Partial discharge image data enhancement method based on deep learning and electronic equipment

Abstract

The invention relates to a partial discharge image data enhancement method based on deep learning and electronic equipment. The method comprises the following steps: constructing a CAE-ACGAN model, wherein the CAE-ACGAN model comprises a generator adopting a convolutional auto-encoder architecture and a discriminator comprising an adversarial branch and a classification branch; training a CAE-ACGAN model: inputting a partial discharge real PRPD spectrogram and an image category label into a generator to learn the internal structure of the PRPD spectrogram, outputting a reconstructed image as the input of a backbone network in a discriminator, and outputting two branches for resistance judgment and auxiliary classification judgment respectively by the backbone network; the enhancement process comprises the following steps: inputting a minority class of real PRPD spectrograms and class labels thereof into a trained generator, and directionally generating new PRPD spectrograms belonging to the class; and repeating the enhancement process until the number of the PRPD spectrograms of the minority class is balanced with the number of the PRPD spectrograms of other classes, and obtaining an enhanced balanced data set of the class. Compared with the prior art, the method has the advantages of stably realizing high-quality and category-controllable partial discharge data enhancement and the like.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and electrical equipment fault diagnosis technology, and in particular to a method for enhancing partial discharge image data based on deep learning. Background Technology

[0002] Partial discharge is an important indicator for assessing the insulation status of high-voltage electrical equipment such as power transformers and gas-insulated switches (GIS). Different types of partial discharge defects (such as tip discharge, floating discharge, surface discharge, and bubble discharge) pose varying degrees of threat to equipment insulation. Therefore, accurate pattern recognition of partial discharge signals is a key step in achieving equipment status early warning and intelligent operation and maintenance.

[0003] With the development of artificial intelligence technology, deep learning-based methods, especially convolutional neural networks (CNNs), have shown great potential in the field of partial discharge pattern recognition. These methods can automatically extract high-dimensional features from partial discharge spectra (such as PRPD spectra) to achieve end-to-end classification, greatly reducing the burden of manual diagnosis. However, the performance of deep learning models is highly dependent on large-scale, high-quality, and class-balanced training datasets. In practical engineering, partial discharge is an intermittent fault, and the number of effective samples that can be collected on-site is very limited. Furthermore, the frequency of occurrence of different types of faults varies greatly, leading to two major challenges in datasets: "small sample size" and "class imbalance." This directly results in overfitting of the trained recognition model, poor generalization ability, and a severe bias towards the majority class with a large sample size. This can easily lead to missed or false positives for minority class faults, limiting the application effectiveness of deep learning technology in real-world scenarios.

[0004] To address this issue, data augmentation techniques have emerged. Traditional image augmentation methods, such as geometric transformations (rotation, flipping), can increase the amount of data but do not increase sample diversity, thus offering limited help in improving model generalization ability. Oversampling methods (such as the SMOTE algorithm) synthesize new samples by performing linear interpolation in the feature space, but they are prone to introducing noise and even blurring class boundaries. In recent years, deep generative models, represented by Generative Adversarial Networks (GANs), have provided new solutions for data augmentation. GANs, through adversarial games between the generator and discriminator, can learn the complex distribution of real data and generate highly realistic new samples. However, the training process of standard GANs is unstable and prone to mode collapse and gradient vanishing problems. Furthermore, unconditional GANs cannot control the class of generated samples, making them unsuitable for scenarios requiring targeted augmentation of specific minority classes.

[0005] Therefore, there is an urgent need for a data augmentation method that can be stably trained, generate high-quality data with controllable categories, in order to effectively solve the data bottleneck problem in partial discharge pattern recognition.

[0006] A search revealed Chinese invention patent application publication number CN115908842A, which discloses a method for augmenting and identifying partial discharge data in transformers. Using partial discharge fault type PRPD maps as the object, real samples are labeled and input into a CGAN model. The trained generator model exhibits stable training and can generate partial discharge map samples of specific types. The partial discharge maps generated using the improved CGAN model under various specific discharge types can expand the original discharge dataset, increasing its diversity and better addressing issues such as imbalance and small sample distribution in the original discharge dataset. However, this existing patent application suffers from problems such as inputting random noise into the generator for training, separating the discriminator's identification of true and false data from the classification process, leading to instability, uncontrollable categories, and inefficiency in data augmentation.

[0007] Achieving stable, efficient, and category-controllable partial discharge image data enhancement has become a technical problem that needs to be solved. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a method for enhancing partial discharge image data based on deep learning.

[0009] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a method for augmenting partial discharge image data based on deep learning is provided, the method comprising: constructing a CAE-ACGAN model: the CAE-ACGAN model includes a generator employing a convolutional autoencoder architecture, and a discriminator including adversarial branches and classification branches; Training the CAE-ACGAN model: Input the real PRPD spectra and image category labels of the partial discharge into the generator to learn the intrinsic structure of the PRPD spectra. The output reconstructed image is used as the input of the backbone network in the discriminator. The backbone network outputs two branches for adversarial judgment and auxiliary classification judgment, respectively. While judging the authenticity, it also learns the features that distinguish different categories of PRPD spectra. Augmentation process: Input a real PRPD spectrogram of the minority class and its class label into the trained generator to generate a new PRPD spectrogram belonging to that class; repeat the augmentation process until the number of minority class PRPD spectrograms is balanced with other classes, resulting in an augmented balanced dataset for that class.

[0010] Preferably, the generator includes a convolutional encoder, an intermediate layer, and a convolutional decoder connected in one pass; A convolutional encoder is used to encode the input real PRPD spectrogram and output a latent code; The convolutional encoder includes two downsampling modules, each containing a convolutional layer with a stride of 2, an instance normalization layer, and a LeakyReLU activation function. The intermediate layer inputs the latent code output by the convolutional encoder into multiple residual connection modules for deep feature transformation; The convolutional decoder includes two upsampling modules, each containing a transposed convolutional layer with a stride of 2, an instance normalization layer, and a ReLU activation function.

[0011] More preferably, the generator further includes label embedding, which converts the category label into a vector with the same number of latent code channels through the embedding layer, and concatenates it with the latent code in the channel dimension to achieve conditional generation.

[0012] Preferably, the adversarial branch is used to determine the authenticity of the PRPD spectrum, and the classification branch is used to assist in determining the partial discharge type to which the PRPD spectrum belongs.

[0013] More preferably, the features extracted from the backbone network in the adversarial branch and classification branch multiplexing discriminator are used.

[0014] More preferably, the classification branch of the discriminator, while judging the authenticity, also learns to distinguish the features of different categories of PRPD spectra, thereby guiding the generator to generate PRPD spectra with clear category features.

[0015] Preferably, the CAE-ACGAN model is trained using a composite loss function consisting of three parts: adversarial loss, reconstruction loss, and auxiliary classification loss. The adversarial loss is used to make the training process smoother, and includes the total loss of the generator and the total loss of the discriminator, both of which are calculated using the least squares loss method. The reconstruction loss constrains the generation process by calculating the difference between the image reconstructed by the generator and the original input image, and is calculated using the L1 norm. The auxiliary classification loss is used to achieve category-controllable generation, and includes the total auxiliary classification loss of the discriminator and the auxiliary classification loss of the generator.

[0016] More preferably, the total auxiliary classification loss of the discriminator is the sum of the cross-entropy loss between the predicted class and the true class label of the real image, and the cross-entropy loss between the predicted class and the true class label of the reconstructed image; the auxiliary classification loss of the generator is the cross-entropy loss between the predicted class and the true class label of the reconstructed image.

[0017] Preferably, the true PRPD spectrum of partial discharge is obtained by preprocessing the partial discharge file, including a two-dimensional histogram and a two-dimensional thermal map.

[0018] According to another aspect of the present invention, an electronic device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the method described thereon.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1) The CAE-ACGAN model of this invention combines the reconstruction capability of CAE and the adversarial generation capability of GAN, and introduces L1 reconstruction loss. The generated PRPD spectrum is highly similar to the real image in terms of structure and detail. The classification constraint introduced in the discriminator makes the generation process guided by the category label, and generates local spectral images of a specified type as needed. It is very suitable for targeted enhancement of the minority class in solving the class imbalance problem. Therefore, it achieves high-quality local spectral image enhancement with strong class controllability.

[0020] 2) The CAE-ACGAN model of this invention is trained using a composite loss function consisting of three parts: adversarial loss, reconstruction loss, and auxiliary classification loss. The adversarial loss uses the least squares loss method instead of the traditional GAN ​​cross-entropy loss, which effectively alleviates the gradient vanishing problem and makes the training process more stable and convergent. The reconstruction loss uses the L1 norm to ensure that the generator can faithfully learn and reproduce the structural information of the PRPD spectral map. The auxiliary classification loss uses cross-entropy loss, which forces the discriminator to become an effective classifier and at the same time drives the generator to produce images with distinct category features that can be correctly classified by the discriminator, so that the training process can converge quickly and stably. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall architecture of the CAE-ACGAN model proposed in this invention; Figure 2 This is a schematic diagram of the network structure of generator G in this invention; Figure 3 This is a schematic diagram of the network structure of the discriminator D in this invention; Figure 4 Examples of comparisons between different types of partial discharge PRPD spectra generated using the method of this invention and real spectra; Figure 5 This is a schematic flowchart of the partial discharge image data enhancement method in this invention. Detailed Implementation

[0022] 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.

[0023] Partial discharge (PRPD) is a low-probability event, and the collected PRPD datasets are often small in scale and unevenly distributed across classes. This can easily lead to deep learning classifiers biasing their identification results towards the majority class, resulting in false positives and false negatives. To address this issue, this invention proposes a stable, efficient, and class-controllable method for augmenting PRPD spectral data.

[0024] Example 1 This embodiment relates to a deep learning-based method for partial discharge image data augmentation. This method innovatively proposes a CAE-ACGAN model, using a convolutional autoencoder (CAE) as the generator G. Unlike traditional generative adversarial networks (GANs) that generate images from random noise, the generator G of this invention takes real PRPD spectrograms as input. This generator is not limited to generating new images, but learns the deep structure and core features of the PRPD spectrogram through an "encode-decode" reconstruction path.

[0025] CAE-ACGAN (Conditional Autoencoder with Auxiliary Classifier Generative Adversarial Network) model, such as Figure 1 It consists of a generator G and a discriminator D.

[0026] a) Generator G: such as Figure 2 The generator employs a convolutional autoencoder (CAE) architecture, taking a real PRPD spectrogram as input. Its internal convolutional encoder compresses the input image into a compact low-dimensional latent code through a series of convolutions with strides (downsampling), capturing the image's core structure and statistical features. The latent code is then processed through multiple deep residual blocks for deep feature transformation and refinement. Finally, the convolutional decoder reconstructs the processed latent code into a pseudo-image of the same size as the input image through a series of transposed convolutions (upsampling). This "reconstruction" generation path forces the generator to learn the intrinsic structure of the PRPD spectrogram, ensuring the realism and structural integrity of the generated samples.

[0027] b) Discriminator D: such as Figure 3The discriminator employs a multi-task convolutional neural network architecture, including a backbone network and two output branches. The discriminator receives an image (real or generated) as input, which is progressively processed by the backbone network to extract features and reduce the feature map size, resulting in two output branches. The first branch output is used for adversarial judgment, determining whether the image is "real" or "fake." The second output is used for classification, predicting the type of local discharge (PD) the image belongs to (e.g., corona discharge, levitation). Compared to using separate discriminators and independent classifiers, this fusion architecture offers several advantages: First, it achieves efficient feature sharing; the discriminator can reuse features extracted from the same backbone network when learning "real / fake" and "category" judgments, making the feature representation more robust. Second, the supervised classification task provides a stable gradient source for adversarial training, acting as a regularization mechanism and contributing to improved overall training stability. This design ensures that the discriminator, while judging real / fake images, must also learn the features that distinguish different categories of PRPD spectral maps, thus guiding the generator to generate images with clearly defined category features.

[0028] To address the problem of unstable model training and ensure generation quality, this invention employs a composite loss function consisting of three parts: adversarial loss, reconstruction loss, and auxiliary classification loss.

[0029] Adversarial Loss: To address the instability issue in standard GAN training, this invention employs Least Squares Loss (LSGAN Loss). Compared to traditional cross-entropy loss, the penalty function of LSGAN is smoother far from the decision boundary, providing a more stable gradient signal to the generator, effectively avoiding gradient vanishing, and making the training process smoother. Its mathematical description is: in, The total loss of discriminator D, The total loss of generator G, Let be the expected value, representing the average of the loss within the parentheses over the entire data distribution. For the discriminator to analyze the input image True or false prediction tags To reconstruct the image The discriminator predicts the authenticity of the reconstructed image using labels.

[0030] Reconstruction Loss: This loss term is the core of the CAE structure of this invention. It calculates the image reconstructed by the generator. Compared with the original input image xThe generation process is constrained by the differences between the pixels. The reconstruction loss of this invention uses the L1 norm (i.e., the sum of absolute differences pixel by pixel) because it produces sharper, less blurred images compared to the L2 norm. This loss ensures that the generator faithfully learns and reproduces the structural information of the PRPD spectral map. Its mathematical description is: in, To reconstruct the loss, x The original input image is shown here as the true PRPD spectrogram. For generator G based on x The reconstructed image, This represents the expected value for the training data.

[0031] Auxiliary Classification Loss: This loss term is used to achieve class-controlled generation. It is generated by the classification branch of the discriminator and calculates the cross-entropy loss between the predicted and true class labels of the real and generated images. This loss forces the discriminator to become an effective classifier and simultaneously drives the generator to produce images with distinct class features that can be correctly classified by the discriminator. Its mathematical description is: in, The total auxiliary classification loss for discriminator D; The auxiliary classification loss for the generator G; CE (⋅,⋅) is the cross-entropy loss function; C (⋅) represents the label output by the discriminator's classification head; y For real category labels; To reconstruct image data; For the expected value of the training data, , x The original input image is the real image.

[0032] Finally, based on the above two design aspects, the model can achieve on-demand generation. After the model training is completed, a real image of a specific minority class and its class label are provided to the generator to generate new samples with diversity belonging to that class, thereby expanding and balancing the training set of partial-amplitude pattern recognition, and thus significantly improving the performance of the recognition model.

[0033] This embodiment also relates to a deep learning-based method for augmenting partial discharge image data, which augments a PRPD image dataset containing four types of partial discharge (corona, floating, inner, and particles) to improve the performance of subsequent classification models.

[0034] This method is as follows Figure 5 This includes the following steps: Step 1: Data preparation and preprocessing, including 11) Obtaining the data source: The raw data is a partial discharge .dat file collected on-site or in the laboratory, which contains pulse amplitude and phase information.

[0035] 12) PRPD Spectrum Generation: For each partial discharge .dat file, the pulse data is mapped to a two-dimensional "phase-amplitude" grid. Specifically, the phase angle ϕ is folded back to the range of [0, 360) degrees, and the pulse amplitude q is non-negatively processed. The pulse count value within each grid point is accumulated on the two-dimensional grid to form the PRPD spectrum, including a two-dimensional histogram and a two-dimensional heatmap. ,in : fell in the j Phase intervals , No. k Amplitude range The total number of pulses within the range is divided into 12 intervals of 30 degrees in the range [0, 360), corresponding to j = 0, 1, ..., 11.

[0036] 13) Standardization and Rendering: To eliminate domain offset caused by factors such as device gain, the pulse amplitude is calibrated before generating the histogram. The generated 2D histogram is rendered as a PNG image using a fixed color map, and irrelevant visual elements such as axes and scales are removed to prevent the model from learning "shortcut" features. All images are uniformly resized to a fixed size (e.g., 128x128 pixels) and normalized using the mean and standard deviation of the ImageNet dataset.

[0037] Step 2: Model architecture construction, proposing a CAE-ACGAN model, such as... Figure 1 The model consists of a generator G and a discriminator D.

[0038] The generator G employs a convolutional autoencoder (CAE) architecture, such as... Figure 2 ,include: The convolutional encoder consists of two downsampling modules that encode the input PRPD spectral map and output a latent code. Each downsampling module contains a convolutional layer with a stride of 2 (halving the feature map size and doubling the number of channels), an instance normalization layer, and a LeakyReLU activation function.

[0039] Intermediate layer: The latent code output by the convolutional encoder is input into 6 residual connection modules for deep feature transformation.

[0040] The convolutional decoder consists of two upsampling modules, and its structure is symmetrical to that of the convolutional encoder. Each upsampling module contains a transposed convolutional layer with a stride of 2 (which doubles the feature map size and halves the number of channels), an instance normalization layer, and a ReLU activation function.

[0041] Label embedding: The category labels (such as one-hot vectors) are converted into vectors with the same number of latent code channels through the embedding layer, and concatenated or added to the latent code in the channel dimension to achieve conditional generation.

[0042] 22) Discriminator D (introducing auxiliary classification AC), such as Figure 3 ,include: The backbone network consists of a series of convolutional downsampling modules that progressively extract features and reduce the size of the feature map.

[0043] Output branches: At the end of the network, two parallel branches are generated using 1x1 convolution: an adversarial branch and a classification branch.

[0044] Adversarial branch: Outputs a scalar value used to determine the authenticity of the input image.

[0045] Classification branch: After global average pooling (GAP), the output is a vector with dimensions equal to the number of classes, used to predict the image's class. The entire discriminator does not use fully connected layers to reduce the number of parameters.

[0046] Step 3: Model Training Loss function: Define the total loss function, which includes LSGAN adversarial loss and L1 reconstruction loss (weights). λ rec Set to 50) and cross-entropy auxiliary classification loss.

[0047] Optimizer: Configure the Adam optimizer for the generator (G) and discriminator (D) respectively, with a learning rate of 2×10−4 and momentum parameters β1=0.5, β2=0.999.

[0048] Training process: In each training iteration, the discriminator and generator are updated, including: Update the discriminator D: Take a batch of real images x and their labels y, and let the generator reconstruct x to obtain G(x). Calculate the total loss of the discriminator on the real images and the reconstructed images (adversarial loss + classification loss), and perform a backpropagation to update the parameters of the discriminator.

[0049] Update generator G: Using a batch of real images x and their labels y, calculate the generator's total loss (adversarial loss + reconstruction loss + classification loss) and perform a backpropagation to update the generator's parameters.

[0050] Training monitoring: During the training process, each loss component is monitored in real time. L AD , L AG , L rec , L CD , L CG Changes in ) . A healthy training state is typically characterized by: Combat losses: L AD and L AG It tends to an equilibrium point (theoretically about 0.25 and 0.125 for least squares loss LSGAN). Reconstruction loss: L rec A steady decrease and convergence to a low value (e.g., 0.03–0.08) indicates that the structure has been sufficiently learned; Auxiliary classification loss: L CD and L CG A steady decline indicates that the classification task learning is effective.

[0051] Step 4: Data Augmentation and Performance Evaluation Generating new samples: After model training converges, the parameters of the generator G are fixed. For classes with fewer samples (e.g., floating), providing the class label and any real image of that class as input will generate a new PRPD spectrogram belonging to that class but with subtle differences from the input image. This process is repeated until the minority class sample size is balanced with the other classes, resulting in an augmented, balanced dataset. Figure 4 As shown, the left image is the input real partial discharge PRPD spectrum, and the right image is the partial discharge PRPD spectrum reconstructed by the generator of this invention. Comparative analysis shows that the reconstructed image is very close to the real data. This is a high-fidelity reconstruction result, indicating that the generator successfully captures the inherent manifold and core semantic features of the PRPD spectrum. The reconstructed image and the original image maintain a high degree of consistency in global structure, including the spatial location, shape, and density of all major pulse clusters, proving that the model effectively learns the low-frequency components representing the partial discharge mode. Although the reconstructed image inevitably exhibits loss of high-frequency details and slight diffusion at the pixel level, this result demonstrates that this invention successfully utilizes reconstruction as a self-supervised surrogate task, forcing the encoder to learn a structurally sound and semantically rich latent code space.

[0052] Performance Evaluation: A standalone partial-amplitude pattern recognition model (e.g., ResNet) was trained using the enhanced balanced dataset. This model was then evaluated on a separate external validation set (original_prpd). The evaluation showed that a stable and high average accuracy (e.g., 82.78%) was achieved even on a cross-domain external validation set, demonstrating that the samples generated by this invention possess good diversity and generalization ability, effectively improving the robustness of the recognition system.

[0053] After using the pseudo-samples generated by this method to augment the imbalanced original dataset, the trained partial discharge pattern recognition model showed significant improvements in various evaluation metrics (such as accuracy and F1 score), verifying the effectiveness and practicality of the present invention.

[0054] 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.

[0055] 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.

[0056] 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).

[0057] 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.

[0058] 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.

[0059] 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.

[0060] 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 deep learning-based partial discharge image data augmentation method, characterized in that, The method comprises: constructing a CAE-ACGAN model; the CAE-ACGAN model comprises a generator adopting a convolutional autoencoder architecture and a discriminator comprising an adversarial branch and a classification branch; training the CAE-ACGAN model: inputting a real PRPD spectrum and an image category label of partial discharge into the generator to learn the internal structure of the PRPD spectrum, outputting a reconstructed image as an input of a backbone network in the discriminator, and outputting two branches of the backbone network for adversarial judgment and auxiliary classification judgment, respectively, to judge authenticity and learn features for distinguishing PRPD spectrums of different categories; an enhancement process: inputting a real PRPD spectrum of a minority category and a category label thereof into the trained generator to directionally generate a new PRPD spectrum belonging to the category; and repeating the enhancement process until the number of PRPD spectrums of the minority category and the number of PRPD spectrums of other categories are balanced, thereby obtaining a balanced data set of the enhanced category.

2. The partial discharge image data augmentation method based on deep learning according to claim 1, characterized in that, The generator comprises a convolutional encoder connected once, an intermediate layer and a convolutional decoder; the convolutional encoder is used for encoding an input real PRPD spectrum and outputting a latent code; the convolutional encoder comprises two down-sampling modules, and each down-sampling module comprises a convolutional layer with a step of 2, an instance normalization layer and a LeakyReLU activation function; the intermediate layer inputs the latent code output by the convolutional encoder into a plurality of residual connection modules for deep feature transformation; the convolutional decoder comprises two up-sampling modules, and each up-sampling module comprises a transposed convolutional layer with a step of 2, an instance normalization layer and a ReLU activation function.

3. The partial discharge image data augmentation method based on deep learning according to claim 2, characterized in that, The generator further comprises a label embedding, which converts a category label into a vector with the same number of channels as the latent code through an embedding layer, and splices the vector with the latent code in the channel dimension to realize conditional generation.

4. The partial discharge image data augmentation method based on deep learning according to claim 1, characterized in that, The adversarial branch is used for judging authenticity of the PRPD spectrum, and the classification branch is used for auxiliary judgment of a partial discharge type to which the PRPD spectrum belongs.

5. The partial discharge image data augmentation method based on deep learning according to claim 4, characterized in that, The adversarial branch and the classification branch reuse features extracted by a backbone network in the discriminator.

6. The partial discharge image data augmentation method based on deep learning according to claim 5, characterized in that, The classification branch of the discriminator learns features for distinguishing PRPD spectrums of different categories while judging authenticity, thereby reversely guiding the generator to generate PRPD spectrums with clear category features.

7. The partial discharge image data augmentation method based on deep learning according to claim 1, characterized in that, The CAE-ACGAN model adopts a composite loss function composed of an adversarial loss, a reconstruction loss and an auxiliary classification loss for training; the adversarial loss is used for making a training process more stable, and comprises a total loss of the generator and a total loss of the discriminator, both of which are calculated by using a least square loss method; the reconstruction loss is used for restraining a generation process by calculating a difference between a reconstructed image and an original input image, and is calculated by using an L1 norm; the auxiliary classification loss is used for realizing category-controllable generation, and comprises a total auxiliary classification loss of the discriminator and an auxiliary classification loss of the generator.

8. The partial discharge image data augmentation method based on deep learning according to claim 7, characterized in that, the total auxiliary classification loss of the discriminator is a sum of a cross-entropy loss between a predicted category of a real image and a real category label and a cross-entropy loss between a predicted category of a reconstructed image and the real category label; and the auxiliary classification loss of the generator is a cross-entropy loss between the predicted category of the reconstructed image and the real category label.

9. The partial discharge image data augmentation method based on deep learning according to claim 1, wherein, The real PRPD spectrum of partial discharge is obtained by preprocessing the partial discharge file, including a two-dimensional histogram and a two-dimensional heat map.

10. An electronic device comprising a memory and a processor, said memory having stored thereon a computer program, characterized in that, The processor implements the method of any one of claims 1-9 when executing the program.

Images