area between the predicted bounding box and the ground truth box, but in some cases, it may not accurately reflect the difference between the predicted bounding box and the ground truth box. Therefore, the EIOU (Efficient Intersection over Union) loss function is introduced in the improved YOLOv8n model. The EIOU loss function adds penalty terms for the center distance and size difference on the basis of the traditional IoU loss, thus providing a more comprehensive and finer-grained way to measure bounding box regression. This not only speeds up the training convergence speed of the network but also optimizes the accuracy of target box regression.

[0044] 5. Pre-training the model with training set data 5.1. Image annotation To train the improved YOLOv8n model, target annotation is performed on the PRPD spectrograms in the training set. The LabelImg annotation platform is used to manually annotate the discharge features in the spectrograms. During the annotation process, the bounding box coordinates of each discharge feature and its corresponding discharge type label are recorded. These annotation information will serve as the supervision signal during model training to help the model learn how to identify different types of partial discharges.

[0045] 5.2. Model pre-training After completing the model improvement, the annotated training set is input into the improved YOLOv8n model for training.

[0046] 6. Adjusting model parameters with validation dataset data During the training process, optimization algorithms such as gradient descent are used to continuously adjust the network weights so that the model can capture the feature and target information in the images. To prevent overfitting, techniques such as data augmentation and Dropout are also adopted. After multiple iterative trainings, the loss function value of the model on the training data gradually converges below a predetermined threshold, indicating that the model has learned effective feature representations and classification rules.

[0047] 7. Evaluating model performance with test set data During the training process, the performance of the model is regularly evaluated using the validation set. Specific methods include calculating metrics such as Precision and Mean Average Precision (mAP). These metrics can quantify the accuracy of the model when predicting positive samples and the overall detection accuracy across all classes. Based on the feedback from the validation set, the hyperparameters of the model (such as learning rate, regularization factor, batch size, etc.) can be appropriately adjusted to further optimize the model's performance. After the training is completed, the test set is input into the final improved model for evaluation, and various performance metrics are calculated. These performance metrics will serve as an important basis for judging whether the model meets the actual application requirements.

[0048] 8. Model deployment and real-time detection Finally, we deploy the trained improved YOLOv8n model to the power system site for real-time detection and identification of partial discharges and their types. The system can collect PRPD spectrogram data of power equipment in real time and input it into the improved YOLOv8n model for classification and identification. Since the model adopts a lightweight design, it can operate efficiently on resource-constrained edge devices to meet the requirements of real-time detection. This system provides reliable technical support for the fault prevention and maintenance of power equipment, effectively improving the safety and stability of the power system.

[0049] Through the above specific implementation methods, the specific implementation process of a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model of PRPD spectrograms is elaborated in detail. This process covers multiple links such as data collection and preprocessing, model lightweight improvement, model training and evaluation, and model deployment and real-time detection, fully demonstrating the innovation and practicality of the present invention.

[0050] Embodiment 2 In another preferred embodiment, on the basis of the above Embodiment 1, this embodiment provides another specific implementation of a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model of PRPD spectrograms. The difference from Embodiment 1 lies in the specific methods of lightweight improvement and data augmentation of the YOLOv8n model.

[0051] In the data preprocessing stage, experiments of simulating four typical discharge types are also carried out in the high-voltage hall, PRPD spectrograms are collected, and normalization and grayscale processing are performed. Then, more diverse data augmentation methods are adopted, such as random cropping, flipping, adding noise, etc., to further enrich the training samples and improve the generalization ability of the model.

[0052] In terms of model lightweight improvement, in addition to using ShuffleNet-V2 as the backbone feature extraction network, more lightweight technologies are introduced. For example, depthwise separable convolution is added after some convolutional layers in the model to further reduce the amount of calculation and the number of parameters. Depthwise separable convolution decomposes the standard convolution into two steps: depthwise convolution and pointwise convolution. Among them, depthwise convolution is performed independently on each input channel, and pointwise convolution is used to combine the outputs of depthwise convolution.

[0053] In addition, in the Neck part, besides introducing the CoordAttention mechanism, an Attention Gating Mechanism is added to better capture the dependencies between features and improve the model's feature extraction ability. The Attention Gating Mechanism generates attention weights by calculating the correlation between feature maps and corrects the input features by weighting, thereby enhancing the model's ability to focus on important features.

[0054] In terms of the loss function, this embodiment also adopts the EIOU loss function to optimize the accuracy of the target box regression. By introducing penalty terms for the center distance and size difference, the EIOU loss function provides a more comprehensive and fine-grained bounding box regression metric, further improving the detection performance of the model.

[0055] The model training, validation, adjustment, and evaluation processes are similar to those in Embodiment 1. The training set is input into the improved YOLOv8n model for training, the validation set is used to evaluate the performance of the model, and the hyperparameters of the model are adjusted according to the feedback of the validation set. After training, the test set is input into the finally improved model for evaluation, and various performance metrics are calculated.

[0056] Finally, the trained improved YOLOv8n model is deployed to the power system site for real-time detection and identification of partial discharges and their types. This system can accurately classify different discharge types while ensuring light weight, providing reliable technical support for the fault prevention and maintenance of power equipment, and effectively improving the safety and stability of the power system.

[0057] Embodiment 3 In another preferred embodiment, based on the above Embodiments 1 and 2, as Figure 6 shown, this embodiment further elaborates and refines the method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram of the present invention.

[0058] 1. Data collection and preprocessing 1.1 Data collection Experiments simulating four typical discharge types, namely tip discharge, floating discharge, air gap discharge, and surface discharge, are conducted in the high-voltage hall, as Figure 2 and Figure 3 shown, to obtain the PRPD spectrograms of the four discharge types, as Figure 4 shown. To reduce the differences between the same discharge types and highlight the discharge characteristics, the PRPD spectrograms are normalized; to avoid the influence of color on recognition and classification, the images are grayscale processed, and the processed PRPD spectrograms are as Figure 5 shown.

[0059] 1.2. Image annotation The LabelImg tool is used to manually annotate the discharge features in the spectrogram to generate annotation files in YOLO format. These files record the bounding box coordinates of each discharge feature and its corresponding discharge type label. The bounding box coordinates of the discharge features in the image are .

[0060] 1.3. Data augmentation and division To improve the generalization ability of the model, data augmentation and expansion processing are performed on the collected spectrograms, including operations such as translation, scaling, and resampling. The dataset is divided into a training set, a validation set, and a test set at a ratio of 80%, 10%, and 10% to ensure the effectiveness of model training and evaluation.

[0061] 2. Lightweight improvement of the YOLOv8n model After completing the data preprocessing, lightweight improvement is carried out on the YOLOv8n model. The YOLOv8 model is an efficient object detection network that can perform feature detection on the input PRPD spectrogram and identify its corresponding class label. However, the model has a high complexity and a large amount of computation. Therefore, the present invention performs lightweight improvement on the YOLOv8n model to meet the classification requirements of PRPD spectrograms of different partial discharge types. First, the ShuffleNet-V2 structure is used to replace the original model as the backbone feature extraction network, which significantly reduces the computational complexity while improving the model's detection and classification accuracy. Secondly, the CoordAttention mechanism is introduced in the Neck to enhance the model's attention ability to important features in the spatial and channel dimensions and the quality of feature fusion, so as to extract effective information in complex backgrounds. Finally, the bounding box regression loss function EIOU is introduced to accelerate network convergence, thereby more efficiently completing training.

[0062] 2.1. ShuffleNet-V2 lightweight module In the traditional YOLOv8n backbone network, multiple convolutional feature extraction modules are used. Its advantage is that it can capture multi-scale features through deep stacking and can maintain high detection performance in relatively complex tasks. However, the number of parameters in the backbone network of the traditional YOLOv8n reaches 3.2M, and the amount of computation (FLOPs, Floating Point Operations Per Second) is about 8.7 GFLOPs, which will consume a large amount of computing resources. The large number of parameters and the amount of computation pose great pressure on the deployment of edge devices and greatly increase the difficulty of real-time detection. Therefore, the present invention selects the lightweight backbone network ShuffleNet-V2 (stride = 1), as Figure 7 shown, which greatly improves the lightweight of the model and is better deployed in the device.

[0063] The core improvement of ShuffleNet-V2 (stride = 1) lies in the introduction of an efficient grouped convolution and channel shuffle mechanism. Grouped convolution divides the input channels into multiple groups, and each group independently performs convolution operations. Channel shuffle compensates for the lack of feature fusion caused by grouped convolution by rearranging the convolutional output channels, enabling the features of different groups to be mixed and interacted. By combining the two, ShuffleNet-V2 becomes an efficient lightweight network structure, significantly improving the network's feature expression and fusion capabilities while significantly reducing the network's computational amount and parameter quantity.

[0064] First, given an input feature map (where is the height, is the width, is the number of channels), the module first evenly divides along the channel dimension into two parts, namely (the number of channels in each part is ); where is directly passed as the shortcut branch, while enters the main branch for a series of transformations. The processing flow of the main branch is as follows: First, perform a 1×1 convolution (point convolution) on , and the formula is expressed as: (1) In the formula, is the output of the point convolution of the main branch, is the 1×1 convolution kernel parameter, represents the convolution operation, represents batch normalization; then, a 3×3 depthwise convolution is adopted, and each channel independently performs convolution while keeping the number of channels unchanged. The processing formula is: (2) In the formula, is the independent convolution output of the channel, represents the depthwise convolution operation; subsequently, another 1×1 convolution is performed to reorganize the features and pass through and ReLU activation, and its formula is: (3) In the formula, is the output of the main branch, is the parameter of the second 1×1 convolution kernel; after completing the transformation of the main branch, the shortcut branch and the main branch output are concatenated in the channel dimension and denoted as : (4) Restore the number of output channels to .

[0065] To further promote the full interaction of information between channels, ShuffleNet-V2 introduces the Channel Shuffle operation. The specific process is to first reshape to the size (where is the number of groups), then transpose the middle two dimensions to get the shape , and finally reshape it back to , thus obtaining the output feature map .

[0066] This ShuffleNet-V2 module divides the input feature Figure 1 into two, performs lightweight convolution operations (including 1×1 point convolution and 3×3 depth convolution) on half of the features, and uses the channel shuffle operation to break the channel independence, thereby significantly reducing the computational amount while ensuring full information interaction and efficient feature expression. This design enables the network to have excellent running efficiency on embedded and mobile devices, while ensuring a relatively high recognition accuracy to a certain extent.

[0067] 2.2. CoordAttention mechanism For the PRPD spectrograms of partial discharges, there are some similarities in the features of different types of spectrograms. Therefore, there are some difficulties in the feature extraction of discharge types. The CoordAttention mechanism is a self-attention mechanism. Different from traditional channel attention mechanisms [such as the SE module (Secure Element, security module), CBAM (Convolutional Block Attention Module)], the CoordAttention mechanism aims to capture global channel dependencies while embedding long-range spatial location information, so as to express features more finely. Its basic idea is to perform one-dimensional global average pooling on the input feature map in the horizontal and vertical directions respectively to obtain two sub-features describing location information, and then through shared transformation and branch processing, generate attention weights for height and width, and finally perform weighted correction on the input respectively, as Figure 8 shown, When the feature map is input into the CoordAttention mechanism, first, the input in two directions, namely , One-dimensional global average pooling is performed separately on the above to obtain two low-dimensional features in the vertical and horizontal directions, namely the vertical direction feature , horizontal direction feature Indicated as: Vertical direction: (4) The obtained Contains the global information of each channel on each row; Horizontal direction: (5) The obtained Captures the global information of each channel on each column; Subsequently, the above two features are concatenated in the spatial dimension. For ease of subsequent processing, usually first and Perform dimensional reshaping (for example, change to , to ), and then concatenate along the spatial dimension to obtain: (7) Next, use a shared 1×1 convolution (essentially a fully connected transformation on the channels) to Reduce the dimension (usually set the reduction ratio to , let the number of intermediate channels be , and connect batch normalization and activation functions (such as ReLU or h-swish), the formula is: (8) In the formula, Is the shared convolution kernel parameter, Is the Sigmoid activation function, which limits the output value to the range [0, 1]; Furthermore, the shared transformed feature Is divided into two parts according to the original spatial dimension length: (1)The first 𝐻 elements correspond to the vertical direction information, denoted as: ; (2)The last 𝑊 elements correspond to the horizontal direction information, denoted as: . Then, perform 1×1 convolution (not sharing parameters) on these two parts respectively and connect the Sigmoid activation function to generate attention weights: (9) (10) In the formula, and Convolution weights in the vertical and horizontal directions respectively; Finally, apply the generated attention weights to the original input on the corresponding spatial dimensions. Specifically, for each position and channel , the output feature is calculated as follows: (11) Thus, the input features obtain dynamic channel recalibration with position information in both the vertical and horizontal directions, achieving more refined feature representation. This design not only introduces global long-range dependencies but also preserves spatial position information, making it very suitable for use in lightweight networks and mobile devices, while effectively enhancing the model's ability to capture detailed features in complex scenarios.

[0068] 2.3, EIoU Loss Function For the detection and classification of PRPD spectrograms of partial discharge types, the morphology of the spectrograms is complex or the sizes are diverse. The traditional IoU (Intersection over Union) loss function mainly focuses on the overlapping area between the predicted box and the ground truth box, but lacks constraints on the center point offset and shape matching of the target box, and the bounding box will also be inaccurately located. Introducing the EIoU loss function can not only significantly improve the regression accuracy of the bounding box but also enhance the detection performance of the target, accelerate the model training convergence speed, and at the same time meet the goal of lightweight design. The calculation process of the entire EIoU loss function can be described as the following steps: First, calculate the intersection over union (IoU) between the predicted box and the ground truth box , and the formula is: (12) In the formula, is the intersection area, is the union area. The traditional IoU loss is usually defined as , but using the IoU loss alone has the problem of discontinuous gradients. Therefore, in order to make the center positions of the predicted box and the ground truth box closer, EIoU introduces a center distance term. Let the center of the predicted box be , and the center of the ground truth box be , and the Euclidean distance between them is: (13) At the same time, calculate the diagonal length 𝑐 of the minimum bounding rectangle containing the predicted box and the ground truth box; the center distance penalty term is ; To further make the shape of the predicted bounding box more consistent with the ground truth box, EIoU penalizes the differences in width and height respectively. Let the width and height of the predicted bounding box be and , and the ground truth box be and ; meanwhile, the width and height of the minimum bounding rectangle are denoted as and . The penalty terms for width and height are: and ; Combining the above parts, the final form of the EIoU loss is: (14) Generally speaking, the EIoU loss function provides a more comprehensive and fine-grained way to measure bounding box regression by adding two important components, the center distance and the size difference, to the traditional IoU loss. It not only solves the problem of insufficient gradient feedback in the IoU loss but also further improves the performance of the detection model in terms of localization accuracy.

[0069] 2.4. Model Detection Based on the Dataset After completing the improvement of the YOLOv8n model, the divided training set, validation set, and test set are input into the model according to the following process: (1) Training set: The training set is input into the improved YOLOv8n model for training. Through optimization algorithms such as gradient descent in the model, the network weights are continuously adjusted so that the model can capture the features and target information in the images. In this stage, the model iteratively learns until it reaches a good convergence state on the training data; (2) Validation set: During the training process, the performance of the model is regularly evaluated using the validation set (such as calculating metrics like precision, mean average precision, etc.). Based on the feedback from the validation set, the hyperparameters of the model (such as learning rate, regularization factor, batch size, etc.) can be appropriately adjusted; (3) Test set: After the model training and hyperparameter tuning are all completed, the test set is input into the final improved model for evaluation, and various performance metrics are calculated.

[0070] The above three types of datasets with clear divisions can comprehensively and objectively measure the improvement effect of the model and provide a reliable guarantee for practical applications.

[0071] 2.5. Evaluation of the Improved Model During the model training and validation stages, multiple evaluation metrics are used to measure the performance of the integrated model, including: (1) Precision: Measures the proportion of correct predictions when the model predicts a certain class, and can quantify the accuracy of the model when predicting positive samples; (2) Mean Average Precision (mAP): Reflects the overall detection accuracy of the model across all classes; (3) FLOPs (Floating Point Operations): Refers to the number of floating point operations the model needs to perform when processing an image, and is an important indicator for measuring the computational complexity and running speed of the model; (4) Model Parameters: The total number of all trainable parameters that make up the neural network, reflecting the capacity and complexity of the model.

[0072] By calculating these metrics on the validation set and test set, it is ensured that the improved model can maintain high classification performance when detecting different types of partial discharges, as Figure 9 shown.

[0073] 3. Model Deployment and Real-time Detection Finally, the improved YOLOv8n model is deployed in the power system for real-time detection of partial discharge types, and the operating conditions of power equipment are monitored and analyzed to ensure the efficient and stable operation of the system in diverse environments.

[0074] Example 4 In another preferred embodiment, based on the above Embodiment 3, as shown in this embodiment, the implementation process of a method for detecting and classifying partial discharge types of a lightweight YOLOv8n model based on PRPD spectrograms of the present invention is further refined, specifically including the following steps: Step 1: Data Collection Scenario Setup: In the high-voltage hall, four typical partial discharge types are simulated using laboratory equipment, namely corona discharge, floating discharge, surface discharge, and air-gap discharge.

[0075] Data Acquisition: Use a high-speed oscilloscope or a dedicated sensor to collect the PRPD spectrograms of each discharge type. 1000 spectrograms are collected for each discharge type, for a total of 4000 pieces of raw data.

[0076] Step 2: Data Preprocessing Normalization Processing: Normalize the collected PRPD spectrograms to unify the pixel value range to [0, 1] to eliminate the influence of different acquisition devices or environmental factors.

[0077] Grayscale Processing: Convert the color PRPD spectrograms to grayscale images to reduce computational complexity while retaining the key information of discharge characteristics.

[0078] Step 3: Data Augmentation and Expansion Processing Translation operation: Randomly translate the spectrogram within a range of ±10% of the original image size to simulate the displacement changes that may occur during the acquisition process of the spectrogram.

[0079] Scaling operation: Randomly scale the spectrogram with a scaling ratio of 0.8 to 1.2 times to enhance the model's recognition ability for discharge characteristics at different scales.

[0080] Resampling operation: Use bilinear interpolation method to resample the scaled spectrogram to ensure image quality.

[0081] Data augmentation result: After the above operations, the number of spectrograms for each discharge type is expanded from 1000 to 5000, and the total dataset reaches 20000.

[0082] Step 4: Object annotation Annotation tool: Use the LabelImg annotation platform to perform object annotation on the PRPD spectrograms in the training set.

[0083] Annotation content: Annotate the bounding box coordinates of the discharge characteristics and the corresponding discharge type labels (corona discharge, floating discharge, surface discharge, air gap discharge).

[0084] Annotation file generation: Generate an XML file containing annotation information, with the format conforming to the Pascal VOC dataset standard.

[0085] Step 5: Model lightweight improvement Backbone network replacement: Use ShuffleNet-V2 to replace the original backbone feature extraction network (CSPDarknet) of YOLOv8n to reduce the number of model parameters and computational complexity.

[0086] Introduction of attention mechanism: Introduce the CoordAttention mechanism in the feature fusion stage of ShuffleNet-V2 to enhance the model's ability to focus on important features and improve detection accuracy.

[0087] Loss function optimization: Introduce the EIOU (Efficient Intersection over Union) loss function to optimize the accuracy of target box regression and accelerate the convergence speed of network training.

[0088] Step 6: Model training Dataset division: Divide the enhanced 20000 spectrograms into a training set, a validation set, and a test set according to the ratio of 8:1:1.

[0089] Training parameter settings: Learning rate: The initial learning rate is set to 0.01, and the cosine annealing strategy is used to dynamically adjust the learning rate.

[0090] Batch size: Set to 16 to adapt to memory limitations.

[0091] Number of training epochs: Train for 100 epochs and evaluate the model performance using the validation set after each epoch.

[0092] Training process: Forward propagation: Input the training set into the model to calculate the prediction results.

[0093] Loss calculation: Use the EIOU loss function to calculate the loss between the prediction results and the true labels.

[0094] Backward propagation: Update the model parameters through the backward propagation algorithm.

[0095] Model saving: After each epoch ends, save the current best model (the lowest validation set loss).

[0096] Step 7: Model evaluation Evaluation metrics: Precision: The ratio of the number of samples predicted as positive and actually positive to the number of samples predicted as positive.

[0097] Mean Average Precision (mAP): The average of the average precisions at different confidence thresholds, reflecting the overall detection performance of the model.

[0098] Floating Point Operations (FLOPs): The number of floating point operations required for the model's forward propagation, measuring the computational complexity of the model.

[0099] Number of model parameters: The total number of trainable parameters in the model, reflecting the storage requirements of the model.

[0100] Evaluation results: Precision: 95% for corona discharge, 92% for floating discharge, 90% for surface discharge, 88% for air gap discharge.

[0101] mAP: 89.5%.

[0102] FLOPs: 12.3 GFLOPs.

[0103] Number of model parameters: 3.2M.

[0104] Step 8: Model deployment and application Model export: Export the trained model to the ONNX format for easy deployment on different platforms.

[0105] Embedded device deployment: Deploy the model to edge computing devices such as NVIDIA Jetson Nano to achieve real-time partial discharge detection.

[0106] Practical application: In the actual detection of power equipment, the model can identify and classify partial discharge types in real time, providing decision-making support for equipment maintenance.

[0107] In the preferred solution, in Step 1, the partial discharge types include tip discharge, floating discharge, air gap discharge, and surface discharge; the above settings can comprehensively cover common partial discharge types, improving the accuracy and reliability of monitoring; at the same time, according to the characteristics of different discharge types, corresponding detection techniques and analysis methods can be adopted to further improve the efficiency and accuracy of fault diagnosis.

[0108] In the preferred solution, the preprocessing in Step 2 includes normalization, grayscale processing, and data augmentation. Among them, the normalization process limits the discharge signal amplitude of the PRPD spectrogram between 0 and 1 through linear transformation; the above settings, the grayscale processing is to convert the PRPD spectrogram into a grayscale image to reduce the data dimension; the data augmentation increases the sample diversity through operations such as rotation, translation, and scaling, thereby improving the generalization ability of subsequent model training.

[0109] In the preferred solution, the data augmentation step includes random translation, scaling, resampling, and rotation operations to increase the diversity of PRPD spectrogram data; the above settings aim to simulate data changes under different acquisition conditions, thereby improving the generalization ability of subsequent deep learning models; in addition, noise addition and filtering processing are introduced to further enrich the dataset and ensure the robustness of model training.

[0110] In the preferred solution, the division ratio of the training set, validation set, and test set in Step 3 is 8:1:1; the above settings aim to ensure that the model is fully trained while retaining sufficient validation and test sets to evaluate the generalization ability; in addition, a cross-validation strategy is adopted to further reduce the risk of overfitting and improve the stability and reliability of the model.

[0111] In the preferred solution, the lightweight improvement of the YOLOv8n model in Step 4 is to use the ShuffleNet-V2 network structure to replace the original backbone feature extraction network of YOLOv8n, and add the CoordAttention mechanism and EIOU loss function to the network; the above settings effectively reduce the number of model parameters, improve the feature extraction ability and localization accuracy, enabling the improved YOLOv8n model to achieve faster inference speed and lower computational resource consumption while maintaining high detection accuracy.

[0112] In the preferred solution, the lightweight improvement in Step 4 includes replacing the original backbone feature extraction network of YOLOv8n with ShuffleNet-V2 and optimizing other parts of the network to reduce the amount of computation and the number of parameters; the ShuffleNet-V2 network structure significantly reduces the amount of computation through the combination of grouped channel convolution and point convolution, and enhances the network's ability to express features through the channel shuffle mechanism; the above settings enable the model to achieve faster inference speed and lower resource occupancy while maintaining high accuracy; in addition, depthwise separable convolution is introduced to further optimize the network, further compressing the model size and enhancing the deployment efficiency and compatibility of the model in practical applications.

[0113] In the preferred solution, the CoordAttention mechanism is introduced into the Neck of the lightweight improved YOLOv8n model in Step 4. The CoordAttention mechanism is used to enhance the ability to focus on important features in the improved YOLOv8n model, improve the feature extraction ability and classification performance in complex backgrounds. This mechanism performs one-dimensional global average pooling on the input feature map in the horizontal and vertical directions respectively to obtain position information and generate attention weights for height and width. Finally, the input feature map is weighted and corrected; the above settings enable the YOLOv8n model to significantly improve the positioning accuracy and recognition accuracy of target objects while remaining lightweight, especially in complex and changing scenarios, effectively reducing false detection and missed detection cases, and enhancing the robustness and practicality of the model.

[0114] In the preferred solution, the EIOU loss function is also introduced into the model improved by lightweight YOLOv8n in Step 4. The EIOU loss function is used to accelerate the training convergence speed of the network and optimize the accuracy of bounding box regression in the improved YOLOv8n model. This function introduces a center distance term and width-height penalty terms on the basis of the traditional IoU loss to more comprehensively measure the difference between the predicted box and the ground truth box; the above settings enable the improved YOLOv8n model to be significantly improved in detection speed and accuracy, especially in the detection tasks of small targets and complex backgrounds, effectively enhancing the generalization ability and robustness of the model.

[0115] In the preferred solution, the target annotation in Step 5 uses the LabelImg annotation platform to generate an annotation file containing the coordinates of the discharge feature bounding box and the corresponding discharge type label; the above settings can efficiently and accurately mark the discharge features in the image, providing a high-quality data set for subsequent model training; at the same time, the ease of use of the LabelImg platform also greatly improves the efficiency of the annotation work.

[0116] In a preferred solution, the training in Step 5 is to iteratively train the improved YOLOv8n model using optimization algorithms such as gradient descent. The network weights are continuously adjusted through the optimization algorithm in the model until the loss function value of the model on the training data converges below a predetermined threshold. The optimization algorithm is the gradient descent algorithm or its variant. The above settings can significantly improve the detection accuracy and generalization ability of the model. At the same time, an early stopping strategy is adopted to prevent overfitting and ensure the stable performance of the model on the validation set. In addition, a learning rate decay mechanism is introduced to further optimize the training process and accelerate the convergence of the model.

[0117] In a preferred solution, the performance evaluation in Step 7 is to comprehensively evaluate the performance of the improved YOLOv8n model in the local discharge type detection and classification task by calculating performance indicators such as precision, mean average precision, computational complexity, and the number of model parameters. The above settings ensure that the model has a high-precision local discharge type recognition ability while maintaining efficient computation, providing reliable technical support for the stable operation of the power system, and effectively improving the efficiency and accuracy of fault detection and diagnosis.

[0118] In summary, the present invention proposes a method for detecting and classifying partial discharge types based on a lightweight YOLOv8n model of PRPD spectrograms. This method effectively solves the problems existing in the field of power equipment monitoring technology, such as low accuracy of partial discharge type detection and classification, high model complexity, large computational amount, and poor real-time performance. Compared with traditional signal analysis or partial discharge detection methods based on machine learning, the present invention first applies the object detection technology in deep learning (especially the YOLOv8n model) to the detection and classification of PRPD spectrograms, realizing high-precision automatic recognition of partial discharge types. It not only improves the accuracy of detection but also significantly enhances the detection efficiency, capable of simultaneously identifying multiple partial discharge types. Aiming at the problem of limited resources of edge computing devices, this solution makes a lightweight improvement to the YOLOv8n model. By adopting ShuffleNet-V2 as the backbone feature extraction network, the computational amount and the number of parameters of the model are significantly reduced, making this method applicable to resource-constrained embedded devices and edge computing environments, meeting the requirements of real-time detection. At the same time, the lightweight model improves the stability and reliability of the system while maintaining high detection accuracy. To further enhance the feature extraction ability and classification performance of the model in complex backgrounds, the present invention introduces the CoordAttention mechanism. This mechanism enables the model to more accurately identify partial discharge features by enhancing the model's attention ability to important features, improving the accuracy of classification. In addition, the introduction of the EIOU (Efficient Intersection over Union) loss function also speeds up the training convergence speed of the network and optimizes the accuracy of target box regression, further improving the performance of the model. The detection and classification method of the present invention has broad application prospects. By deploying the trained model to the power system site, real-time monitoring and fault prevention of power equipment can be achieved, providing reliable technical support for the safe operation of the power system. At the same time, this method can also be extended to other fields that require high-precision target detection, such as industrial detection, security monitoring, etc., providing important technical guarantees for the safe production and intelligent management of all walks of life.

Claims

1. A method for detecting and classifying partial discharge types based on the PRPD spectrogram using a lightweight YOLOv8n model, characterized in that, It includes the following steps: Step1: Collect PRPD spectrogram data of partial discharge types; Step2: Preprocess the PRPD spectrogram; Step3: Divide the training set, validation set, and test set; Step4: Lightweight improvement of the YOLOv8n model; Step5: Use the annotation platform to perform target annotation on the PRPD spectrograms in the training set, and input the annotated training set into the improved YOLOv8n model for training; Step6: Use the validation set data to evaluate the model performance and adjust the hyperparameters; Step7: Use the test set to evaluate the performance of the improved model; Step8: Deploy the trained improved YOLOv8n model to the power system site for real-time detection and identification of partial discharge types in power equipment.

2. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 1, characterized in that: The partial discharge types in Step1 include tip discharge, floating discharge, air gap discharge, and surface discharge.

3. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 1, characterized in that: The preprocessing in Step2 includes normalization, grayscale processing, and data augmentation. Among them, the normalization process limits the discharge signal amplitude of the PRPD spectrogram between 0 and 1 through linear transformation.

4. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 3, characterized in that: The data augmentation steps include random translation, scaling, resampling, and rotation operations.

5. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 1, characterized in that: The lightweight improvement of the YOLOv8n model in Step4 adopts the ShuffleNet-V2 network structure, and adds the CoordAttention mechanism and the EIOU loss function to the network. The ShuffleNet-V2 network structure reduces the computational amount through the combination of channel group convolution and point convolution, and improves the network's ability to express features through the channel shuffle mechanism.

6. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 5, characterized in that: The CoordAttention mechanism performs one-dimensional global average pooling on the input feature map in the horizontal and vertical directions respectively to obtain position information, generates attention weights for height and width, and finally performs weighted correction on the input feature map.

7. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 6, characterized in that: The EIOU loss function introduces a center distance term and a width-height penalty term on the basis of the traditional IoU loss to more comprehensively measure the difference between the predicted box and the true box.

8. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 1, wherein: The target annotation in Step5 uses the LabelImg annotation platform to generate an annotation file containing the coordinates of the discharge feature bounding box and the corresponding discharge type label.

9. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 8, characterized in that: The training in Step5 iteratively trains the improved YOLOv8n model using optimization algorithms such as gradient descent. The network weights are continuously adjusted through the optimization algorithm in the model until the loss function value of the model on the training data converges below a predetermined threshold. The optimization algorithm is the gradient descent algorithm or its variant.

10. The method for detecting and classifying partial discharge types of the lightweight YOLOv8n model based on the PRPD spectrogram according to claim 1, characterized in that: The performance evaluation in Step7 comprehensively evaluates the performance of the improved YOLOv8n model in the partial discharge type detection and classification task by calculating performance indicators such as precision, mean average precision, computational amount, and number of model parameters.

Images