Transmission Line Insulator Fault Diagnosis Method Based on Deep Learning

Abstract

The present invention relates to a transmission line insulator fault diagnosis method based on deep learning, which includes the following steps: step 1: collecting and processing image information and operating condition information of the transmission line insulator; step 2: training an insulator fault diagnosis model based on the insulator image data and the insulator operating condition feature matrix data; step 3: determining the training effect of the transmission line insulator fault diagnosis model and saving the trained model; and step 4: applying the insulator fault diagnosis model to online fault diagnosis of transmission line insulators. The transmission line insulator fault diagnosis method based on deep learning provided by the present invention can accurately classify different fault types and significantly improve the speed of transmission line insulator fault diagnosis. By integrating a data preprocessing process and an end-to-end deep learning network, end-to-end fault diagnosis of transmission line insulators is achieved, the fault diagnosis process is simplified, and a higher fault diagnosis accuracy rate is achieved.

Description

Technical Field

[0001] The present application relates to the technical field related to physical component fault diagnosis, and specifically to a transmission line insulator fault diagnosis method based on deep learning. Background Art

[0002] An insulator is a physical component installed between conductors at different potentials, or between a conductor and a grounded structure, capable of withstanding voltage and mechanical stress. Insulators come in a wide variety of types and shapes. While their structure and appearance vary significantly, they all consist of two main components: the insulating element and the connecting fittings. Transmission line insulators are the largest component in the entire power system and are directly linked to transmission line failures. Insulator failures directly or indirectly contribute to the vast majority of transmission line failures. Transmission line insulators come in many different types. Based on their material, they can be categorized into three main types: porcelain, glass, and composite. Each type of insulator is susceptible to distinct types of failures. Over the long-term operation of transmission lines, insulators are subject to numerous environmental influences, such as climate and temperature fluctuations. Variations in the current and voltage of the transmission line can also affect their condition. Furthermore, mechanical loads can also affect the insulator's condition. At present, the possible fault types of transmission line insulators under various working conditions can be summarized into six types: arc creep, string drop, self-explosion, fracture, resistance degradation and surface contamination.

[0003] Traditional methods for diagnosing transmission line insulator faults include the following: the spark gap method, which relies on determining whether the insulator can generate a discharge, a key indicator of faulty insulators; the ball discharge method, which measures the voltage distribution of the insulator by measuring the discharge distance between balls at both ends of the insulator. Both methods suffer from the inability to determine the specific fault type and low diagnostic accuracy; the infrared thermal imaging method, which relies primarily on the thermal effect of the insulator surface. The surface temperature of a faulty insulator will be lower than that of a normal insulator. Although this method has high accuracy, it still suffers from the inability to diagnose specific fault types; and the leakage current detection method, which uses a current sensor to measure the leakage current flowing through the insulator to diagnose insulator faults. However, this method cannot perform live detection and is prohibitively expensive. Fault diagnosis technology based on deep learning visual recognition has developed rapidly and has also been applied to transmission line insulator fault diagnosis. However, existing methods only diagnose multiple fault types based on insulator images, failing to consider the impact of operating conditions on insulator failures. Furthermore, they are completely unable to detect faults that are not visually apparent, such as resistance degradation.

[0004] With the increasing application of attention mechanism networks in deep learning, visual question answering task networks that combine attention mechanisms with deep convolutional neural networks have been developed; compared with diagnostic methods that rely solely on images, deep neural networks based on attention mechanisms and convolutional neural networks can realize the combined analysis of image features and insulator operating conditions, and have more advantages in the field of insulator fault diagnosis. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the deep learning-based transmission line insulator fault diagnosis method provided by the present invention can accurately classify different fault types and significantly improve the speed of transmission line insulator fault diagnosis. By integrating the data preprocessing process and the end-to-end deep learning network, end-to-end fault diagnosis of transmission line insulators is realized, the fault diagnosis process is simplified, and a higher fault diagnosis accuracy is achieved.

[0006] To achieve the above objectives, the solution adopted by the present invention is:

[0007] A method for diagnosing transmission line insulator faults based on deep learning, comprising the following steps:

[0008] Step 1: Collect and process image information and working condition information of transmission line insulators;

[0009] Collect image information of transmission line insulators, adjust the image scale through pixel sampling, and perform normalization processing to obtain insulator image data;

[0010] Collecting eight-dimensional data on the operating conditions of transmission line insulators, including insulator material, transmission voltage, transmission current, insulator mechanical load, steel cap temperature, insulator temperature, ambient temperature, and weather conditions. This eight-dimensional data is then standardized and filled with background data to obtain a uniformly dimensioned insulator operating condition feature matrix.

[0011] Step 2: Train the insulator fault diagnosis model based on the insulator image data and the insulator working condition feature matrix data;

[0012] Training data is formed based on the insulator image data and the insulator working condition feature matrix data in step 1, the training data is divided into a training data set and a validation data set in proportion, and the training data set is input into a transmission line insulator fault diagnosis model based on deep learning for training; the transmission line insulator fault diagnosis model based on deep learning is trained by: an ICN deep learning module, an ECNN deep learning module, a TSAN deep learning module, a common attention mechanism layer, and an output fully connected layer;

[0013] The key model structure of the TSAN deep learning module is the self-attention mechanism layer, which is expressed as follows:

[0014]

[0015] Where: Attention(Q, K, V) represents the self-attention mechanism function; Q represents the first intermediate data of the self-attention mechanism layer; K represents the second intermediate data of the self-attention mechanism layer; V represents the third intermediate data of the self-attention mechanism layer; Swish represents the first activation function of the self-attention mechanism layer; W i represents the first learning parameter of the self-attention mechanism layer; b i represents the second learning parameter of the self-attention mechanism layer; d represents the vector length of the first intermediate data Q and the second intermediate data K of the self-attention mechanism layer; x represents the input of the self-attention mechanism layer of the TSAN deep learning module; i represents different parameter numbers;

[0016] The joint attention mechanism layer can achieve the fusion of insulator image features and insulator working condition semantic features. The joint attention module is trained together with the feature extraction network and automatically optimizes the learning parameters. The expression of the joint attention mechanism layer is as follows:

[0017]

[0018] Where: α represents the first intermediate data of the common attention mechanism layer; U represents the first learning parameter of the common attention mechanism layer; y represents the output of the deep learning module of the ICN network and the ECNN network; l represents the second learning parameter of the common attention mechanism layer; β represents the second intermediate data of the common attention mechanism layer; sigmoid represents the first activation function of the common attention mechanism layer; Swishb represents the second activation function of the common attention mechanism layer; z represents the output of the self-attention mechanism layer of the TSAN deep learning module; m represents the fourth learning parameter of the common attention mechanism layer; output represents the output of the common attention mechanism layer;

[0019] The activation function in the output fully connected layer is softmax, and the specific expression is as follows:

[0020]

[0021] Where: j represents the neuron number of the fully connected layer; C j represents the output of the jth neuron; ω j represents the first learning parameter of the output fully connected layer; β j Represents the second learning parameter of the output fully connected layer; class j represents the probability that the input data belongs to the j-th defect category; * represents matrix multiplication; softmax represents the activation function in the output fully connected layer;

[0022] Step 3: Determine the training effect of the transmission line insulator fault diagnosis model and save the trained model;

[0023] The training effect of the transmission line insulator fault diagnosis model is judged based on the validation data set in step 2. The transmission line insulator fault diagnosis model based on deep learning outputs the fault type of the abnormal insulator. When the mean absolute error of the validation data set is less than 0.9%, the model training is completed and the trained transmission line insulator fault diagnosis model parameters are saved. The mean absolute error calculation formula is as follows:

[0024]

[0025] Where: Lmp represents the mean absolute error of the data set; N represents the number of batches in the data set; k represents the batch number; ACC k Indicates the absolute accuracy of the network inference result in the kth batch;

[0026] Step 4: Apply the insulator fault diagnosis model to online fault diagnosis of transmission line insulators;

[0027] The input data for online fault diagnosis must first undergo the same data preprocessing as the training data in step 1, and then be passed into the deep learning-based insulator fault diagnosis model to obtain the transmission line insulator fault type, ultimately completing the transmission line insulator fault diagnosis.

[0028] Preferably, the training data in step 2 needs to use the fault diagnosis results of professional transmission line engineers to produce real fault condition labels. The fault conditions of insulators include six categories: arc creep, string drop, self-explosion, fracture, resistance degradation and surface contamination.

[0029] Preferably, the feature extraction network of the transmission line insulator fault diagnosis model based on deep learning in step 2 includes an ICN deep learning module consisting of a convolution layer, a deconvolution layer, a Concat mechanism, an activation function, a pooling layer and a fully connected layer, an ECNN deep learning module and a TSAN deep learning module, a common attention mechanism layer and an output fully connected layer.

[0030] Preferably, the transmission line insulator fault diagnosis model based on deep learning in step 2 needs to calculate the cross entropy loss function as shown below:

[0031]

[0032] Where: L represents the cross entropy loss function; M represents the total number of fault categories; y i represents the confidence of the i-th fault category; y i Indicates whether the fault actually occurs in this fault category.

[0033] Preferably, the insulator image data in step 2 needs to be input into an ICN network for feature integration, the output data of the ICN network is then input into an ECNN network to obtain insulator image feature data, the insulator working condition feature data is input into a TSAN deep learning module to obtain insulator working condition semantic feature data, and then the insulator image feature data and the insulator working condition semantic feature data are input into a common attention mechanism module, and finally the diagnosis result is output through the output fully connected layer;

[0034] Preferably, the ICN deep learning module and the ECNN deep learning module in step 2 are specifically:

[0035] The ICN deep learning module uses three downsampling convolutional layers and a maximum pooling with a stride of 2 and a deconvolution layer corresponding to the downsampling convolution;

[0036] The ECNN deep learning module contains two convolution layers with a convolution kernel size of 5 and a convolution step size of 1. It uses two maximum pooling series structures with a step size of 2, with a total of four pooling layers.

[0037] Preferably, the influence degree of the insulator working condition information is obtained in the common attention mechanism layer in step 2 through the fully connected layer and the Sigmoid function, and then the real image feature data of the insulator obtained by the ICN and ECNN deep learning modules are combined to obtain the affected integrated feature data and output the fault diagnosis result through the Softmax function.

[0038] Preferably, the TSAN deep learning module uses a multi-head self-attention mechanism with 8 heads, a feature expansion layer and a fully connected layer; the TSAN deep learning module can automatically extract fault diagnosis sensitive information from insulator operating condition data, and TSAN is trained together with the entire network to automatically optimize learning parameters; the diagnostic sensitive information of insulator operating condition data obtained by the TSAN deep learning module can affect the final fault diagnosis result through the common attention mechanism module.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] (1) The network model proposed in the present invention uses an insulator working condition data feature extraction network based on the self-attention mechanism, which can optimize parameters together with the overall network to achieve quantification of the impact of the insulator working condition on the diagnosis result; using a fault diagnosis result output network based on the joint attention mechanism, the network output module can optimize parameters together with the overall network, realizing the association and result output of the insulator working condition feature influence and the insulator image feature.

[0041] (2) The deep learning-based transmission line insulator fault diagnosis method provided by the present invention can significantly improve the speed of transmission line insulator fault diagnosis and accurately classify different fault types, providing a reference for subsequent fault resolution and troubleshooting, and helping to quickly restore power supply. The deep learning-based transmission line insulator fault diagnosis method provided by the present invention achieves end-to-end fault diagnosis of transmission line insulators by integrating a data preprocessing process and an end-to-end deep learning network.

[0042] (3) The present invention provides users with a more convenient diagnostic method that can be operated without having to master a large amount of professional transmission engineering knowledge. It simplifies the insulator fault diagnosis process and enables ordinary workers to perform insulator fault diagnosis after completing data collection. By using the designed network structure, a higher fault diagnosis accuracy is achieved than that of existing deep learning networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 This is a control block diagram of a transmission line insulator fault diagnosis method based on deep learning according to an embodiment of the present invention;

[0044] Figure 2 This is a simplified flowchart of the steps of a transmission line insulator fault diagnosis method based on deep learning according to an embodiment of the present invention;

[0045] Figure 3 This is a diagram of the network model structure of an embodiment of the present invention;

[0046] Figure 4 This is a computational graph of the common attention mechanism layer in the network model of an embodiment of the present invention;

[0047] Figure 5 This is a computational graph for the multi-head self-attention layer in the network model of an embodiment of the present invention;

[0048] Figure 6 This is a graph showing the change in loss value and accuracy during the network training process according to an embodiment of the present invention;

[0049] Figure 7 This is a confusion matrix diagram of the network reasoning results considering only two fault types in an embodiment of the present invention;

[0050] Figure 8 This is an example of a real image of an insulator used in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

[0052] The network model proposed in the embodiment of the present invention uses an insulator working condition data feature extraction network based on the self-attention mechanism to quantify the impact of the insulator working condition on the diagnosis result, and uses a fault diagnosis result output network based on the joint attention mechanism to achieve the association and result output of the insulator working condition feature impact and the insulator image feature; the transmission line insulator fault diagnosis method based on deep learning can greatly improve the transmission line insulator fault diagnosis speed, and can accurately classify different fault types, provide a reference for subsequent fault resolution and elimination, and help to quickly restore power supply. The transmission line insulator fault diagnosis method based on deep learning provided by the present invention realizes end-to-end fault diagnosis of transmission line insulators by integrating the data preprocessing process and the end-to-end deep learning network; this case simplifies the insulator fault diagnosis process, so that ordinary workers can realize insulator fault diagnosis after completing data collection, and improves the fault diagnosis accuracy. As Figure 1 Shown is a control block diagram of a transmission line insulator fault diagnosis method based on deep learning according to an embodiment of the present invention.

[0053] The embodiment of the present invention provides a method for diagnosing transmission line insulator faults based on deep learning, such as Figure 2 The figure shows a simplified flowchart of the steps of a transmission line insulator fault diagnosis method based on deep learning according to an embodiment of the present invention. To demonstrate the applicability of the present invention, it is applied to an example, specifically comprising the following steps:

[0054] S1: Collect and process image information and working condition information of transmission line insulators;

[0055] The image information of the transmission line insulator is collected, the image scale is adjusted by pixel sampling, and normalization is performed to obtain the insulator image data. The input image of the network is as follows: Figure 8 Shown is an example of a real image of an insulator used in an embodiment of the present invention.

[0056] The working condition information of transmission line insulators is collected, including 8-dimensional data: insulator material, transmission voltage, transmission current, insulator mechanical load, steel cap temperature, insulator temperature, ambient temperature and weather conditions. The 8-dimensional data is standardized and filled with background data to obtain the insulator working condition characteristic matrix data with unified dimensions. The insulator working condition data is a one-dimensional array containing eight data elements.

[0057] S2: Train the insulator fault diagnosis model based on the insulator image data and the insulator working condition feature matrix data;

[0058] Training data is constructed based on the insulator image data and insulator operating condition feature matrix data in S1. The training data is divided into a training dataset and a validation dataset in proportion. The training dataset is passed to the deep learning-based transmission line insulator fault diagnosis model for training. The training data needs to use the fault diagnosis results of professional transmission line engineers to produce real fault condition labels. The insulator fault conditions include six categories: arc creep, string drop, self-explosion, fracture, resistance degradation, and surface contamination. The category labels of the dataset are given in the form of six-dimensional unique hot encoding.

[0059] The transmission line insulator fault diagnosis model based on deep learning is trained including: ICN deep learning module, ECNN deep learning module, TSAN deep learning module, common attention mechanism layer and output fully connected layer; Figure 4 The figure shows the calculation diagram of the common attention mechanism layer in the network model of the embodiment of the present invention. The feature extraction network of the transmission line insulator fault diagnosis model based on deep learning includes an ICN deep learning module composed of a convolution layer, a deconvolution layer, a Concat mechanism, an activation function, a pooling layer and a fully connected layer, an ECNN deep learning module and a TSAN deep learning module, a common attention mechanism layer and an output fully connected layer. Figure 3 The network model structure diagram of the embodiment of the present invention is shown. The TSAN deep learning module uses a multi-head self-attention mechanism with 8 heads, a feature expansion layer, and a fully connected layer; Figure 5 The figure shows the computational diagram of the multi-head self-attention layer in the network model of an embodiment of the present invention. The TSAN deep learning module can automatically extract sensitive information for fault diagnosis from insulator operating data. TSAN is trained together with the entire network and can automatically optimize learning parameters. The diagnostic sensitive information of the insulator operating data obtained by the TSAN deep learning module can influence the final fault diagnosis results through the joint attention mechanism module. When the fault category is 6 and the batch size is set to 2, an example of the final result output by the network is as follows:

[0060]

[0061] Afterwards, conditional statements are used to convert the output data into fault category output to complete the model inference process.

[0062] The insulator image data needs to be input into the ICN network for feature integration, and the output data of the ICN network is then input into the ECNN network to obtain the insulator image feature data. The insulator working condition feature data is input into the TSAN deep learning module to obtain the insulator working condition semantic feature data. The insulator image feature data and the insulator working condition semantic feature data are then passed into the common attention mechanism module, and finally the diagnosis results are output through the output fully connected layer; the ICN deep learning module uses three downsampling convolution layers and a maximum pooling with a stride of 2 and a deconvolution layer corresponding to the downsampling convolution; the ECNN deep learning module contains two convolution layers with a convolution kernel size of 5 and a convolution step of 1, and uses two maximum pooling series structures with a stride of 2, for a total of four pooling layers.

[0063] In the common attention mechanism layer, the influence degree of the insulator working condition information is obtained through the fully connected layer and the Sigmoid function. Then, the real image feature data of the insulator obtained by the ICN and ECNN deep learning modules are combined to obtain the affected integrated feature data and output the fault diagnosis result through the Softmax function.

[0064] The key model structure of the TSAN deep learning module is the self-attention mechanism layer, which is expressed as follows:

[0065]

[0066] Where: Attention(Q, K, V) represents the self-attention mechanism function; Q represents the first intermediate data of the self-attention mechanism layer; K represents the second intermediate data of the self-attention mechanism layer; V represents the third intermediate data of the self-attention mechanism layer; Swish represents the first activation function of the self-attention mechanism layer; W i represents the first learning parameter of the self-attention mechanism layer; b i represents the second learning parameter of the self-attention mechanism layer; d represents the vector length of the first intermediate data Q and the second intermediate data K of the self-attention mechanism layer; x represents the input of the self-attention mechanism layer of the TSAN deep learning module; i represents different parameter numbers.

[0067] The joint attention mechanism layer can achieve the fusion of insulator image features and insulator working condition semantic features. The joint attention module is trained together with the feature extraction network and automatically optimizes the learning parameters. The expression of the joint attention mechanism layer is as follows:

[0068]

[0069] Where: α represents the first intermediate data of the common attention mechanism layer; U represents the first learning parameter of the common attention mechanism layer; y represents the deep learning module output of the ICN network and the ECNN network; l represents the second learning parameter of the common attention mechanism layer; β represents the second intermediate data of the common attention mechanism layer; sigmoid represents the first activation function of the common attention mechanism layer; Swishb represents the second activation function of the common attention mechanism layer; z represents the self-attention mechanism layer output of the TSAN deep learning module; m represents the fourth learning parameter of the common attention mechanism layer; output represents the output of the common attention mechanism layer.

[0070] The activation function in the output fully connected layer is softmax, and the specific expression is as follows:

[0071]

[0072] Where: j represents the neuron number of the fully connected layer; C j represents the output of the jth neuron; ω j represents the first learning parameter of the output fully connected layer; β j Represents the second learning parameter of the output fully connected layer; class j represents the probability that the input data belongs to the j-th defect category; * represents matrix multiplication; softmax represents the activation function in the output fully connected layer.

[0073] The transmission line insulator fault diagnosis model based on deep learning requires the calculation of the cross entropy loss function, such as Figure 6 The figure shows the change of loss value and accuracy rate during the network training process of the embodiment of the present invention; as shown below:

[0074]

[0075] Where: L represents the cross entropy loss function; M represents the total number of fault categories; y i represents the confidence of the i-th fault category; y i Indicates whether the fault actually occurs in this fault category.

[0076] S3: Determine the training effect of the transmission line insulator fault diagnosis model and save the trained model;

[0077] The training effect of the transmission line insulator fault diagnosis model is judged based on the validation data set in S2. The transmission line insulator fault diagnosis model based on deep learning outputs the fault type of the abnormal insulator. When the mean absolute error of the validation data set is less than 0.9%, the model training is completed and the trained transmission line insulator fault diagnosis model parameters are saved. The mean absolute error calculation formula is as follows:

[0078]

[0079] Where: Lmp represents the mean absolute error of the data set; N represents the number of batches in the data set; k represents the batch number; ACC k Indicates the absolute accuracy of the network inference result in the kth batch.

[0080] S4: Apply the insulator fault diagnosis model to online fault diagnosis of transmission line insulators;

[0081] The input data of online fault diagnosis first needs to undergo the same data preprocessing operation as the training data in S1, and then be input into the insulator fault diagnosis model based on deep learning to obtain the transmission line insulator fault type, and finally complete the transmission line insulator fault diagnosis. Figure 7 The figure shows a confusion matrix diagram of the network reasoning results of an embodiment of the present invention that only considers two fault types. Through the analysis of the confusion matrix, it can be seen that the present invention has a good effect on the diagnosis of transmission line insulator faults and can meet actual use requirements.

[0082] In summary, the prediction results of the transmission line insulator fault diagnosis method based on deep learning in this case have proved to be very effective.

[0083] (1) The embodiment of the present invention processes actual transmission line data through an insulator operating condition data feature extraction network based on a self-attention mechanism, which can optimize parameters together with the overall network to achieve quantification of the impact of the insulator operating condition on the diagnosis result; using a fault diagnosis result output network based on a joint attention mechanism, the network output module can optimize parameters together with the overall network, realizing the association and result output of the insulator operating condition feature impact and insulator image features.

[0084] (2) The deep learning-based transmission line insulator fault diagnosis method provided by the embodiment of the present invention can significantly improve the speed of transmission line insulator fault diagnosis and accurately classify different fault types, providing a reference for subsequent fault resolution and troubleshooting, and helping to quickly restore power supply. The deep learning-based transmission line insulator fault diagnosis method provided by the present invention achieves end-to-end fault diagnosis of transmission line insulators by integrating a data preprocessing process and an end-to-end deep learning network.

[0085] (3) The embodiment of the present invention simplifies the insulator fault diagnosis process, which can be operated without mastering a large amount of professional transmission engineering knowledge. It enables ordinary workers to perform insulator fault diagnosis after completing data collection. By comparing the diagnosis results, it can be seen that this case can improve the accuracy of fault diagnosis and meet actual use needs.

[0086] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A transmission line insulator fault diagnosis method based on deep learning, characterized in that: It includes the following steps: Step 1: Collect and process image information and working condition information of transmission line insulators; Collect image information of transmission line insulators, adjust the image scale through pixel sampling, and perform normalization processing to obtain insulator image data; Collecting eight-dimensional data on the operating conditions of transmission line insulators, including insulator material, transmission voltage, transmission current, insulator mechanical load, steel cap temperature, insulator temperature, ambient temperature, and weather conditions. This eight-dimensional data is then standardized and filled with background data to obtain a uniformly dimensioned insulator operating condition feature matrix. Step 2: Train a transmission line insulator fault diagnosis model based on deep learning based on insulator image data and insulator working condition feature matrix data; According to the insulator image data and the insulator working condition feature matrix data in step 1, the training data is divided into a training data set and a validation data set in proportion, and the training data set is input into the transmission line insulator fault diagnosis model based on deep learning for training; The transmission line insulator fault diagnosis model based on deep learning includes: an ICN deep learning module, an ECNN deep learning module, a TSAN deep learning module, a common attention mechanism layer and an output fully connected layer; The key model structure of the TSAN deep learning module is the self-attention mechanism layer, which is expressed as follows: Where: Attention(Q, K, V) represents the self-attention mechanism function; Q represents the first intermediate data of the self-attention mechanism layer; K represents the second intermediate data of the self-attention mechanism layer; V represents the third intermediate data of the self-attention mechanism layer; Swish represents the first activation function of the self-attention mechanism layer; W i represents the first learning parameter of the self-attention mechanism layer; b i represents the second learning parameter of the self-attention mechanism layer; d represents the vector length of the first intermediate data Q and the second intermediate data K of the self-attention mechanism layer; x represents the input of the self-attention mechanism layer of the TSAN deep learning module; i represents different parameter numbers; The insulator image data needs to be input into the ICN deep learning module for feature integration. The output data of the ICN deep learning module is then input into the ECNN deep learning module to obtain insulator image feature data. The insulator working condition feature data is then input into the TSAN deep learning module to obtain insulator working condition semantic feature data. The insulator image feature data and insulator working condition semantic feature data are then passed into the joint attention mechanism layer, and finally the diagnosis results are output through the output fully connected layer. The joint attention mechanism layer can achieve the fusion of insulator image features and insulator working condition semantic features. The joint attention mechanism layer is trained together with the entire network of the model and automatically optimizes the learning parameters. The expression of the joint attention mechanism layer is as follows: Where: α represents the first intermediate data of the common attention mechanism layer; U represents the first learning parameter of the common attention mechanism layer; y represents the output of the ICN deep learning module and the ECNN deep learning module; l represents the second learning parameter of the common attention mechanism layer; β represents the second intermediate data of the common attention mechanism layer; sigmoid represents the first activation function of the common attention mechanism layer; Swishb represents the second activation function of the common attention mechanism layer; z represents the output of the self-attention mechanism layer of the TSAN deep learning module; m represents the fourth learning parameter of the common attention mechanism layer; output represents the output of the common attention mechanism layer; The activation function in the output fully connected layer is softmax, and the specific expression is as follows: Where: j represents the neuron number of the fully connected layer; C j represents the output of the jth neuron; ω j represents the first learning parameter of the output fully connected layer; β j Represents the second learning parameter of the output fully connected layer; class j represents the probability that the input data belongs to the j-th defect category; * represents matrix multiplication; softmax represents the activation function in the output fully connected layer; Step 3: Determine the training effect of the deep learning-based transmission line insulator fault diagnosis model and save the trained model; The training effect of the deep learning-based transmission line insulator fault diagnosis model is determined based on the validation data set in step 2. The deep learning-based transmission line insulator fault diagnosis model outputs the fault type of the abnormal insulator. When the mean absolute error of the validation data set is less than 0.9%, the model training is completed and the trained parameters of the deep learning-based transmission line insulator fault diagnosis model are saved. The mean absolute error calculation formula is as follows: Where: Lmp represents the mean absolute error of the data set; N represents the number of batches in the data set; k represents the batch number; ACC k Indicates the absolute accuracy of the network inference result in the kth batch; Step 4: Apply the deep learning-based transmission line insulator fault diagnosis model to online fault diagnosis of transmission line insulators; The input data for online fault diagnosis must first undergo the same data preprocessing as the training data in step 1, and then be passed into the deep learning-based transmission line insulator fault diagnosis model to obtain the transmission line insulator fault type, and finally complete the transmission line insulator fault diagnosis.

2. The transmission line insulator fault diagnosis method based on deep learning according to claim 1, characterized in that: The training data in step 2 needs to use the fault diagnosis results of professional transmission line engineers to produce real fault condition labels. The fault conditions of insulators include six categories: arc creep, string drop, self-explosion, fracture, resistance degradation and surface contamination.

3. The transmission line insulator fault diagnosis method based on deep learning according to claim 1, characterized in that: The ICN deep learning module consists of convolutional layers, deconvolutional layers, Concat mechanisms, activation functions, pooling layers, and fully connected layers.

4. The transmission line insulator fault diagnosis method based on deep learning according to claim 1, characterized in that: The deep learning-based transmission line insulator fault diagnosis model in step 2 requires the calculation of the cross entropy loss function as shown below: Where: L represents the cross entropy loss function; M represents the total number of fault categories; y i represents the confidence of the i-th fault category; Indicates whether the fault actually occurs in this fault category.

5. The transmission line insulator fault diagnosis method based on deep learning according to claim 1, characterized in that: The ICN deep learning module and the ECNN deep learning module in step 2 are specifically: The ICN deep learning module uses three downsampling convolutional layers and a maximum pooling layer with a stride of 2 and a deconvolution layer corresponding to the downsampling convolution; The ECNN deep learning module contains two convolution layers with a convolution kernel size of 5 and a convolution step size of 1. It uses two maximum pooling series structures with a step size of 2, for a total of four pooling layers.

6. The transmission line insulator fault diagnosis method based on deep learning according to claim 1, characterized in that: In the common attention mechanism layer in step 2, the influence degree of the insulator working condition information is obtained through the fully connected layer and the Sigmoid function, and then the insulator image feature data obtained by the ICN deep learning module and the ECNN deep learning module are combined to obtain the affected integrated feature data and output the fault diagnosis result through the Softmax function.

7. The transmission line insulator fault diagnosis method based on deep learning according to claim 3, characterized in that: The TSAN deep learning module uses a multi-head self-attention mechanism with 8 heads, a feature expansion layer, and a fully connected layer; the TSAN deep learning module can automatically extract fault diagnosis sensitive information from insulator operating condition data. The TSAN deep learning module is trained together with the entire network and can automatically optimize learning parameters; the diagnostic sensitive information of the insulator operating condition data obtained by the TSAN deep learning module can affect the final fault diagnosis result through the common attention mechanism layer.

Images