Detection method for GIS partial discharge PRPD atlas under small sample condition

Abstract

A detection method for a GIS partial discharge PRPD atlas under a small sample condition realizes high-precision partial discharge detection by combining an improved CoTDCGAN model of a deep convolutional generative adversarial network (DCGAN) and an improved YOLO v8 model based on a VanilaNet network structure, utilizes a context converter attention module to enhance atlas data features through the CoTDCGAN model, and improves the partial discharge detection accuracy. Through the improved YOLO v8 model, the Normalization-based Attention Module (NAM) and the Lion optimizer are integrated, the recognition performance is optimized, rapid and accurate PRPD atlas intelligent recognition is provided, the accuracy and training efficiency of the model are improved, and the method is particularly suitable for actual power system monitoring with the limited sample number.

Description

Technical Field

[0001] The invention belongs to the field of power equipment fault detection, and in particular relates to a detection method for GIS partial discharge PRPD spectrum under small sample conditions. Background Art

[0002] Gas insulated switchgear (GIS) is an important component of modern power systems, with the advantages of low maintenance cost, small footprint, and strong anti-interference ability. However, partial discharge (PD) phenomenon occurring during the operation of GIS equipment can cause equipment performance degradation or even damage. Therefore, timely and accurate detection of GIS PD is crucial to the stable operation of the power system.

[0003] Currently, the commonly used method for detecting partial discharge in GIS is mainly the discharge phase distribution (PRPD) spectrum analysis method. Due to the sealing characteristics of GIS equipment and the high voltage environment, it is very difficult to obtain high-quality PRPD spectrum data, resulting in a limited number of samples. It is difficult to meet the data set requirements of existing intelligent detection algorithms to achieve real-time and accurate fault detection. Therefore, this application provides a detection method based on small samples. Summary of the invention

[0004] The purpose of the present invention is to provide a method for efficiently detecting GIS partial discharge PRPD spectra under small sample conditions. The method uses deep learning technology to expand and accurately identify limited data to improve the accuracy and efficiency of detection.

[0005] A detection method for GIS partial discharge PRPD spectrum under small sample conditions, characterized in that the method comprises the following steps:

[0006] S1, collect the original PRPD spectrum data of GIS partial discharge;

[0007] S2. We apply a CoTDCGAN model based on a deep convolutional generative adversarial network (DCGAN) to augment the original PRPD atlas data. The model integrates a context transformer attention (CoTAttention) module to improve the quality and relevance of the generated data by enhancing the context relevance of the data feature representation;

[0008] S3. Apply the YOLO v8 model based on the improved VanillaNet network structure to the PRPD map after data expansion. The model integrates the Normalization-based Attention Module (NAM) attention mechanism and optimizes parameters through the Lion optimizer to achieve training and high-precision intelligent recognition of the PRPD map.

[0009] S4. Analyze the results of intelligent identification, determine the type and status of partial discharge in GIS equipment, and provide accurate diagnostic information;

[0010] The data augmentation step of the CoTDCGAN model includes feature enhancement of the input graph using a context transformer attention module and adjusting the learning rate to optimize the quality and similarity of the generated graph;

[0011] The intelligent recognition step further includes introducing a NAM attention mechanism into the YOLOv8 model to enhance the model's ability to recognize key features of the PRPD graph, and using the Lion optimizer to adjust network parameters to optimize the speed and accuracy of the recognition process;

[0012] The NAM attention mechanism includes channel attention and spatial attention modules, calculates the importance of weights through the scaling factor in batch normalization, and combines sparse regularization terms to enhance the feature extraction capability of the model;

[0013] The Lion optimizer calculates updates through symbolic operations to achieve a uniform update amplitude and reduce memory overhead to improve the generalization ability and training efficiency of the model;

[0014] The data-expanded PRPD atlas is used to train the improved YOLO v8 model to ensure that the model has high recognition accuracy and fast response capability in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of a method for detecting a PRPD spectrum of a GIS partial discharge in an embodiment of the present invention.

[0016] Figure 2 It is a structural diagram of the CoTAttention attention module in an embodiment of the present invention.

[0017] Figure 3 Schematic diagram of the structure of the CoTDCGAN model framework in an embodiment of the present invention.

[0018] Figure 4 It is a structural framework diagram of the YOLOv8 model improved based on VanillaNet in an embodiment of the present invention.

[0019] Figure 5 It is the overall architecture diagram of the NAM attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The specific implementation of the present invention is described in detail below with reference to the accompanying drawings. To ensure the clarity and accuracy of the implementation, the embodiments of the present invention will gradually introduce the design and optimization details of each step and model.

[0021] 1. Collect the original PRPD spectrum data of GIS partial discharge.

[0022] 2. Data preprocessing and grayscale processing

[0023] Considering that the coordinates of the PRPD spectrum are fixed, its shape and background color have little effect on the identification of the discharge type, so the present invention performs grayscale processing on the PRPD spectrum. Specifically, the weighted average method is used to convert the color spectrum into a black and white spectrum, and the formula is as follows:

[0024] γ=αR+βG+γB

[0025] Among them, γ represents the gray value, R, G, B represent the intensity values ​​of the red, green and blue channels respectively. α, β, γ are the weight coefficients of the corresponding color channels;

[0026] The memory usage of a single atlas is about 4-6KB. This processing method can not only significantly reduce data redundancy, but also effectively shorten the time required for model training and improve the overall training efficiency. The processed atlas is stored in corresponding named folders according to the fault category, and divided into training set and verification set in a ratio of 7:3, laying an efficient data foundation for subsequent model training and analysis.

[0027] 3. Apply a CoTDCGAN model improved by deep convolutional generative adversarial network (DCGAN) to expand the original PRPD atlas data

[0028] In traditional machine learning tasks, insufficient data often leads to unsatisfactory model training results. To solve this problem, the present invention adopts a CoTDCGAN model based on the deep convolutional generative adversarial network (DCGAN) for data expansion. The uniqueness of the CoTDCGAN model is that it integrates a context transformer attention (CoTAttention) module, which can enhance feature capture based on the context information of the graph. Specifically, during the generation of the graph, the CoTAttention module can extract more refined local features by modeling the context of the original graph, thereby improving the quality and diversity of the generated graph.

[0029] At the same time, the generation quality is further improved by adjusting the initial learning rate. In the initial stage of training, the learning rate of the discriminator is appropriately reduced to give the generator more room for improvement; and the learning rate of the discriminator is increased to enhance its discrimination ability;

[0030] The present invention sets the discriminator learning rate to 0.0001 and the discriminator learning rate to 0.0004. This strategy helps to establish a more effective adversarial balance between the generator and the discriminator, and ultimately improves the overall quality of the generated image;

[0031] Through the above method, the CoTDCGAN model can iteratively generate a large number of PRPD data maps during the data expansion process, and select high-quality and high-similarity PRPD maps from them to significantly increase the amount of sample data, overcome the limitations of small sample data, and improve the training effect of subsequent models.

[0032] 4. After data expansion, apply the YOLO v8 model based on the improved VanillaNet network structure

[0033] After data expansion, the YOLOv8 model based on the improved VanillaNet structure is used. YOLOv8 is an advanced target detection model. In the present invention, it is used to identify the local discharge features in the PRPD map. In order to further enhance the performance of the model in feature recognition, the application integrates the Normalization-based Attention Module (NAM) attention mechanism in YOLOv8.

[0034] The NAM attention mechanism mainly optimizes the model's attention to graph features at different scales through channel attention and spatial attention modules;

[0035] In the channel dimension, this module uses the scaling factor γ of batch normalization (BN) to calculate the importance of each channel and measures the significance of each channel through the variance metric. The formula is as follows:

[0036] W γ =γ i / ∑ j＝0 γ j

[0037] Among them, γ is the scaling factor of BN. This method avoids introducing additional fully connected or convolutional layers, thus keeping the network lightweight.

[0038] In the spatial dimension, NAM measures the importance of each pixel by the scaling factor of BN, thereby constructing the so-called "pixel normalization". The weights in this module are calculated by the following formula:

[0039] W λ =λi / ∑ j＝0 λ j

[0040] In addition, the present invention also uses Lion optimizer to optimize the YOLOv8 model. Lion optimizer has the advantages of uniform update amplitude, regularization effect, momentum tracking and fewer hyperparameters. It calculates updates through symbolic operations, reduces memory overhead, and achieves regularization by introducing noise, thereby improving the generalization performance of the model;

[0041] The Lion optimizer uses the default exponential moving average (EMA) factor to track momentum and balances the weights of the current gradient and momentum through an interpolation factor, thereby speeding up training and improving the generalization ability of the model.

[0042] After a series of improvements, the network structure of the YOLOv8 model has been significantly optimized and its network parameters have been significantly reduced. These optimizations not only improve the model's effectiveness in extracting features, but also enhance its generalization capabilities in different data environments.

[0043] 5. Training environment and parameter settings

[0044] The software environment used for the training experiment of the present invention is a deep learning framework built with Windows 10, Python 3.8, Torch 1.10.0, CUDA v10.2, and cuDNN v7615, and the version used is Ultralytics YOLO v8.0.196; the hardware environment is that the CPU is Intel (R) Core (TM) i7-8750H @ 2.20GHz, the graphics card is NVIDIA GeForce GTX1050, the memory (RAM) is 16GB DDR4, and the storage is 128GB SSD + 1TB HDD; the input image size is 224 pixels × 224 pixels, the training rounds (Epochs) are 200 rounds, the batch size (Batch Size) is 256, and the other training parameters are kept as default.

[0045] 6. Analyze the results of the intelligent identification to determine the type and operating status of partial discharge in the GIS equipment

[0046] The deep learning model trained above can accurately detect and classify the input partial discharge map;

[0047] Specifically, the graph to be tested is input into the trained model. The model first standardizes the input graph, including resizing and normalization operations, to match the input requirements of the network. Then, the multi-level features of the input graph are extracted through the forward propagation process, and the detection results are generated based on the fully connected classification layer.

[0048] The detection result is a multi-dimensional probability vector, each dimension of which corresponds to the predicted probability of a partial discharge type. By analyzing the probability vector, the category with the highest probability value is determined as the final classification result, and the confidence score of the category is output.

[0049] Specifically, if the test result is (partial discharge type A, 0.92), it means that the spectrum is predicted to be "partial discharge type A" and the prediction confidence of the model is 92%. Combined with the test result, the type and operating status of partial discharge in GIS equipment can be further comprehensively analyzed and diagnosed to provide technical support for the operation and maintenance of the equipment.

[0050] Although the present invention has been described in detail with reference to the drawings and embodiments, those skilled in the art may make various improvements and changes to the present invention without departing from the scope of the claims, and these improvements and changes should be regarded as the protection scope of the present invention.

Claims

1. A method for detecting PRPD spectrum of partial discharge in GIS under small sample conditions, characterized by: The method comprises the following steps: S1, collect the original PRPD spectrum data of GIS partial discharge; S2. Using the CoTDCGAN model improved based on the deep convolutional generative adversarial network (DCGAN), the original PRPD atlas data is expanded, wherein the CoTDCGAN model integrates the context transformer attention (CoTAttention) module; S3. After data expansion, the YOLO v8 model based on the improved VanillaNet network structure is applied. The model integrates the Normalization-based Attention Module (NAM) attention mechanism and is optimized by the Lion optimizer to achieve training and intelligent recognition of PRPD maps. S4. Analyze the results of the intelligent identification to determine the type and state of partial discharge in the GIS equipment.

2. The detection method for GIS partial discharge PRPD spectrum under small sample conditions according to claim 1, characterized in that: The data expansion step S2 includes: a) using the CoTDCGAN model to perform feature enhancement on the input graph through the context transformer attention module; b) adjusting the learning rate to optimize the quality and similarity of the generated graph.

3. The detection method for GIS partial discharge PRPD spectrum under small sample conditions according to claim 1 is characterized in that: The intelligent recognition further includes: a) introducing the NAM attention mechanism into the YOLOv8 model to enhance the model's recognition ability for key features of the PRPD graph; b) using the Lion optimizer to adjust network parameters to optimize the speed and accuracy of the recognition process.

4. The detection method for GIS partial discharge PRPD spectrum under small sample conditions according to claim 1 is characterized in that: The NAM attention mechanism includes channel attention and spatial attention modules, calculates the importance of weights through the scaling factor in batch normalization, and combines with sparse regularization terms to enhance the feature extraction capability of the model.

5. The detection method for GIS partial discharge PRPD spectrum under small sample conditions according to claim 1, characterized in that: The Lion optimizer calculates updates through symbolic operations to achieve a uniform update amplitude and reduce memory overhead to improve the generalization ability and training efficiency of the model.

6. The detection method for GIS partial discharge PRPD spectrum under small sample conditions according to claim 1, characterized in that: The data-expanded PRPD atlas is used to train the improved YOLO v8 model to ensure that the model has high recognition accuracy and fast response capability in practical applications.

Images