Partial discharge mode identification method based on transfer learning

Abstract

The invention relates to a partial discharge mode recognition method based on transfer learning. The method comprises the following steps: preprocessing original partial discharge pulse data into a PRPD spectrogram or a time-frequency image; the method comprises the following steps: constructing an EfficientNet-ECA model, embedding an ECA module into the model behind each convolution block of an EfficientNet-B0 network, dynamically carrying out channel dimension re-calibration on a feature map extracted by the EfficientNet-B0 network through the ECA module, and calculating an attention weight so as to re-scale the input feature map; migration training: adopting parameters of an ImageNet data set pre-training model to initialize network parameters of an OfficientNet-ECA model, and performing fine tuning through a hierarchical learning rate strategy; and inputting the PRPD spectrogram or the time-frequency image to be identified into the migrated and trained OfficientNet-ECA model, and outputting the partial discharge type to which the PRPD spectrogram or the time-frequency image belongs. Compared with the prior art, the method has the advantage of efficiently realizing accurate identification of the partial discharge mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence and electrical equipment fault diagnosis, in particular to a partial discharge pattern recognition method based on transfer learning. BACKGROUND

[0002] Partial discharge is an important indicator of insulation deterioration of high-voltage electrical equipment, and its pattern recognition is of great significance to the evaluation of the running state of the equipment. Traditional recognition methods rely on manual feature extraction and classifier design, which have the problems of strong subjectivity and poor adaptability. With the rapid development of artificial intelligence technologies such as deep learning and knowledge graph, data-driven pattern recognition has changed the state monitoring of high-voltage equipment. In recent years, although the recognition method based on deep learning has made progress, its performance is highly dependent on large-scale labeled data. For example: the deep learning technology based on traditional fully connected convolutional neural network can tap the value of massive multi-source heterogeneous electrical data, can realize noise suppression and data enhancement of partial discharge signals, and can effectively identify different types of partial discharge signals, but in actual engineering, the number of partial discharge samples is small and the classes are unbalanced, which leads to easy overfitting of the model and poor generalization ability.

[0003] In the prior art, although some researchers have tried to use data augmentation, generative adversarial networks (GAN) and other methods to expand the samples, there are still problems such as unstable quality of generated samples, difficulty in class control, etc. Therefore, although the data quantity can be expanded, the diversity of samples is not increased, which is of limited help to improve the generalization ability of the model. In summary of the above points, partial discharge pattern recognition is a typical small sample learning problem in the power system. In the field of artificial intelligence and deep learning, in the face of the challenge of small sample learning with scarce samples and high labeling cost, transfer learning is one of the mainstream solutions. Its main idea is to transfer labeled data or knowledge structure from related fields to complete or improve the learning effect of the target field or task. By transferring the knowledge of the source domain (Source Domain) to the target domain, transfer learning breaks the assumption of "training-test data independent and identically distributed" in traditional machine learning, and has obvious advantages in small sample learning conditions. However, although transfer learning can improve the small sample learning ability with the help of pre-trained models, it is easy to have negative transfer when the domain difference is large and the feature distribution is inconsistent, which affects the recognition accuracy and makes parameter tuning difficult, and there are still challenges in industrial applications. Therefore, there is an urgent need for a method that can stably and efficiently realize partial discharge pattern recognition under the condition of a small number of samples.

[0004] A search revealed Chinese invention patent application publication number CN118484711A, which discloses a partial discharge pattern recognition method based on deep transfer learning networks and model fusion. The method includes: S1: acquiring time-domain waveform data S(t) of partial discharge under different defect types using a UHF sensor; S2: extracting corresponding numerical time-series feature parameter vectors; S3: constructing a deep transfer learning residual shrinking network model; S4: constructing a one-dimensional improved deep residual network model; S5: using the original time-series signal data and feature parameter vectors of partial discharge as inputs to the deep transfer learning residual shrinking network model and the one-dimensional improved deep residual network model, respectively; using the Adam algorithm as the optimization algorithm for the partial discharge (PD) classification model; training to obtain the optimal parameters of the PD classification model; and using the Softmax function to output the partial discharge type judgment result. This existing patent application suffers from inaccurate partial discharge pattern recognition.

[0005] How to achieve efficient and accurate partial discharge pattern recognition has become a technical problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a partial discharge pattern recognition method based on transfer learning.

[0007] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a partial discharge pattern recognition method based on transfer learning is provided, the method comprising: Data preprocessing: The raw partial discharge pulse data is preprocessed into PRPD spectra or time-frequency images; Construct an EfficientNet-ECA model, in which an ECA module is embedded after each convolutional block of the EfficientNet-B0 network; Transfer training: The parameters of the EfficientNet-ECA model are initialized using the parameters of the pre-trained model on the ImageNet dataset, and fine-tuned using a hierarchical learning rate strategy; The PRPD spectrum or time-frequency image to be identified is input into the transfer-trained EfficientNet-ECA model. The ECA module dynamically recalibrates the channel dimension of the feature map extracted by the EfficientNet-B0 network, calculates attention weights to rescale the input feature map, and finally outputs the corresponding partial discharge type.

[0008] As a preferred technical solution, the ECA module operates on the channel dimension of the input feature map X, including: Global average pooling is applied to aggregate spatial information and generate channel descriptors; Capture local cross-channel interactions through fast one-dimensional convolutions with adaptive kernel size; The original input feature map X is rescaled using the calculated attention weights to obtain the output of the ECA module.

[0009] As a preferred technical solution, the adaptive kernel size is adaptively determined by the channel dimension.

[0010] As a preferred technical solution, the adaptive kernel size is specifically expressed as follows: , in, k For the size of the core, Ψ This represents the mapping function, where C is the channel dimension. This represents the odd-number operator, which takes the nearest odd integer. Indicates the scaling factor. Indicates bias.

[0011] As a preferred technical solution, the attention weights are calculated using the sigmoid activation function.

[0012] As a preferred technical solution, the transfer training includes: freezing the entire backbone network in the initial stage and training only the classification head; and unfreezing the deeper parts of the backbone network in the subsequent stage and performing end-to-end training using a decreasing learning rate.

[0013] As a preferred technical solution, the transfer training adopts a joint loss function, including source classification loss, domain difference loss and target supervision loss.

[0014] As a preferred technical solution, the joint loss function is specifically expressed as follows: , in, For source domain samples; The true labels of the source domain samples; For the target domain sample; The true labels of the target domain samples; These represent the feature distributions of the source and target domains, respectively. L CE (⋅,⋅) is the cross-entropy loss function. Representative source classification loss; The representative domain difference loss, here referring to the maximum mean difference, is used to calculate the feature distribution of the source domain. and target domain feature distribution The differences between them; Representative target monitoring loss; , Represents the transfer weight coefficient. Controlling the intensity of source knowledge transfer, The strength of the control domain alignment, 1- α - β Used to control the strength of the monitoring signal in the target domain; and These represent the predicted outputs of the source domain samples and the target domain samples after passing through the model, respectively.

[0015] As a preferred technical solution, the EfficientNet-B0 network systematically scales the network's depth, width, and resolution using a composite scaling method.

[0016] As a preferred technical solution, the output partial discharge type includes corona discharge, suspension discharge, internal discharge, and particle discharge.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1) This invention proposes a lightweight EfficientNet-ECA model that integrates an efficient channel attention mechanism for cases with a small number of partial discharge samples. An ECA module is embedded after each convolutional block of the EfficientNet-B0 network. The model initializes the network model with parameters from a model pre-trained on the ImageNet dataset. The ECA module dynamically recalibrates the channel dimensions of the feature map extracted by the EfficientNet-B0 network, calculates attention weights to rescale the input feature map, and finally outputs the type of partial discharge. Through transfer learning and the lightweight EfficientNet-ECA model, efficient and accurate identification of partial discharge patterns is achieved.

[0018] 2) In this invention, the ECA module captures local cross-channel interactions through fast one-dimensional convolution with adaptive kernel size, ensuring that the coverage of cross-channel interactions is proportional to the channel dimension. The original input feature map is rescaled using the calculated attention weights to obtain the output of the ECA module, allowing the model to dynamically recalibrate the channel feature response with minimal computational overhead, thereby significantly improving its ability to identify subtle discharge patterns in noise. Therefore, the ECA module is deeply embedded and integrated with the EfficientNet-B0 network, which can improve the efficiency and accuracy of partial discharge patterns. Compared with traditional fully connected neural network methods, the accuracy is improved by about 10%.

[0019] 3) This invention addresses the partial discharge scenario with a small number of target samples by designing a joint loss function that comprehensively considers source classification loss, domain difference loss, and target supervision loss. This function can optimize the model's classification performance and feature distribution alignment in the target domain, better utilize the general features and network parameters in the pre-trained model of a large-scale natural image dataset, and quickly and stably train a model for partial discharge model recognition, thereby achieving fast and accurate pattern recognition. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the partial discharge pattern recognition method based on transfer learning according to the present invention. Figure 2 This is a schematic diagram of the EfficientNet-ECA model structure in this invention; Figure 3 This is an example of a PRPD spectrum in one embodiment of the present invention; Figure 4 This is a visualization of clustering diagrams of features extracted by the model in one embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] This invention aims to address the problem of poor model generalization ability caused by a small number of samples and class imbalance in partial discharge identification. It proposes a lightweight convolutional neural network model that integrates an efficient channel attention mechanism and extracts general features from a pre-trained model on a large-scale natural image dataset through transfer learning. This model is suitable for the identification task of partial discharge PRPD spectrum in insulation condition diagnosis and fault early warning of high-voltage equipment such as power transformers and GIS.

[0023] Example 1 This embodiment relates to a partial discharge pattern recognition method based on transfer learning, which includes four stages: data preprocessing, model building, transfer training and classification recognition. It can achieve high-precision and high-robust partial discharge pattern recognition under conditions of a small number of samples, providing technical support for intelligent operation and maintenance of power equipment.

[0024] This invention improves the model's feature extraction and classification capabilities under limited sample conditions by introducing an efficient channel attention mechanism and a hierarchical fine-tuning strategy. For example... Figure 1 The technical solution of the present invention includes the following steps: S1, Data Preprocessing: Convert the original partial discharge pulse data into PRPD spectra or time-frequency images, and perform denoising, phase alignment, color mapping and size normalization to generate image data suitable for input to convolutional neural networks.

[0025] S2, EfficientNet-ECA model construction: as follows Figure 2This invention builds a lightweight convolutional neural network model based on EfficientNet-B0, incorporating efficient channel attention (ECA). The EfficientNet-B0 network achieves an excellent balance between computational efficiency and performance through compound scaling, making it highly suitable as a base model for resource-constrained industrial deployments, such as partial discharge pattern recognition. However, when directly applying the EfficientNet-B0 network to partial discharge PRPD spectral recognition, its inherent channel weighting mechanism limits its ability to distinguish subtle features of partial discharges. Under conditions of scarce samples, this deficiency is amplified, making it difficult for the model to focus on the most discriminative features. To address these shortcomings, the innovation of this invention is specifically reflected in: This invention proposes a deep embedded integration method for EfficientNet-B0 and Efficient Channel Attention (ECA): An Efficient Channel Attention (ECA) module is innovatively embedded after each convolutional block of the EfficientNet-B0 network. The ECA module dynamically calibrates channel feature responses through a lightweight, dimensionality-reduction-free attention mechanism. This design ensures that the output feature map can be immediately recalibrated in channel dimensions after each critical stage of feature extraction in the base network. This allows the model to "actively" focus on more informative channel features and suppress unimportant channel features, achieving progressive and refined optimization of channel attention.

[0026] Based on a deep embedded integration method, a novel network structure (EfficientNet-ECA) specifically designed for partial discharge spectrum recognition is constructed: by precisely embedding the ECA module into each convolutional block, this invention cleverly combines EfficientNet-B0 with the ECA module, thereby forming a new, lightweight network structure, EfficientNet-ECA, which is sensitive to channel features. This structure is specifically designed to capture discharge features with low signal-to-noise ratio and subtle patterns in PRPD spectra.

[0027] EfficientNet-B0: A convolutional neural network renowned for its parameter efficiency and high performance. This network systematically scales its depth, width, and resolution using a compound scaling method to maximize model accuracy within a fixed resource budget. Its optimal configuration is determined by a compound coefficient φ that uniformly scales the three dimensions. , , , , , Where d, w, and r represent network depth, width, and resolution, respectively. , , These are constants determined by small-grid search. This results in a range of models that are not only more accurate than previous convolutional neural networks, but also smaller and faster, making them particularly suitable for deployment scenarios with computational constraints.

[0028] The ECA module, or Efficient Channel Attention (ECA) module, aims to improve the network's representational power by emphasizing informative features and suppressing less useful features through a lightweight, parameter-free attention mechanism. It operates on the channel dimension of the input feature map X, first applying Global Average Pooling (GAP) to aggregate spatial information and generate channel descriptors z. c Its mathematical description is as follows: , in, H Represents the height of the feature map. W The width represents the feature map. Representing the The feature map of each channel in spatial location The value at that location.

[0029] Subsequently, ECA does not use a dimension-reduced fully connected layer, but instead uses a kernel size of k Fast one-dimensional convolutions capture local cross-channel interactions, where the kernel size... k From the channel dimension C Adaptive determination: , here, Ψ Represents a mapping function. This represents the odd-number operator, which takes the nearest odd integer. Indicates the scaling factor; Indicates bias.

[0030] This adaptive kernel selection ensures that the coverage of cross-channel interactions is proportional to the channel dimension. The final attention weights for each channel... It is calculated using the sigmoid activation function: , in, Represents one-dimensional convolution. It is the sigmoid activation function.

[0031] Then, by using the calculated attention weights The original input feature map is rescaled to obtain the output of the ECA module. This process allows the model to dynamically recalibrate the channel feature responses with minimal computational overhead, thereby significantly improving its ability to identify subtle discharge patterns in noise.

[0032] S3, Transfer Learning: Initialize network parameters using parameters from a pre-trained model on the ImageNet dataset, and fine-tune them using a hierarchical learning rate strategy: initially, the parameters are frozen as follows... Figure 2 The entire backbone network, including the main structure of the EfficientNet-B0 network and all its embedded ECA modules, is trained only on the classification head to avoid overfitting in the early stages and accelerate training; in subsequent stages, the deeper parts of the backbone network are unfrozen, i.e. Figure 2 The EfficientNet-B0 network utilizes convolutional modules (i.e., the set of convolutional modules close to the output) and their corresponding ECA modules in the backend to adapt to and transfer high-level features. End-to-end training with a decreasing learning rate is employed for fine-tuning parameters, avoiding the destruction of learned useful features. The EfficientNet-ECA model is lightweight, training stably, and suitable for practical engineering deployment. This method eliminates the need for manual feature engineering, enabling automated end-to-end recognition and improving diagnostic efficiency and objectivity.

[0033] Loss Function Design: This invention combines cross-entropy classification loss and maximum mean difference domain adaptation loss to optimize the model's classification performance and feature distribution alignment capability in the target domain. Addressing the core issue of how to simultaneously utilize source domain knowledge and perform domain alignment with a small number of target samples, this invention innovatively designs a specific loss function structure and its weight coefficients. This allows it to work in conjunction with the phased, hierarchical learning rate fine-tuning strategy in transfer training, achieving a tailored solution for the "partial discharge pattern recognition" scenario.

[0034] The joint loss function for transfer learning combines the source supervision loss and the domain alignment loss, and its mathematical description is as follows: , in, For source domain samples; The true labels of the source domain samples; For the target domain sample; The true labels of the target domain samples; These represent the feature distributions of the source and target domains, respectively. L CE (⋅,⋅) is the cross-entropy loss function. Representative source classification loss; The representative domain difference loss, here referring to the maximum mean difference, is used to calculate the feature distribution of the source domain. and target domain feature distribution The differences between them; Representative target monitoring loss; , Represents the transfer weight coefficient ( Controlling the intensity of source knowledge transfer, The strength of the control domain alignment, 1- α - β (used to control the strength of the target domain monitoring signal); and These represent the predicted outputs of the source domain samples and the target domain samples after passing through the model, respectively.

[0035] S5, Classification and Recognition: Input the PRPD spectrum to be identified into the trained EfficientNet-ECA model and output its local discharge type (such as corona discharge, suspension discharge, internal discharge, particle discharge, etc.).

[0036] Example 2 This embodiment also relates to a partial discharge pattern recognition method based on transfer learning. Taking the PRPD spectrum recognition of four typical partial discharge types (corona discharge, suspension, internal discharge, and particulate discharge) as an example, the implementation process of the method of the present invention is explained.

[0037] Step 1: Data Preparation Partial discharge pulse data are collected from the field or laboratory to generate PRPD spectra, such as... Figure 3 As shown, the image was denoised, cropped, and resized to 224×224 pixels, and then normalized using ImageNet mean and standard deviation.

[0038] Step 2: Building the EfficientNet-ECA model The EfficientNet-B0 network is used as the backbone, with an ECA module embedded after each convolutional block. The classification head uses global average pooling and fully connected layers to output four class probabilities.

[0039] Step 3: Transfer Training Using the Adam optimizer, the initial learning rate is 10. -4 The backbone network learning rate is 5×10 -5 The learning rate for the classification head is 10. -3 Training for 100 rounds, with an early stop mechanism patience of 15.

[0040] Step 4: Performance Evaluation Using 5-fold cross-validation, the model achieved an average accuracy of 96.67% on the test set, significantly outperforming traditional fully connected network methods.

[0041] Step 5: Practical Application The trained model is deployed in a power equipment condition monitoring system to achieve real-time, automatic identification of partial discharge types, such as... Figure 4 As shown, the t-SNE (t-Distributed Stochastic Neighbor Embedding) algorithm, based on nonlinear dimensionality reduction technology, is mainly used to map high-dimensional data to a low-dimensional space. The feature points of various partial discharges show obvious clustering differences in the two-dimensional space, demonstrating that after the t-SNE algorithm reduces the dimensionality of partial discharge features, the features of different discharge types have good distinguishability in the low-dimensional space, and can intuitively show the distribution characteristics and category boundaries of various partial discharge features.

[0042] Example 3 The electronic device of this invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM may also store various programs and data required for device operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0043] Multiple components in the device are connected to the I / O interface, including: input units such as keyboards and mice; output units such as various types of displays and speakers; storage units such as disks and optical discs; and communication units such as network interface cards (NICs), modems, and wireless transceivers. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0044] The processing unit performs the various methods and processes described above. For example, in some embodiments, the methods may be implemented as computer software programs tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program may be loaded and / or installed on the device via ROM and / or a communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the CPU may be configured to execute the methods by any other suitable means (e.g., by means of firmware).

[0045] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0046] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0047] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0048] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A partial discharge pattern recognition method based on transfer learning, characterized in that, The method includes: Data preprocessing: The raw partial discharge pulse data is preprocessed into PRPD spectra or time-frequency images; Construct an EfficientNet-ECA model in which an ECA module is embedded after each convolutional block of the EfficientNet-B0 network; Transfer training: The parameters of the EfficientNet-ECA model are initialized using the parameters of the pre-trained model on the ImageNet dataset, and fine-tuned using a hierarchical learning rate strategy; The PRPD spectrum or time-frequency image to be identified is input into the transfer-trained EfficientNet-ECA model. The ECA module dynamically recalibrates the channel dimension of the feature map extracted by the EfficientNet-B0 network, calculates attention weights to rescale the input feature map, and finally outputs the corresponding partial discharge type.

2. The partial discharge pattern recognition method based on transfer learning according to claim 1, characterized in that, The ECA module operates on the channel dimension of the input feature map X, including: Global average pooling is applied to aggregate spatial information and generate channel descriptors; Capture local cross-channel interactions through fast one-dimensional convolutions with adaptive kernel size; The original input feature map X is rescaled using the calculated attention weights to obtain the output of the ECA module.

3. The partial discharge pattern recognition method based on transfer learning according to claim 2, characterized in that, The adaptive kernel size is adaptively determined by the channel dimension.

4. The partial discharge pattern recognition method based on transfer learning according to claim 3, characterized in that, The adaptive kernel size is specifically expressed as follows: ， in, k For the size of the core, ψ This represents the mapping function, where C is the channel dimension. This represents the odd-number operator, which takes the nearest odd integer. Indicates the scaling factor. Indicates bias.

5. The partial discharge pattern recognition method based on transfer learning according to claim 2, characterized in that, Attention weights are calculated using the sigmoid activation function.

6. The partial discharge pattern recognition method based on transfer learning according to claim 1, characterized in that, The transfer training includes: in the initial stage, freezing the entire backbone network and training only the classification head; in the subsequent stage, unfreezing the deeper parts of the backbone network and performing end-to-end training using a decreasing learning rate.

7. The partial discharge pattern recognition method based on transfer learning according to claim 1, characterized in that, The transfer training described uses a joint loss function, which includes source classification loss, domain difference loss, and target supervision loss.

8. The partial discharge pattern recognition method based on transfer learning according to claim 7, characterized in that, The joint loss function is specifically expressed as follows: ， in, For source domain samples; The true labels of the source domain samples; For the target domain samples; The true labels of the target domain samples; These represent the feature distributions of the source and target domains, respectively. L CE (⋅,⋅) is the cross-entropy loss function. Representative source classification loss; The representative domain difference loss, here referring to the maximum mean difference, is used to calculate the feature distribution of the source domain. and target domain feature distribution The differences between them; Representative target monitoring loss; , Represents the transfer weight coefficient. Controlling the intensity of source knowledge transfer, The strength of the control domain alignment, 1- α - β Used to control the strength of the monitoring signal in the target domain; and These represent the predicted outputs of the source domain samples and the target domain samples after passing through the model, respectively.

9. The partial discharge pattern recognition method based on transfer learning according to claim 1, characterized in that, The EfficientNet-B0 network described above systematically scales the network's depth, width, and resolution using a composite scaling method.

10. The partial discharge pattern recognition method based on transfer learning according to claim 1, characterized in that, The output partial discharge types include corona discharge, suspension discharge, internal discharge, and particle discharge.

Images