Switch cabinet fault diagnosis method and device based on deep learning

Abstract

The invention discloses a switch cabinet fault diagnosis method and device based on deep learning, and belongs to the technical field of power equipment. The method comprises the following steps: S1, acquiring PD data; s2, pre-processing the PD data; the preprocessing content comprises the following steps: performing global statistical feature extraction on the PD data to obtain global statistical features; carrying out data type conversion on the PD data to obtain a grayscale image; s3, processing the grayscale image by adopting a ResNet model combined with a CBAM attention mechanism, and extracting key potential features as key local discrimination features; s4, performing feature splicing on the global statistical features and the key local discrimination features to obtain multi-scale features; and S5, performing classification processing by using an MLP classifier to obtain a switch cabinet fault diagnosis result. According to the method, signal features can be analyzed from different levels and scales through multi-scale feature fusion, and the fault diagnosis precision and adaptability are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power equipment, and specifically relates to a switch cabinet fault diagnosis method and device based on deep learning. Background Art

[0002] In the process of power generation, transmission, distribution and energy conversion in the power system, the main function of the switchgear is to open and close, control and protect the electrical equipment, which directly affects the safety and stability of the high-voltage transmission process. However, in the process of switchgear from production to transportation, assembly, operation to aging, some minor physical defects are inevitable, which can easily lead to the distortion of the field strength of the power transmission inside the switchgear equipment and the occurrence of partial discharge (PD). As the discharge phenomenon gradually worsens, it will seriously affect the stability of power transmission and even cause large-scale power outages.

[0003] Li Zhaoshu and others designed an online monitoring system for partial discharge of high-voltage switchgear based on the pulse current method to monitor the operating status of the high-voltage switchgear and determine whether it is in a fault state. The system collects ultra-high frequency signals of the high-voltage switchgear through an ultra-high frequency sensor module, and uses the second-generation wavelet denoising technology to denoise the signal. At the same time, the optimal binary tree complex wavelet decomposition method is used to extract the characteristics of the discharge signal, and finally a partial discharge feature fusion recognition model based on neural network and evidence theory is constructed.

[0004] Jia Kunqi et al. collected ultrasonic signals from high-voltage switchgear, extracted waveform feature vectors, the first and second correlation coefficients, and the peak values ​​of each signal segment, and determined the local discharge feature vectors. Subsequently, they combined the ultra-high frequency detection method with these feature vectors for multi-information fusion, thereby realizing the diagnosis of local discharge fault types in high-voltage switchgear.

[0005] These algorithms involve complex feature extraction processes and related domain knowledge, and usually require expert knowledge guidance to achieve accurate pattern recognition, and rely heavily on experience and technology. In addition, manual feature extraction alone does not allow the model to fully understand and utilize the original data features. When the local discharge patterns are similar, it is difficult to achieve good fault diagnosis results. Summary of the invention

[0006] The present invention aims to solve one of the technical problems in the above-mentioned related art at least to a certain extent.

[0007] To this end, the purpose of the present invention is to provide a switch cabinet fault diagnosis method and device based on deep learning, which can analyze signal characteristics from different levels and scales through multi-scale feature fusion, thereby improving the fault diagnosis accuracy and adaptability.

[0008] In order to solve the above-mentioned technical problems, the present invention is achieved as follows:

[0009] An embodiment of the present invention provides a switch cabinet fault diagnosis method based on deep learning, the method comprising:

[0010] S1. Obtain PD data;

[0011] S2, preprocessing the acquired PD data;

[0012] The preprocessing includes: extracting global statistical features from the PD data to obtain global statistical features; and converting the data type of the PD data to obtain a grayscale image;

[0013] S3, using a ResNet model combined with a CBAM attention mechanism to process the grayscale image and extract key potential features as key local discriminant features;

[0014] S4, performing feature splicing on the global statistical features and the key local discriminant features to obtain multi-scale features;

[0015] S5. Use an MLP classifier to classify the multi-scale features to obtain a switch cabinet fault diagnosis result.

[0016] In addition, the switch cabinet fault diagnosis method based on deep learning according to the present invention may also have the following additional technical features:

[0017] In some of the embodiments, the global statistical features include the number of pulses, the average absolute amplitude of the pulses, the standard deviation of the pulses, and the maximum and minimum values ​​of the pulses.

[0018] In some embodiments, the PD data is a partial discharge phase-resolved pulse sequence or phase-resolved partial discharge data.

[0019] In some of the embodiments, the CBAM attention mechanism includes a channel attention mechanism and a spatial attention mechanism. By combining the channel attention mechanism and the spatial attention mechanism, the key potential features related to the PD pattern can be effectively extracted.

[0020] In some of the embodiments, the MLP classifier is a multi-layer perceptron designed based on multi-scale features, and the switch cabinet fault diagnosis result determined by the MLP classifier includes the presence or absence of a fault and the type of fault.

[0021] In some of the embodiments, the data type conversion in step S2 includes: organizing the PD data into a two-dimensional data matrix, normalizing the data matrix, multiplying the normalized data by 255 and converting it into an integer format to adapt to the grayscale image pixel value range of 0 to 255, thereby obtaining a grayscale image that can clearly display the spatial distribution and intensity of local discharge activity.

[0022] In some of the embodiments, the CBAM attention mechanism is set in the basic block of the ResNet model.

[0023] In some of the embodiments, the channel attention mechanism includes an average pooling layer and a maximum pooling layer, a fully connected layer, a ReLU module and a fully connected layer arranged in parallel;

[0024] The formula of the channel attention mechanism is:

[0025]

[0026] Where F represents the input feature map, σ represents the sigmoid activation function, W0 and W1 represent the weights of the two fully connected layers respectively, and M c Represents the output channel attention feature map.

[0027] In some of the embodiments, the spatial attention mechanism includes a maximum pooling layer, an average pooling layer, and a convolutional layer that are arranged sequentially and adjacently;

[0028] The formula of the spatial attention mechanism is:

[0029]

[0030] In the formula, F represents the input feature map, σ represents the sigmoid activation function, and f 7×7 represents a 7x7 convolution operation, M s Represents the output spatial attention feature map.

[0031] An embodiment of the present invention also provides a switch cabinet fault diagnosis device based on deep learning, including a processor and a memory, wherein a program is stored in the memory, and when the processor runs the program in the memory, the content of the switch cabinet fault diagnosis method based on deep learning as described in any one of the above can be implemented.

[0032] Compared with the prior art, the present invention has at least the following beneficial effects:

[0033] In the embodiment of the present invention, the switch cabinet fault diagnosis method based on deep learning is provided, which can effectively mine the key features of the partial discharge mode by making full use of the information in the original data, thereby enhancing the recognition ability of the model;

[0034] In the embodiment of the present invention, the switch cabinet fault diagnosis method based on deep learning is provided, which enables the model to analyze signal features from different levels and scales through multi-scale feature fusion, thereby improving the adaptability to complex patterns;

[0035] In the embodiment of the present invention, the switch cabinet fault diagnosis method based on deep learning is provided, which adopts the automatic feature extraction process of deep learning, reduces the dependence on domain knowledge, and improves the versatility and application potential of the system;

[0036] In the embodiment of the present invention, the switch cabinet fault diagnosis method based on deep learning is provided, which combines global statistical features with local discriminant features extracted by deep learning, reflects the different dimensional states of partial discharge signals, can effectively reduce the influence of noise and abnormal data on the recognition results, and ensures reliable fault recognition in complex environments;

[0037] In the embodiment of the present invention, the switch cabinet fault diagnosis method based on deep learning is provided, which shows high accuracy in fault detection and type recognition through the combination of multi-level feature extraction, attention mechanism and deep learning model. Through precise feature expression and information integration, the model can quickly identify potential problems and provide accurate fault analysis, thereby significantly improving the diagnostic efficiency.

[0038] In the embodiments of the present invention, the switch cabinet fault diagnosis method based on deep learning is provided. Due to the universality of the structural design and feature extraction method, the model can adapt to the monitoring needs of different environments and can be extended to fault diagnosis in different application scenarios and environments, and has good application potential.

[0039] The switch cabinet fault diagnosis device based on deep learning of the present invention can implement the content of the switch cabinet fault diagnosis method based on deep learning, and thus at least has all the features and advantages of the switch cabinet fault diagnosis method based on deep learning, which will not be repeated here. Additional aspects and advantages of the present invention will be partially given in the following description, and part will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A flow chart of a switch cabinet fault diagnosis method based on deep learning disclosed in one embodiment of the present invention;

[0041] Figure 2 A schematic diagram of data preprocessing disclosed in one embodiment of the present invention;

[0042] Figure 3 A schematic diagram of the ResNet18 network structure disclosed in one embodiment of the present invention;

[0043] Figure 4 A schematic diagram of the CBAM attention mechanism structure disclosed in one embodiment of the present invention;

[0044] Figure 5 A schematic diagram of the channel attention mechanism structure disclosed in one embodiment of the present invention;

[0045] Figure 6 A schematic diagram of the structure of a spatial attention mechanism disclosed in one embodiment of the present invention;

[0046] Figure 7 A PRPS fault diagnosis confusion matrix disclosed in one embodiment of the present invention;

[0047] Figure 8 A PRPS fault mode classification confusion matrix disclosed in one embodiment of the present invention;

[0048] Fig. 9 A PRPD fault diagnosis confusion matrix disclosed in one embodiment of the present invention;

[0049] Fig.10 A PRPD fault mode classification confusion matrix disclosed in one embodiment of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] The embodiments of the present invention are described in detail below through specific embodiments and application scenarios in conjunction with the accompanying drawings.

[0052] See also Figure 1 As shown, in some embodiments of the present invention, a switch cabinet fault diagnosis method based on deep learning is provided, which uses a deep neural network for feature learning, combines channel attention and spatial attention, focuses on the discriminative potential features related to local discharge patterns, and combines artificially extracted signal statistical features with deep potential features, so as to determine whether there is a fault, and further determine the fault type when a fault exists.

[0053] Based on the phase resolved pulse sequence (PRPS) or phase resolved partial discharge (PRPD) data collected on site, the following steps are performed.

[0054] (1) PD data preprocessing, mainly including statistical feature extraction and local data conversion, extracting the global statistical information of PD data and converting the original data into a single-channel three-dimensional image matrix, i.e., grayscale image data;

[0055] (2) For the converted grayscale image, feature learning is performed in a ResNet network combined with an attention mechanism. In order to solve the problem of insufficient feature extraction capability of the traditional ResNet network, the present invention introduces the CBAM (Convolutional Block Attention Module) attention mechanism, focusing on useful information related to the PD pattern, and obtaining key local discriminant features expressing the PD category pattern through deep feature extraction. The statistical feature variables expressing the global information of the PD signal and the deep extraction information expressing the local information are spliced ​​and fused to obtain multi-scale features.

[0056] (3) Based on multi-scale features, a multilayer perceptron (MLP) is designed to diagnose the presence and type of fault simultaneously.

[0057] In some embodiments of the present invention, the content of data preprocessing includes:

[0058] 1) Global statistical information extraction

[0059] The number of pulses has significant advantages as a global feature for partial discharge identification. It directly reflects the activity of partial discharge and provides an important indication of the health status of the equipment. By monitoring the changes in the number of pulses per unit time, potential faults can be identified in real time and trends can be analyzed to determine whether the discharge phenomenon is gradually intensifying. In addition, the calculation of the number of pulses is relatively simple, easy to understand, and convenient for quick decision-making. It is not easily affected by single abnormal data, thereby improving the reliability of identification. Comparison of the number of pulses between different devices can also help quickly identify abnormal situations and improve monitoring efficiency.

[0060] The average absolute amplitude of the pulse refers to the average value of the absolute amplitudes of all partial discharge pulses within a certain time window, which can effectively reflect the overall strength of the partial discharge signal. By calculating the absolute amplitude of each pulse and averaging it, this feature not only provides an important basis for fault diagnosis, but also has a strong anti-interference ability, because the use of absolute values ​​can reduce the impact of negative amplitudes on the results. In addition, the calculation of the average absolute amplitude is simple and easy to understand, which helps to quickly evaluate the operating status of the equipment, so it has important application value in partial discharge feature extraction.

[0061] The standard deviation of the pulse can effectively reflect the fluctuation of the amplitude of the partial discharge signal and provide a strong indication of the health status of the equipment. A higher standard deviation usually means that the equipment has a greater risk of instability or failure. In addition, this feature is easy to calculate and understand in feature analysis, which helps to quickly evaluate the operating status of the equipment. In general, the standard deviation of the pulse is an important indicator in partial discharge identification and fault diagnosis.

[0062] The maximum and minimum values ​​of the pulse can help quickly identify abnormal behavior and fault conditions of the equipment. The height of the maximum value can indicate whether there is a serious fault, while the minimum value can provide information on signal stability. These two features are simple and easy to calculate, which helps in real-time monitoring and quick decision-making, and therefore have important application value in partial discharge identification and fault diagnosis.

[0063] 2) Data image format conversion

[0064] During the monitoring process, PD signals are usually collected by sensors or detection equipment. These signals can be organized into a two-dimensional data matrix. In order to effectively convert the data into an image, the data matrix needs to be normalized. This operation scales the amplitude of the signal to the range of [0,1] to facilitate subsequent image generation. The normalization step involves finding the minimum and maximum values ​​of the data matrix, and then applying the following formula:

[0065]

[0066] By normalizing, the contrast of the image can be enhanced, making the signal features more obvious in the image. The normalized data matrix can be converted into a grayscale image through the image processing library. In this step, the normalized data is multiplied by 255 and converted to integer format (uint8) to fit the pixel value range of the grayscale image (0 to 255). The resulting image can clearly show the spatial distribution and intensity of the local discharge activity. The process is as follows Figure 2 shown.

[0067] In some embodiments of the present invention, the content of the ResNet model combined with the CBAM attention mechanism includes:

[0068] The deep neural network used for PD signal pattern feature extraction is as follows Figure 3The ResNet18 structure shown in the figure inputs the preprocessed grayscale image data into the ResNet model, and the image passes through multiple residual blocks, each of which consists of multiple convolutional layers, which are responsible for extracting features at different levels. By introducing jump connections, the network can retain input information during training, ensuring that each layer can learn useful features. During the feature extraction process, batch normalization is applied after the convolutional layer to accelerate training and stabilize the performance of the model. It standardizes the output of each layer so that the distribution of the input of each layer is more consistent, thereby reducing fluctuations during training. In addition, the ReLU activation function introduces nonlinear features, allowing the network to learn more complex patterns. As the image is passed layer by layer in the network, the model gradually extracts multi-level information from low-level features (such as edges and textures) to high-level features (such as partial discharge patterns).

[0069] In order to improve the model's ability to extract key features in large-scale data processing, the CBAM attention module is introduced in ResNet. Figure 4 As shown in the figure, this module mainly includes channel attention mechanism and spatial attention mechanism. Through the combination of the two attention mechanisms, they complement each other and effectively extract key information related to the PD pattern.

[0070] The operating mechanism of channel attention is as follows Figure 5 As shown in the figure, the input feature map passes through the parallel average pooling layer (AvgPool) and the maximum pooling layer (MaxPool), compresses the spatial size while keeping the channel unchanged to obtain two feature maps, and then passes through the shared MLP module to compress and expand the channels of the feature map to the original number of channels, and enhances the channel feature extraction capability through dimensional transformation. After that, the ReLU activation function is passed and element-by-element addition is performed. In order to convert the feature into an attention score, a sigmoid activation function is used for calculation. The attention formula of this process is shown as follows:

[0071]

[0072] Where F represents the input feature map, σ represents the sigmoid activation function, W0 and W1 represent the weights of the two fully connected layers respectively, and M c Represents the output channel attention feature map.

[0073] Global average pooling and maximum pooling reflect the importance of different channels in the feature extraction process. The model uses these descriptors to calculate the weight of each channel, so that the network can focus on the channels that are most critical for distinguishing local discharge features. This focus not only enhances the response to valid signals, but also suppresses the impact of redundant or noise information.

[0074] In addition to the above channel-level compression and expansion, in order to extract the key features of the PD signal at the spatial level, the output features of the channel attention mechanism are transmitted into Figure 6 The spatial attention module shown in the figure first compresses the feature channel through maximum pooling and average pooling to reflect the importance of each spatial position in the feature map, then concatenates the two features through the Concat operation and converts them into a feature map with a channel of 1 through the convolution layer. The feature variables are converted into attention scores through the sigmoid function. The network can emphasize the area with the most concentrated local discharge activity, thereby extracting spatial features more effectively. The formula is shown in the following formula:

[0075]

[0076] In the formula, F represents the input feature map, σ represents the sigmoid activation function, and f 7×7 represents a 7x7 convolution operation, M s Represents the output spatial attention feature map.

[0077] Finally, the CBAM module was placed in the Basic Block in the ResNet network to weaken the model's attention to irrelevant information. The model can not only identify local discharge patterns more accurately, but also maintain robustness in complex background and noisy environments, reducing the risk of false alarms and missed alarms in actual monitoring data processing, thereby improving the overall diagnostic accuracy.

[0078] In some embodiments of the present invention, the content of the MLP classifier based on multi-scale features includes:

[0079] Based on the global statistical features of partial discharge (PD) signals and the key local discriminant features extracted by deep neural networks, a diagnostic classifier using MLP is designed. The classifier mainly includes a shared fully connected layer and two independent fully connected classifiers, which are used to determine whether the PD signal has a fault and determine the specific type of fault.

[0080] 1) Shared fully connected layer: This layer receives input from the feature extraction part, performs preliminary feature fusion, and extracts more abstract feature representations. The shared layer can make full use of the complementary information of global statistical features and local discriminant features to enhance the learning ability of the model.

[0081] 2) Fully connected classifier: The fault detection classifier is used to determine whether the PD signal has a fault. It helps to quickly identify potential problems by outputting a binary classification result (faulty / no fault). The fault type classifier is used to further subdivide the fault type, such as internal discharge, external discharge, and corona discharge. This layer provides multi-level fault classification, which helps to more accurately locate and diagnose specific fault causes.

[0082] Through this multi-level classification design, the model can not only improve the accuracy of fault detection, but also conduct in-depth analysis on fault type identification. The shared fully connected layer effectively integrates different features, enhances the model's ability to recognize complex patterns, and ultimately improves the overall diagnostic accuracy. The flexibility and hierarchy of this structure enable the model to make quick and accurate judgments when faced with diverse PD signals, thereby providing more reliable support in power equipment maintenance.

[0083] The present invention combines global statistical features with local discriminant features extracted by deep learning, and captures the multi-dimensional information of the signal through multi-level information fusion, which helps the model to understand the local discharge pattern more comprehensively and enhances the feature expression ability of the model. The adopted CBAM attention module effectively enhances the extraction of key information by combining channel attention and spatial attention. The channel attention mechanism weights the features according to the importance of different channels, and the spatial attention mechanism emphasizes the important spatial positions in the feature map. This mechanism ensures that the model can focus on the areas where the local discharge activity is most concentrated. The present invention adopts the design of shared fully connected layers and independent classifiers, so that the model has higher flexibility in fault detection and type recognition. The shared layer can fuse different features and enhance learning ability, while the independent classifier handles the existence of faults and type recognition respectively, which improves the performance of the model in complex modes.

[0084] The following are two experiments and results of the practical application of the present invention.

[0085] Example 1:

[0086] Based on the PRPS data collected by the ultrasonic sensor, deep feature extraction and pattern recognition tasks are performed, and the model parameter settings are shown in Table 1.

[0087] Table 1 PRPS data experimental model parameter settings

[0088]

[0089] The MLP classifier designed with global features and deep key features can effectively perform fault diagnosis and pattern classification on the equipment data of the actual substation switchgear. Specifically, PRPS data monitoring was performed on three different switchgears on site. It is known that the local discharge modes of the three switchgears are different, which are divided into no local discharge, low-intensity local discharge and high-intensity local discharge. Feature extraction and fault diagnosis are performed on the PRPS data of the three different switchgears. The input data is a grayscale image of the PRPS data after normalization and image conversion. After deep feature extraction by deep neural network, the discriminative latent features and the global statistical features of the PRPS data are fused and input into the MLP classifier. The classifier can effectively perform fault diagnosis and pattern recognition. The recognition results are shown in Table 2.

[0090] Table 2 PRPS data experimental pattern recognition accuracy

[0091]

[0092] Among them, the fault diagnosis accuracy rate reached 99.71%, the fault classification accuracy rate reached 98.84%, and the fault diagnosis confusion matrix was as follows: Figure 7 As shown, the failure mode classification confusion matrix is ​​as follows Figure 8 As shown, 0 represents no fault and no partial discharge occurs, 1 represents partial discharge occurs in the switch cabinet and the partial discharge intensity is low, and 2 represents partial discharge occurs in the switch cabinet and the partial discharge intensity is high. The partial discharge pattern recognition algorithm proposed in the present invention can effectively classify the monitored PRPS data for partial discharge with high classification accuracy.

[0093] Example 2:

[0094] Based on the PRPD data collected by the ultrasonic sensor, deep feature extraction and pattern recognition tasks are performed, and the model parameter settings are shown in Table 3.

[0095] Table 3 PRPD data experimental model parameter settings

[0096]

[0097] The MLP classifier designed by global features and deep key features can effectively perform fault diagnosis and pattern classification on the equipment data of the actual substation switch cabinet. Specifically, PRPD data monitoring was performed on three different switch cabinets on site. It is known that the local discharge modes of the three switch cabinets are different, which are divided into no local discharge, low-intensity local discharge and high-intensity local discharge. Feature extraction and fault diagnosis are performed on the PRPD data of the three different switch cabinets. The input data is the grayscale image of the PRPD data after normalization and image conversion. After deep feature extraction by deep neural network, the discriminative potential features and the global statistical features of the PRPD data are fused and input into the MLP classifier. The classifier can effectively perform fault diagnosis and pattern recognition. The recognition results are shown in Table 4.

[0098] Among them, the fault diagnosis accuracy rate reached 99.71%, the fault classification accuracy rate reached 98.55%, and the fault diagnosis confusion matrix was as follows: Fig. 9 As shown, the failure mode classification confusion matrix is ​​as follows Fig.10 As shown, 0 represents no fault and no partial discharge occurs, 1 represents partial discharge occurs in the switch cabinet and the partial discharge intensity is low, and 2 represents partial discharge occurs in the switch cabinet and the partial discharge intensity is high. The partial discharge pattern recognition algorithm proposed in the present invention can effectively classify the monitored PRPS data for partial discharge with high classification accuracy.

[0099] Table 4 PRPD data experimental pattern recognition accuracy

[0100]

[0101] The parts of the present invention that are not described in detail may refer to the prior art or are known to those skilled in the art.

[0102] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A switch cabinet fault diagnosis method based on deep learning, characterized in that: The method comprises: S1. Obtain PD data; S2, preprocessing the acquired PD data; The preprocessing includes: extracting global statistical features from the PD data to obtain global statistical features; and converting the data type of the PD data to obtain a grayscale image; S3, using a ResNet model combined with a CBAM attention mechanism to process the grayscale image and extract key potential features as key local discriminant features; S4, performing feature splicing on the global statistical features and the key local discriminant features to obtain multi-scale features; S5. Use an MLP classifier to classify the multi-scale features to obtain a switch cabinet fault diagnosis result.

2. The switch cabinet fault diagnosis method based on deep learning according to claim 1 is characterized in that: The global statistical features include the number of pulses, the average absolute amplitude of the pulses, the standard deviation of the pulses, and the maximum and minimum values ​​of the pulses.

3. The switch cabinet fault diagnosis method based on deep learning according to claim 1 is characterized in that: The PD data is a partial discharge phase-resolved pulse sequence or phase-resolved partial discharge data.

4. The switch cabinet fault diagnosis method based on deep learning according to claim 1 is characterized in that: The CBAM attention mechanism includes a channel attention mechanism and a spatial attention mechanism. By combining the channel attention mechanism and the spatial attention mechanism, the key potential features related to the PD pattern can be effectively extracted.

5. The switch cabinet fault diagnosis method based on deep learning according to claim 1 is characterized in that: The MLP classifier is a multi-layer perceptron designed based on multi-scale features, and the switch cabinet fault diagnosis result determined by the MLP classifier includes the presence or absence of a fault and the type of fault.

6. The switch cabinet fault diagnosis method based on deep learning according to claim 1, characterized in that: The data type conversion in step S2 includes: organizing the PD data into a two-dimensional data matrix, normalizing the data matrix, multiplying the normalized data by 255 and converting it into an integer format to adapt to the grayscale image pixel value range of 0 to 255, thereby obtaining a grayscale image that can clearly display the spatial distribution and intensity of partial discharge activity.

7. The switch cabinet fault diagnosis method based on deep learning according to claim 1 is characterized in that: The CBAM attention mechanism is set in the basic block of the ResNet model.

8. The switch cabinet fault diagnosis method based on deep learning according to claim 4 is characterized in that: The channel attention mechanism includes an average pooling layer and a maximum pooling layer, a fully connected layer, a ReLU module and a fully connected layer set in parallel; The formula of the channel attention mechanism is: Where F represents the input feature map, σ represents the sigmoid activation function, W0 and W1 represent the weights of the two fully connected layers respectively, and M c Represents the output channel attention feature map.

9. The switch cabinet fault diagnosis method based on deep learning according to claim 4 is characterized in that: The spatial attention mechanism includes a maximum pooling layer, an average pooling layer, and a convolutional layer that are arranged sequentially and adjacently; The formula of the spatial attention mechanism is: In the formula, F represents the input feature map, σ represents the sigmoid activation function, and f 7×7 represents a 7x7 convolution operation, M s Represents the output spatial attention feature map.

10. A switch cabinet fault diagnosis device based on deep learning, comprising a processor and a memory, wherein a program is stored in the memory, characterized in that: When the processor runs the program on the memory, it can implement the content of the switch cabinet fault diagnosis method based on deep learning as described in any one of claims 1 to 9.

Images