Substation equipment partial discharge detection method and device based on improved CSP-GCNet and medium

Abstract

A transformer partial discharge detection method based on improved CSP-GCNet, collects original PD signals, and carries out VMD-Hilbert post-processing; the original database is processed and sample expansion is carried out to increase the data quantity and diversity; the improved CSP-GCNet model is built by combining the CSPnet network structure, GhostConv ghost convolution, Focus network structure and SiLU activation function.The characteristic set of the input atlas can be effectively extracted by the feature extraction process based on the improved CSP-GCNet, and the key features of the transformer partial discharge mode can be captured.The algorithm has important application value in the field inspection of power distribution equipment, and helps to monitor the running state of the transformer in real time.

Description

Technical Field

[0001] This invention belongs to the field of power equipment monitoring technology, specifically relating to a method, device and medium for detecting partial discharge in substation equipment based on an improved lightweight cross-level local network CSP-GCNet. Background Technology

[0002] With the continuous expansion and development of power systems, the requirements for the safe operation of power lines and the reliability of power supply are increasing. Power lines play a crucial role in the power grid, and their safe and stable operation has a decisive impact on ensuring the integrity of the power grid structure. In power transformers, the insulation condition is closely related to the overall operation of the power grid. Among them, partial discharge (PD) is a phenomenon caused by the deterioration of transformer insulation performance, which effectively reflects an important indicator of insulation defects inside and outside the transformer. Different insulation deterioration mechanisms lead to different types of discharge, and these discharge types also differ in their external manifestations and severity. Therefore, accurately identifying and monitoring partial discharge types is of vital importance for assessing the insulation condition of equipment, and is also a research hotspot in the field of transformer fault diagnosis and location. However, traditional methods suffer from slow computation speed and low accuracy in map classification. Summary of the Invention

[0003] To overcome the shortcomings of the existing technology and achieve partial discharge mode classification, this invention proposes a partial discharge detection method for power equipment based on an improved CSP-GCNet. By utilizing a series of key technologies and network structures, this method reflects the insulation defects existing inside and outside the transformer, achieving efficient and accurate spectrum classification and rapid and accurate detection of transformer partial discharge.

[0004] The core idea of ​​this invention is as follows:

[0005] By acquiring the raw signals of partial discharge in transformers and utilizing a specially designed network structure and feature extraction method, accurate classification of partial discharge modes in transformers can be achieved.

[0006] Firstly, partial discharge is a key indicator during transformer inspection. Partial discharge refers to the partial discharge phenomenon present in the transformer's insulation system. Its characteristic signals can be used to determine whether the transformer has insulation defects and to perform corresponding fault diagnosis. Therefore, during the inspection of power distribution equipment, it is necessary to first collect the raw signals of partial discharge in the transformer.

[0007] Secondly, a specially designed network structure and feature extraction method were employed. A feature extraction process based on an improved CSP-GCNet model was established. This model combines key technologies such as the CSPnet network structure, the Focus network structure, and the introduction of the SiLU activation function. Through the CSP-GCNet model, the feature set of the input map can be effectively extracted, thereby capturing the key features of transformer partial discharge modes. In terms of network structure, a residual convolution module (Residual) was introduced to enhance the network's ability to represent features.

[0008] Furthermore, the CSPnet network structure, combined with the Focus network structure, is adopted to further optimize the feature extraction process. The SPP structure can adaptively divide the input features into blocks and perform max pooling on the features within each block, thereby achieving effective fusion of features at different scales. The SiLU activation function is employed. Compared to the traditional ReLU activation function, the SiLU activation function has better smoothness and non-linearity, effectively improving the model's classification performance.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] On the one hand, this invention provides a transformer partial discharge detection method based on an improved CSP-GCNet, characterized in that it includes:

[0011] Acquire raw partial discharge signals from the transformer, including discharge signals under normal and abnormal conditions.

[0012] The acquired raw signal is decomposed into several intrinsic mode components with different center frequencies by variational mode decomposition (VMD-Hilbert post-processing). Then, Hilbert transform is performed to obtain the Hilbert spectrum. The Hilbert spectrum is integrated over the entire time domain to obtain the Hilbert spectrum of the original signal.

[0013] Based on the Hilbert spectrum of the original signal obtained in the previous step, an image classification database is constructed.

[0014] A model based on the improved CSP-GCNet network was built for transformer partial discharge pattern classification;

[0015] Based on the established image classification database, the model is trained, validated, and its pattern recognition accuracy is evaluated using average precision. Finally, the task of identifying the partial discharge pattern of a transformer is completed.

[0016] Building an image classification database from the original atlas, specifically including:

[0017] The Hilbert spectra of the acquired raw signals are scaled down to ensure that the spectra have a uniform size and resolution; and the scaled spectra data are divided into training set, test set and validation set in an 8:1:1 ratio to construct a spectra classification dataset.

[0018] Graph preprocessing techniques, including cropping, rotation, flipping, and deformation operations, are employed to increase the diversity and richness of the graph classification dataset.

[0019] The spectral data is filtered and cleaned, and classified and labeled according to specific standards to build an original spectral database.

[0020] Furthermore, a transformer partial discharge spectrum classification method based on the improved CSP-GCNet is constructed, specifically including:

[0021] A residual convolution module, Residual, is introduced, which includes a backbone for extracting map features and residual edge parts for input and output of the backbone, and is connected by skip connections.

[0022] Building the CSPnet network structure includes:

[0023] -Feature extraction: The input map is processed through convolutional layers and pooling layers to extract high-level map features;

[0024] - Divide the feature map into two parts. One part is stacked with continuous residual blocks of the backbone, and the other part is directly connected to the final output of the network as residual edges to construct the CSP structure.

[0025] - The features of the backbone and residual edges are fused using channel-level overlay or weighted summation;

[0026] Based on the CSPnet network structure, the GhostConv module is introduced to build the CSP-GCNet network structure.

[0027] Based on the CSP-GCNet network structure, a Cross Stage Partial module is introduced to divide the input features into two branches and introduce a feature reuse mechanism between the branches;

[0028] A Focus network structure is built by reducing the size and number of channels of the input feature map;

[0029] Introducing the SiLU activation function, the formula is as follows:

[0030] f(x) = x·sigmoid(x)

[0031] In the formula, x is the input value, and sigmoid is the sigmoid activation function;

[0032] An SPP network structure is constructed to perform multi-scale feature fusion. The input features are adaptively divided into blocks, and the features within each block are max-pooled to achieve the fusion of features at different scales.

[0033] Furthermore, based on the established image classification database, the model is trained and validated. The average precision is used to evaluate the model's pattern recognition accuracy. Finally, the transformer partial discharge pattern recognition task is completed, specifically:

[0034] During training, the optimized CSP-GCNet model is applied to feature extraction from the input map;

[0035] After training, the CSP-GCNet model is evaluated using a validation set. The performance of the CSP-GCNet model is assessed by calculating its average accuracy on the validation set, thereby continuously optimizing the CSP-GCNet model and improving its accuracy in transformer partial discharge pattern mapping tasks. On the other hand, this invention also provides a transformer partial discharge detection device based on the improved lightweight cross-level local network CSP-GCNet, characterized by comprising:

[0036] The signal acquisition unit is used to acquire the original partial discharge signal of the transformer, including discharge signals under normal and abnormal conditions;

[0037] The signal processing unit is used to decompose the acquired raw signal into several intrinsic mode components with different center frequencies through variational mode decomposition (VMD-Hilbert post-processing), then perform Hilbert transform to obtain the Hilbert spectrum, and integrate the Hilbert spectrum over the entire time domain to obtain the Hilbert spectrum of the original signal.

[0038] The image classification module is used to build an image classification database based on the original signal Hilbert spectrum obtained in the previous step.

[0039] The partial discharge map classification module is used to build a model based on the improved CSP-GCNet network for transformer partial discharge map classification.

[0040] The task recognition module is used to train and validate the model based on the established image classification database, evaluate the model's pattern recognition accuracy using average precision, and finally complete the transformer partial discharge pattern recognition task.

[0041] The present invention also provides a storage medium storing an instruction set, wherein the instruction set, when executed by a processor, implements the above-described transformer partial discharge detection method.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0043] 1) Introducing deep learning technology into the power industry's power distribution equipment inspection has significant application value in power distribution equipment inspection and fault diagnosis. The introduced Focus network structure can improve the accuracy and efficiency of graph classification, providing important technical support for the safe operation of power systems. Simultaneously, it also promotes progress in transformer insulation condition assessment and fault diagnosis research, playing a positive role in the development of the power industry.

[0044] 2) Simultaneously, the GhostConv module introduced in this invention can efficiently and accurately extract features from the partial discharge mode map of transformers and reliably classify them. It has broad application prospects in the transformer industry and other signal processing and mode classification tasks. Attached Figure Description

[0045] Figure 1 This is a flowchart of the transformer partial discharge detection method based on the improved CSP-GCNet of the present invention.

[0046] Figure 2 This is an improved overall architecture diagram of CSP-GCNet.

[0047] Figure 3 This is a schematic diagram of the Residual Network.

[0048] Figure 4 This is a schematic diagram of the CSPnet network structure.

[0049] Figure 5 This is a schematic diagram of the GhostConv convolutional module.

[0050] Figure 6 This is a schematic diagram of the Focus network structure.

[0051] Figure 7 This is a data diagram illustrating the training process. Detailed Implementation

[0052] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the scope of protection of the present invention.

[0053] Acquiring raw partial discharge signals from transformers is crucial, as partial discharge is a key indicator of insulation defects both inside and outside the transformer. During power distribution equipment inspection, acquiring raw partial discharge signals from transformers is a primary task. To further improve acquisition accuracy, environmental noise interference is considered. By appropriately setting filters and gain control, external interference signals are suppressed, maintaining the purity and reliability of the acquired partial discharge signals. Through the acquisition of raw partial discharge signals from a large number of transformers, combined with relevant annotation information, a high-quality partial discharge dataset is established, containing rich partial discharge patterns and features, providing a reliable foundation for subsequent algorithm training and performance evaluation.

[0054] Please refer to the figure. Figure 1 This is a flowchart of the transformer partial discharge detection method based on the improved CSP-GCNet of the present invention. As shown in the figure, a method for detecting partial discharge in power equipment based on the improved CSP-GCNet includes the following steps:

[0055] Step (1): Collect partial discharge signals of the transformer, including discharge signals under normal and abnormal conditions.

[0056] During transformer inspection, partial discharge is a key indicator. Collecting the raw signals of partial discharge reveals insulation defects both inside and outside the transformer. Partial discharge refers to the partial discharge phenomenon within the transformer's insulation system. Its characteristic signals can be used to determine the presence of insulation defects and perform corresponding fault diagnosis. Therefore, collecting the raw signals of partial discharge from transformers is necessary during the inspection of power distribution equipment.

[0057] Step (2): Perform VMD-Hilbert post-processing on the acquired raw signal. The VMD algorithm decomposes the signal into several intrinsic mode components with different center frequencies and finite bandwidths, which have good noise robustness. Then, perform Hilbert transform on the VMD algorithm result to obtain the Hilbert spectrum, and integrate the Hilbert spectrum over the entire time domain to obtain the Hilbert spectrum of the original signal.

[0058] Step (2.1): By employing the VMD (Variational Mode Decomposition) algorithm, the original PD signal is decomposed into finite-bandwidth intrinsic mode components with different center frequencies, thereby achieving spectral decomposition of the signal. This characteristic makes the VMD algorithm more advantageous in processing transformer partial discharge mode pattern classification. The noise robustness of the VMD algorithm can reduce the impact of interference on the signal decomposition results, thereby improving the accuracy and reliability of the algorithm.

[0059] Step (2.2): By applying the Hilbert transform to the output of the VMD algorithm, the Hilbert spectrum of the signal is obtained. The Hilbert transform is an effective frequency domain analysis method that can convert a signal into an analytic signal and provide signal amplitude and phase information. By analyzing the Hilbert spectrum, important information about the characteristics of the original signal can be obtained.

[0060] Step (2.3): Integrate the Hilbert spectrum over the entire time domain to obtain the Hilbert spectrum of the original signal. By integrating the Hilbert spectrum, we can obtain the energy distribution information of the signal in the time domain, which helps to more accurately describe the partial discharge mode of the transformer.

[0061] This invention employs VMD-Hilbert post-processing on the original PD discharge signal. By decomposing the signal into intrinsic mode components with finite bandwidth and extracting the Hilbert spectrum and energy distribution information, it achieves accurate classification and analysis of the transformer partial discharge mode map. This algorithm can not only be applied to the transformer industry but also extended to signal processing and mode classification tasks in other fields.

[0062] Step (3): Create a graph classification dataset, scale the dataset appropriately, divide it into training set, test set and validation set in a ratio of 8:1:1, and build the original graph database.

[0063] Step (3.1): Create a graph classification dataset for training, testing, and validation, and scale it appropriately to ensure data accuracy. To evaluate the algorithm's performance, the dataset was divided into training, testing, and validation sets in an 8:1:1 ratio.

[0064] Step (3.2): During the dataset creation process, graph preprocessing techniques are employed, including operations such as cropping, rotating, flipping, and transforming, to increase the diversity and richness of the dataset. Through these operations, we can expand the scale of the dataset, enhance the robustness of the model, and thus better adapt to real-world application scenarios.

[0065] Step (3.3): The data was filtered and cleaned, and classified and labeled according to specific criteria to facilitate subsequent model training and performance evaluation. This step ensures the accuracy and consistency of the dataset, laying the foundation for the effective implementation of the algorithm.

[0066] By implementing the above steps, the reliability and accuracy of the graph classification algorithm can be improved. Creating appropriate datasets, applying diverse graph preprocessing techniques, and filtering and cleaning the data help optimize algorithm performance and improve its applicability in practical applications. The graph classification algorithm of this invention has broad application prospects in transformer partial discharge pattern recognition and other related fields.

[0067] Step (4): Build a transformer partial discharge map classification algorithm based on the improved CSP-GCNet. This algorithm adopts a specially designed network structure and feature extraction method to achieve efficient and accurate map classification. A feature extraction process based on the improved CSP-GCNet model was built. This model combines key technologies such as the CSPnet network structure, GhostConv phantom convolution, Focus network structure, and the introduction of the SiLU activation function. Through the CSP-GCNet model, we can effectively extract the feature set of the input map, thereby capturing the key features of the transformer partial discharge mode. In terms of network structure, a residual convolution module Residual is introduced to enhance the network's ability to represent features. This module, by introducing residual connections in the network, realizes cross-layer feature transfer and information fusion, improving the network's deep feature expression capability. The CSPnet network structure was adopted, the GhostConv phantom convolution module was introduced, and the Focus network structure was combined to further optimize the feature extraction process. The CSP-GCNet network architecture introduces a CSP (Cross Stage Partial) module, dividing the input features into two branches and introducing a feature reuse mechanism between branches, effectively improving the expressive power of the features. The Focus network architecture, on the other hand, reduces the computational complexity of the network by decreasing the size and number of channels of the input feature map, thus accelerating the training and inference speed of the model. For the activation function, the SiLU activation function is adopted. Compared with the traditional ReLU activation function, the SiLU activation function has better smoothness and non-linear expressive power, effectively improving the model's classification performance. The SPP (Spatial Pyramid Pooling) structure is used for multi-scale feature fusion. The SPP structure can adaptively divide the input features into blocks and perform max pooling on the features within each block, thereby achieving effective fusion of features at different scales.

[0068] Step (4.1): Introduce a Residual Network, such as Figure 3 As shown, in CSPDarknet, residual convolution consists of two parts: the backbone and the residual edge part. The backbone includes one 1x1 convolution and one 3x3 convolution, while the residual edge part directly combines the input and output of the backbone without any processing. The entire backbone is constructed using residual convolution.

[0069] Residual networks are a type of deep neural network architecture inspired by the concept of residual learning. In traditional neural network training, increasing the number of layers leads to the vanishing gradient problem, making the network difficult to converge and optimize. Residual networks effectively alleviate this problem by introducing skip connections and residual blocks.

[0070] Specifically, the residual block consists of two parts: the backbone and the residual edge parts. The backbone consists of one 1x1 convolution and one 3x3 convolution, used to extract spectral features. The residual edge parts directly combine the input and output of the backbone without any processing. This structure allows information to flow directly between the backbone and the residual edge parts, avoiding the attenuation problem of information passing through layers in the network.

[0071] Skip connections are a key component of residual networks. They add the network's input directly to the output of the residual block, creating a shortcut. In this way, the residual network can learn the difference between the input and output, passing the residual as key information. This skip connection design allows gradients to propagate more easily through the network, avoiding the vanishing gradient problem in deep networks, thus improving the model's optimization performance.

[0072] Because residual networks are easy to optimize and train, their accuracy can be improved by increasing network depth. Compared to traditional neural network structures, residual networks can better capture details and complex features in the partial discharge map, thus improving the performance of map classification tasks. Therefore, residual networks are introduced into the transformer partial discharge map classification algorithm based on the improved CSP-GCNet of this invention to fully utilize their advantages and alleviate the gradient vanishing problem in deep neural networks through skip connections. This design makes the algorithm easier to converge during training and achieves better map classification results. By introducing residual convolution and skip connections, the training effect and accuracy of the algorithm are improved. This algorithm has the potential for wide application and has important practical significance in fields such as transformer partial discharge pattern recognition.

[0073] Step (4.2): Build the CSPnet network structure, such as... Figure 4 As shown.

[0074] The CSPnet architecture splits the original residual block stack into two parts. The backbone continues the original residual block stacking operation to extract spectral features. The other part, similar to a residual edge, is directly connected to the network's final output after minimal processing. This part in CSP can be considered a large residual edge. Specifically, the CSPnet network structure includes the following key steps:

[0075] 1. Feature extraction: The input map is processed through convolutional layers and pooling layers to extract high-level map features.

[0076] 2. CSP structure: The feature map is divided into two parts. One part is stacked with continuous residual blocks of the backbone, and the other part is directly connected to the final output of the network as residual edges.

[0077] 3. Feature fusion: The features of the main body and the residual edges are fused, which can be done by channel-level superposition or weighted summation.

[0078] By using the CSPnet network structure, the graph classification algorithm of this invention can better utilize the advantages of residual learning, and by introducing a large residual edge through the stacking of split residual blocks, it further improves the performance of graph classification. Furthermore, the design of the CSPnet structure also has low computational complexity and a small number of parameters, making the algorithm feasible for practical applications.

[0079] In summary, the transformer partial discharge pattern classification algorithm proposed in this invention, based on an improved CSP-GCNet, effectively improves the accuracy and performance of the pattern classification task by optimizing the stacking method of residual blocks and introducing large residual edges under the guidance of the CSPnet network structure. This algorithm has broad application prospects and significant practical value in fields such as transformer partial discharge pattern classification.

[0080] Step (4.3): Based on the CSPnet of Step (4.2) above, introduce GhostConv to build the CSP-GCNet network. The GhostConv convolutional module, as follows: Figure 5 As shown.

[0081] GhostConv technology reduces computational resource requirements while maintaining model performance. GhostConv performs initial extraction of input feature maps using a small number of convolutional kernels, and then applies cheaper linear transformation operations to further process these feature maps, ultimately generating a concatenated feature map.

[0082] Traditional convolutional operations can learn redundant information, leading to a waste of computational resources. However, the GhostConv method combines a small number of convolutional kernels with inexpensive linear transformation operations to extract key features while reducing the computational cost of redundant features. In this way, the GhostConv model significantly reduces computational resource requirements without compromising model performance, providing a more efficient solution for power distribution equipment inspection.

[0083] The improved CSP-GCNet model of this invention also introduces the CSPnet network structure, the Focus network structure, and the SiLU activation function to enhance performance in the feature extraction process. The CSPnet network structure employs cross-stage connections and separable convolutions, increasing the network's capacity and receptive field. The Focus network structure, by introducing a multi-scale receptive field, captures detailed and global information from the transformer partial discharge pattern spectrum. The SiLU activation function has good nonlinear fitting capabilities, further improving the network's feature extraction performance.

[0084] Step (4.4): Build the Focus network structure, such as... Figure 6 As shown

[0085] By manipulating the input map and sampling at one-pixel intervals at each pixel location, four independent feature layers are obtained. The width and height information of the map are consolidated into the channel information, quadrupling the number of input channels. These four independent feature layers are then stacked together to form a concatenated feature layer, increasing the number of channels to twelve compared to the original three-channel feature layer.

[0086] By using the improved Focus network structure, the graph classification algorithm of this invention can make fuller use of the detailed information within the graph to achieve more accurate classification results. The Focus network structure effectively extracts spatial and channel information from the graph through the stacking and concatenation of multi-channel features, providing richer feature representation capabilities for subsequent classification tasks. At the same time, the design of this network structure also has low computational complexity and a small number of parameters, making the algorithm highly efficient in practical applications.

[0087] In summary, guided by the Focus network structure, the stacking and splicing of multi-channel feature layers effectively improves the accuracy and performance of image classification tasks. This algorithm has broad application prospects in fields such as transformer partial discharge image classification and can provide strong technical support for related industries.

[0088] Step (4.5): Introduce the SiLU activation function, as shown in the formula:

[0089] f(x) = x·sigmoid(x)

[0090] Where x is the input value and sigmoid is the sigmoid activation function.

[0091] The SiLU (Sigmoid-Weighted Linear Unit) activation function is an improvement on the traditional Sigmoid and ReLU activation functions. The SiLU function has the following characteristics: it has both an unbounded upper bound and a finite lower bound, ensuring that the activation value does not increase or decrease indefinitely; its output is continuous, smooth, and non-monotonic, enhancing the expressive power of neural networks. Compared to the traditional ReLU activation function, the SiLU function exhibits better performance in deep models. To some extent, the SiLU function can be considered a smoothed ReLU activation function.

[0092] In this invention, the SiLU activation function is introduced, which benefits the spectral classification algorithm. The use of the SiLU activation function enables the neural network to have better output capabilities when processing spectral features, and can more accurately capture important features in the partial discharge mode spectral data of transformers. Compared with traditional Sigmoid and ReLU functions, the SiLU activation function plays a positive role in improving classification accuracy and model stability.

[0093] Step (4.6): Construct the SPP structure. SPP (Spatial Pyramid Pooling) extracts features through max pooling operations with different pooling kernel sizes to improve the receptive field of the neural network.

[0094] The use of the SPP structure is of great significance to the graph classification algorithm of this invention. By employing max pooling operations with pooling kernels of different sizes, the SPP structure can extract graph features at different scales, thereby expanding the receptive field of the neural network. This multi-scale feature extraction helps to capture detailed and global information in the partial discharge mode graph of transformers, improving the accuracy and generalization ability of graph classification.

[0095] In this invention, the SPP structure is embedded in the backbone feature extraction network, enabling the map classification algorithm to better utilize multi-scale information. By introducing the SPP module into the feature extraction process of transformer partial discharge pattern maps, the algorithm of this invention can effectively capture features at different scales and comprehensively consider their contributions during the classification process. This design makes the algorithm more robust and adaptable to transformer partial discharge pattern maps of various sizes and shapes.

[0096] In summary, this invention provides a transformer partial discharge map classification algorithm based on an improved CSP-GCNet, which incorporates an SPP structure. The SPP structure extracts features through multi-scale max-pooling operations, expanding the receptive field of the neural network and effectively improving the accuracy and generalization ability of map classification. This algorithm brings a new technical approach to the field of transformer fault diagnosis and has broad application prospects.

[0097] Step (5): Train the model, validate the model, calculate the average accuracy, and complete the transformer partial discharge pattern recognition task.

[0098] Step (5.1): During training, we apply the optimized CSP-GCNet model to feature extraction from the input map. This model combines GhostConv technology, the CSPnet network structure, the Focus network structure, and the SiLU activation function to improve feature extraction performance. GhostConv technology reduces the computational cost of redundant features by using a small number of convolutional kernels and inexpensive linear transformation operations. The CSPnet network structure introduces cross-stage connections and separating convolutions, increasing the network's capacity and receptive field. The Focus network structure captures the details and global information of the map by introducing a multi-scale receptive field. The SiLU activation function has good non-linear fitting capabilities.

[0099] Step (5.2): After training, we evaluate the model using a validation set. By calculating the model's average accuracy on the validation set, we can objectively assess the model's performance. Through repeated adjustments to the model's parameters and structure, we continuously optimize the model to improve its accuracy in the transformer partial discharge pattern map classification task.

[0100] Extensive experiments have demonstrated that our transformer partial discharge pattern map classification algorithm based on the improved CSP-GCNet performs exceptionally well in transformer partial discharge detection tasks. Its efficient feature extraction capabilities and accurate classification performance make it an ideal choice for on-site inspection of power distribution equipment. We believe that this algorithm will provide a reliable guarantee for the normal operation of power distribution equipment and bring more advanced and reliable solutions to related fields.

[0101] In summary, the transformer partial discharge pattern classification algorithm based on the improved CSP-GCNet proposed in this invention innovates and optimizes the network structure and feature extraction method. This algorithm can efficiently and accurately extract features from transformer partial discharge pattern maps and reliably classify them. This algorithm has applications in the transformer industry...

[0102] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific embodiments of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

[0103] Comparative ablation experiment

[0104]

[0105] CSP-GCNet-lite refers to CSP-GCNet-lite constructed by introducing GhostConv (phantom convolution) as described in step (4.3) into CSP-Net.

[0106] CSP-GCNet-SiLU refers to CSP-GCNet-SiLU constructed by introducing the SiLU activation function described in step (4.5) into CSP-Net.

[0107] CSP-GCNet-Full refers to the CSP-GCNet-Full network model built by introducing GhostConv (phantom convolution), the Swish activation function, and the Focus network structure described in step (4.4) in CSP-Net.

[0108] In this paper, we present a series of comparative ablation experiments to verify the performance of the improved CSP-GCNet model in transformer partial discharge pattern classification. The experimental design follows rigorous evaluation criteria, including Top1 accuracy, Top5 accuracy, number of parameters, and precision.

[0109] Experimental results show that the proposed versions of CSP-GCNet outperform the traditional ResNet and DenseNet models on multiple metrics. Specifically, the ResNet model achieves a Top-1 accuracy of 81.2%, a Top-5 accuracy of 88.5%, 2.6M parameters, and an accuracy of 0.831. In contrast, the DenseNet model performs slightly better, achieving a Top-1 accuracy of 83.7% and a Top-5 accuracy of 89.4%, but with an increased number of parameters (8.1M) and an accuracy of 0.844.

[0110] In the CSP-GCNet series, we first introduced GhostConv, forming CSP-GCNet-lite. This model achieves 86.1% Top-1 accuracy and 91.2% Top-5 accuracy, with 9.8M model parameters and an accuracy of 0.882, demonstrating a significant improvement over ResNet and DenseNet.

[0111] Subsequently, we introduced the SiLU activation function on top of CSP-Net, forming CSP-GCNet-SiLU, which improved the Top1 accuracy to 86.3%, the Top5 accuracy to 91.6%, the model parameters to 9.9M, and the precision to 0.896.

[0112] Ultimately, the CSP-GCNet-Full model integrates GhostConv (phantom convolution), the Swish activation function, and the Focus network structure. These improvements resulted in a Top-1 accuracy of 87.6%, a Top-5 accuracy of 92.5%, and a precision of 0.912 with 10.2M model parameters. These results fully demonstrate the superior performance of CSP-GCNet-Full in the transformer partial discharge map classification task.

[0113] In summary, through this series of comparative ablation experiments, we can clearly see that the transformer partial discharge map classification algorithm based on the improved CSP-GCNet proposed in this invention surpasses current mainstream image classification network models in terms of accuracy, precision, and model efficiency. These experimental results not only verify the effectiveness of the improved algorithm but also provide reliable technical support for future transformer condition monitoring and fault prediction.

Claims

1. A transformer partial discharge detection method based on an improved lightweight cross-level local network CSP-GCNet, characterized in that, include: Acquire raw partial discharge signals from the transformer, including discharge signals under normal and abnormal conditions. The acquired raw signal is decomposed into several intrinsic mode components with different center frequencies by variational mode decomposition (VMD-Hilbert post-processing). Then, Hilbert transform is performed to obtain the Hilbert spectrum. The Hilbert spectrum is integrated over the entire time domain to obtain the Hilbert spectrum of the original signal. Based on the Hilbert spectrum of the original signal obtained in the previous step, an image classification database is built; A model based on the improved CSP-GCNet network was built for transformer partial discharge pattern classification; Based on the established image classification database, the model is trained, validated, and the average precision is used to evaluate the model's pattern recognition accuracy. Finally, the transformer partial discharge pattern recognition task is completed. A transformer partial discharge spectrum classification method based on an improved CSP-GCNet is constructed, specifically including: A residual convolution module, Residual, is introduced, which includes a backbone for extracting map features and residual edge parts for input and output of the backbone, and is connected by skip connections. Building the CSPnet network structure includes: -Feature extraction: The input map is processed through convolutional layers and pooling layers to extract high-level map features; - The feature map is divided into two parts. One part is stacked with continuous residual blocks of the backbone, and the other part is directly connected to the final output of the network as residual edges to construct the cross-level local network CSPnet structure. - The features of the backbone and residual edges are fused using channel-level overlay or weighted summation; Based on the CSPnet network structure, the GhostConv module is introduced to build the CSP-GCNet network structure. Based on the CSP-GCNet network structure, a Cross Stage Partial module is introduced to divide the input features into two branches and introduce a feature reuse mechanism between the branches; By reducing the size and number of channels of the input feature map, a focus network structure is constructed; Introducing the SiLU activation function, the formula is as follows: In the formula, x is the input value, and sigmoid is the sigmoid activation function; An SPP network structure is constructed to perform multi-scale feature fusion. The input features are adaptively divided into blocks, and the features within each block are max-pooled to achieve the fusion of features at different scales.

2. The transformer partial discharge detection method based on the improved CSP-GCNet according to claim 1, characterized in that: Building an image classification database from the original atlas, specifically including: The Hilbert spectra of the acquired raw signals are scaled down to ensure that the spectra have a uniform size and resolution; and the scaled spectra data are divided into training set, test set and validation set in an 8:1:1 ratio to construct a spectra classification dataset. Graph preprocessing techniques, including cropping, rotation, flipping, and deformation operations, are employed to increase the diversity and richness of the graph classification dataset. The spectral data is filtered and cleaned, and classified and labeled according to specific standards to build an original spectral database.

3. The transformer partial discharge detection method based on the improved CSP-GCNet according to claim 1, characterized in that, The process involves training the model, validating the model, calculating the average accuracy, and completing the transformer partial discharge pattern recognition task. Specifically: During training, the optimized CSP-GCNet model is applied to feature extraction from the input map; After training, the CSP-GCNet model is evaluated using a validation set by calculating the average accuracy of the CSP-GCNet model on the validation set.

4. A transformer partial discharge detection device based on an improved lightweight cross-level local network CSP-GCNet for implementing the method as described in any one of claims 1-3, characterized in that, include: The signal acquisition unit is used to acquire the original partial discharge signal of the transformer, including discharge signals under normal and abnormal conditions; The signal processing unit is used to decompose the acquired raw signal into several intrinsic mode components with different center frequencies through variational mode decomposition (VMD-Hilbert post-processing), then perform Hilbert transform to obtain the Hilbert spectrum, and integrate the Hilbert spectrum over the entire time domain to obtain the Hilbert spectrum of the original signal. The image classification module is used to build an image classification database based on the original signal Hilbert spectrum obtained in the previous step. The partial discharge map classification module is used to build a model based on the improved CSP-GCNet network for transformer partial discharge map classification. The task recognition module is used to train and validate the model based on the established image classification database, evaluate the model's pattern recognition accuracy using average precision, and finally complete the transformer partial discharge pattern recognition task.

5. A storage medium, characterized in that, The storage medium stores an instruction set, wherein the instruction set, when executed by a processor, implements the transformer partial discharge detection method as described in any one of claims 1-3.

Images