Fault Arc Protection Method Based on AlexNet

Abstract

The present invention discloses a fault arc protection method based on AlexNet, and the method comprises the following steps: Step 1, generating a fault arc data set; Step 2, establishing a fault arc diagnosis model structure and training parameters; Step 3, diagnosing the fault arc. The beneficial effect of the present invention is that it has extremely high recognition accuracy and can achieve adaptive protection for fault arcs.

Description

Technical Field

[0001] The present invention relates to the technical field of fault protection, in particular to a fault arc protection method based on AlexNet. Background Art

[0002] The time-frequency domain characteristics of fault arcs for different types of loads are not exactly the same; even for the same type of load, the time-frequency domain characteristics of arc currents at different times may also vary. In practical applications, it is difficult to enumerate the time-frequency domain characteristics of fault arcs under each load condition. Therefore, the traditional fault arc protection method based on arc characteristics and a preset threshold has relatively poor adaptability and is difficult to meet the new requirements for fault arc identification and protection under variable non-linear load conditions or unknown loads. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and design a fault arc protection method based on AlexNet.

[0004] The fault arc protection method based on AlexNet includes the following steps:

[0005] Step 1, generating a fault arc data set, and the process is as follows:

[0006] (1) Data cleaning

[0007] There is diversity in fault arc tests, and the test data is collected from multiple laboratories and at multiple time points. These objective reasons may cause abnormal points in the arc test data. In order to eliminate the inconsistency of similar data and improve the data quality, it is necessary to "clean" the test data through manual intervention before training and delete the abnormal data points in the data. The types of data that need to be cleaned mainly include data with zero sampling due to improper sampling device settings or other data that is obviously irrelevant to the arc current signal;

[0008] (2) Data segmentation

[0009] In order to increase the number of training samples and improve the generalization ability of the fault arc diagnosis model, the data is segmented by means of sliding window sampling. At the same time, in order to ensure the consistency of the samples, it is stipulated that all samples are.csv files with a fixed number of sampling points. Taking every 4 power frequency cycle current data as a sample, the sliding offset is 1 power frequency cycle, the test sampling rate is 100 Ks / s, and each sample includes 8000 sampling points;

[0010] (3) Data annotation

[0011] After the data cleaning and data segmentation processes, there are 5000 groups of single-load fault arc data and normal state data each; before model training, the data set is divided into a test set, a training set, and a validation set according to a certain proportion;

[0012] All normal state data is stored in the folder labeled "normal", and the faulty arc state data is stored in the folder labeled "faulty". At this time, the data is in.csv format; all.csv format data is converted into an image format of 227×227×3. 227×227 represents the image pixels, and 3 represents the RGB three-channel image input;

[0013] Step 2: Establish the structure and training parameters of the faulty arc diagnosis model. The process is as follows:

[0014] The faulty arc diagnosis model is established based on the optimization of the classical AlexNet model. The faulty arc diagnosis model optimized based on AlexNet consists of 6 convolutional layers, 3 pooling layers, and 3 fully connected layers. The input of the model is the faulty arc data set, and the output is the line state;

[0015] The parameter configuration used in the training process is as follows: The activation function selects the ReLU function, uses dropout regularization, and the dropout rate is set to 0.5; the optimization algorithm is the SGD algorithm, the initial learning rate is 0.01, and if the accuracy of the validation set no longer changes with the increase of the number of iterations, the learning rate is reduced to one-tenth of the previous value; the loss function is the cross-entropy loss function applicable to binary classification problems;

[0016] Step 3: Diagnose the faulty arc. The diagnosis process of the faulty arc diagnosis model based on AlexNet is as follows:

[0017] (1) Perform data preprocessing on the collected current signal to obtain the faulty arc data set. The data is divided into three groups according to a certain ratio, which are used as the training set, validation set, and test set respectively;

[0018] (2) Initialize the network parameters. The network model randomly extracts a sample from the training set, calculates the output sequentially from front to back, adjusts the network parameters and updates the network model through backpropagation of the error. Then input the next sample until the error reaches the minimum, and then output the faulty arc diagnosis model;

[0019] (3) Input the test set into the faulty arc diagnosis network model to obtain the recognition accuracy rate, that is, the recognition result of the model can be evaluated.

[0020] The classical AlexNet model includes 5 convolutional layers and 3 fully connected layers.

[0021] The first convolutional layer of the classical AlexNet model uses 96 convolutional kernels of size 11×11 to extract image features of size 227×227 pixels; the second convolutional layer contains 256 convolutional kernels of size 5×5; the third and fourth convolutional layers each include 384 kernels of size 3×3, and the fifth convolutional layer has 256 kernels of size 3×3.

[0022] In the fault arc diagnosis model optimized based on AlexNet, the 5×5 convolutional kernels in the second convolutional layer of the classical AlexNet model are replaced with two 3×3 convolutional kernels.

[0023] Beneficial effects

[0024] The fault arc protection method based on AlexNet made by using the technical solution of the present invention has the following advantages:

[0025] This method has an identification accuracy higher than 96%. It can accurately identify fault arcs of irrelevant feature quantities and load types on the premise of ensuring the complete original information of arc current, solve the problems of difficult extraction of feature quantities and complex threshold setting caused by the complex fault arc current characteristics under non-linear load conditions, and can realize adaptive protection of fault arcs. Brief description of the drawings

[0026] Figure 1 is a schematic flow chart of the fault arc protection method based on AlexNet of the present invention;

[0027] Figure 2 is a structure parameter table of the AlexNet fault arc diagnosis model of the present invention;

[0028] Figure 3 is a structure diagram of the fault arc diagnosis model of the present invention;

[0029] Figure 4 is a parameter setting table of the arc fault diagnosis model of the present invention;

[0030] Figure 5 is a division table of datasets with different ratios of the present invention;

[0031] Figure 6 is a relationship diagram of batch size, training duration and accuracy of the present invention; Detailed implementation manners

[0032] The present invention will be specifically described below with reference to the drawings, as Figure 1-6 shown;

[0033] The creative point of this application is to generate a fault arc dataset, and the process is as follows:

[0034] (1) Data cleaning

[0035] The fault arc test has diversity. The test data is collected from multiple laboratories and multiple time points. These objective reasons may cause abnormal points in the arc test data. In order to eliminate the inconsistency of similar data and improve the data quality, it is necessary to "clean" the test data through manual intervention before training and delete the abnormal data points in the data. The data types that need to be cleaned mainly include the data with zero sampling caused by improper setting of the sampling device or other data that is obviously irrelevant to the arc current signal;

[0036] (2) Data segmentation

[0037] In order to increase the number of training samples and improve the generalization ability of the fault arc diagnosis model, the data is segmented in the way of sliding window sampling. At the same time, in order to ensure the consistency of the samples, it is stipulated that all samples are.csv files with a fixed number of sampling points. Taking the current data of every 4 power frequency cycles as a sample, the sliding offset is 1 power frequency cycle, the test sampling rate is 100 Ks / s, and each sample contains 8000 sampling points;

[0038] (3) Data annotation

[0039] After the data cleaning and data segmentation processes, there are 5000 groups of single-load fault arc data and normal state data respectively; before model training, the data set is divided into a test set, a training set and a validation set according to a certain proportion;

[0040] All normal state data is stored in the folder labeled "normal", and the fault arc state data is stored in the folder labeled "fault". At this time, the data is in.csv format; convert all.csv format data into an image format of 227×227×3. 227×227 represents the image pixels, and 3 represents the RGB three-channel image input;

[0041] The creative point of this application also lies in establishing the fault arc diagnosis model structure and training parameters, and the process is as follows:

[0042] The fault arc diagnosis model is established based on the optimization of the classic AlexNet model. The fault arc diagnosis model optimized based on AlexNet is composed of 6 convolutional layers, 3 pooling layers and 3 fully connected layers. The input of the model is the fault arc data set, and the output is the line state;

[0043] The parameter configuration used in the training process is as follows: The ReLU function is selected as the activation function, dropout regularization is used, and the dropout rate is set to 0.5; The optimization algorithm is the SGD algorithm, and the initial learning rate is 0.01. If the accuracy of the validation set no longer changes with the increase of the number of iterations, the learning rate is reduced to one-tenth of the previous value; The loss function is the cross-entropy loss function applicable to binary classification problems;

[0044] The creative point of this application also lies in diagnosing the faulty arc. The diagnostic process of the faulty arc diagnosis model based on AlexNet is as follows:

[0045] (1) Preprocess the collected current signal to obtain the faulty arc data set. Divide the data into three groups according to a certain ratio, which are used as the training set, validation set, and test set respectively;

[0046] (2) Initialize the network parameters. The network model randomly extracts a sample from the training set, calculates the output sequentially from front to back, adjusts the network parameters and updates the network model through backpropagation of errors. Then input the next sample until the error reaches the minimum, and then output the faulty arc diagnosis model;

[0047] (3) Input the test set into the faulty arc diagnosis network model to obtain the recognition accuracy rate, that is, the recognition result of the model can be evaluated.

[0048] The creative point of this application also lies in that the classic AlexNet model includes 5 convolutional layers and 3 fully connected layers; The first convolutional layer of the classic AlexNet model uses 96 convolutional kernels of size 11×11 to extract image features with a pixel size of 227×227; The second convolutional layer contains 256 convolutional kernels of size 5×5; The third and fourth convolutional layers each include 384 kernels of size 3×3, and the fifth convolutional layer has 256 kernels of size 3×3; The faulty arc diagnosis model optimized based on AlexNet replaces the 5×5 convolutional kernels in the second convolutional layer of the classic AlexNet model with two 3×3 convolutional kernels.

[0049] In the implementation process of the technical solution of this application, the classic AlexNet model includes 5 convolutional layers and 3 fully connected layers; The network input is the image pixels, and the output is the classification result.

[0050] The first convolutional layer of the network structure of the classic AlexNet model uses 96 convolutional kernels of size 11×11 to extract image features with a pixel size of 227×227. The second convolutional layer contains 256 convolutional kernels of size 5×5. The third and fourth convolutional layers each include 384 kernels of size 3×3, and the fifth convolutional layer has 256 kernels of size 3×3..

[0051] Although the classic AlexNet model does not have deep network stacking, it has strong image recognition ability and is not prone to overfitting. The advantages of its training strategy are as follows:

[0052] (1) Application of dropout regularization

[0053] To prevent overfitting and reduce the time and labor costs for training, the idea of regularization (dropout) is adopted. Simply put, the role of dropout is to force certain neurons to stop working with a certain probability p and hide the feature expression. Dropout regularization can achieve better training results mainly for the following two reasons:

[0054] 1) Dropout randomly suppresses the feature expressions of different neurons. Each training process is equivalent to taking the average of the classification results of multiple different-structured networks, and the output result no longer depends on the expressions of some local specific features;

[0055] (2) The application of Dropout makes the weight update process no longer dominated by some nodes with fixed connection relationships. Therefore, the network learning process is not overly sensitive to some specific features, but learns more robust features. Generally, the dropout rate is set to 0.5, aiming to generate the most network structures with fewer parameters.

[0056] (2) Application of the ReLU non-linear activation function

[0057] The commonly used non-linear activation functions in CNN are the Sigmoid function, the hyperbolic tangent function tanh, and the ReLU function.

[0058] The expression of the Sigmoid function is:

[0059]

[0060] The expression of the hyperbolic tangent function tanh is:

[0061]

[0062] The expression of the rectified linear unit RuLE is:

[0063] f(x) = max(0, x).

[0064] Both the Sigmoid function and the tanh function are saturated non - linear functions, but they both have the problems of slow convergence speed and gradient dispersion. The derivative values of the Sigmoid and tanh functions are close to 0 when the input value is large, resulting in gradient dispersion during the weight update process, which causes the error value not to be propagated downward and the learning ability of the underlying network to be insufficient. Compared with saturated non - linear functions, the non - saturated function ReLU can solve the problem of gradient disappearance, and at the same time can accelerate the convergence speed, significantly improving the performance of CNN. The activation function in the AlexNet model is the ReLU function, which can set the numbers less than zero to zero and keep the numbers greater than zero unchanged, having sparsity.

[0065] (3) Application of Stochastic Gradient Descent Optimization Algorithm

[0066] The main idea of the gradient descent algorithm is to find the minimum value of the loss function from the gradient direction or the opposite direction of the loss function through the backpropagation algorithm, that is, to minimize the error between the output value and the actual value. The Stochastic Gradient Descent (SGD) method initializes the weights by selecting a small batch of samples (batch) at each iteration, calculates the gradient of the loss function and updates the weight coefficients until the value of the loss function reaches the minimum and then stops updating. Since SGD randomly selects a group of samples for weight update each time, the computational amount is significantly reduced. Fault arc recognition is a binary classification problem, that is, the prediction results are only two results: "normal" and "fault". Therefore, the loss function in the SGD algorithm is the cross - entropy loss function (Cross - Entropy loss) applicable to binary classification problems, as shown in the following formula

[0067] L(y,f(x;θ)=-logf y (x;θ)

[0068] Among them, L is the cross - entropy loss function. x is the input sample, θ is the coefficient to be updated, y is the label vector of the sample, and fy(x;θ) is the label probability distribution predicted by the model. The weight update of SGD is shown in the following formula:

[0069]

[0070] Among them, θt is the parameter value at the t - th iteration, L(θ) is the loss function, and α is the learning rate.

[0071] The convergence speed of the SGD algorithm is related to the learning rate. If the learning rate is set appropriately, it will accelerate the convergence speed of the algorithm. The learning rate of the fault arc diagnosis model in the paper is adjustable. If the accuracy of the validation set no longer changes, the learning rate is reduced to 1 / 10 of the previous value, and the learning rates of all layers are the same.

[0072] The generation of the fault arc dataset includes the following three steps:

[0073] (1) Data cleaning

[0074] The fault arc test has diversity. The test data is collected from multiple laboratories and at multiple time points. These objective reasons may cause abnormal points in the arc test data. In order to eliminate the inconsistency of similar data and improve the data quality, it is necessary to "clean" the test data through manual intervention before training and delete the abnormal data points in the data. The data types that need to be cleaned mainly include data with zero sampling due to improper sampling device settings or other data that is obviously irrelevant to the arc current signal.

[0075] (2) Data segmentation

[0076] In order to increase the number of training samples and improve the generalization ability of the fault arc diagnosis model, the data is segmented by means of sliding window sampling. At the same time, in order to ensure the consistency of the samples, it is stipulated that all samples are.csv files with a fixed number of sampling points.

[0077] Taking into account the real-time requirements of fault arc diagnosis, every 4 power frequency cycle current data is used as a sample, and the sliding offset is 1 power frequency cycle. According to what is described in Section 2.2.1, the test sampling rate is 100 Ks / s, and each sample contains 8000 sampling points.

[0078] (3) Data annotation

[0079] After the data cleaning and data segmentation processes, there are 5000 groups of single-load fault arc data and normal state data respectively. Before model training, the data set is divided into a test set, a training set and a validation set according to a certain ratio.

[0080] All normal state data is stored in a folder labeled "normal", and the fault arc state data is stored in a folder labeled "fault". At this time, the data is in.csv format. Convert all.csv format data into an image format of 227×227×3. 227×227 represents the image pixels, and 3 represents the RGB three-channel image input. Take the resistor load and the induction cooker load as examples.

[0081] The input data of the fault arc diagnosis model based on AlexNet undergoes a "signal-image" conversion process, retains the characteristics of the original current signal, and reduces the dependence on experience in the process of manually extracting feature quantities. At the same time, through the data cleaning process, the interference of noise on the arc current is reduced, which is conducive to improving the recognition accuracy of the diagnosis model.

[0082] Although the classic AlexNet has a simple network model structure and is applicable to the field of fault arc diagnosis, it has two relatively large convolutional kernels, resulting in a large number of model parameters. Therefore, the optimization method of the fault arc diagnosis model based on AlexNet is to optimize the network architecture and reduce network parameters on the premise of meeting the fault recognition accuracy rate.

[0083] Related research shows that using multiple small-sized convolutional kernels stacked instead of a large convolutional kernel for feature extraction can save network parameters and accelerate the network convergence speed without affecting the data feature expression, thus shortening the training duration. The effect is significant when the connectivity remains unchanged. At the same time, the Inception module structure used in GoogLeNet also shows that replacing a large convolutional kernel with a stack of small convolutional kernels can improve the classification accuracy. Therefore, in order to reduce model parameters, enhance the non-linear expression ability of the model, and improve the recognition accuracy, drawing on the idea of the Inception series network, small convolutional kernels are used to replace large convolutional kernels to optimize the architecture of the classic AlexNet model.

[0084] The classic AlexNet model has two large convolutional kernels. One is an 11×11 convolutional kernel used to extract general features such as the underlying texture of images, which has little relevance to the fault arc current features. Therefore, its parameters and structure are retained. Analyzing from the convolution principle, using a 3×3 convolutional kernel to perform convolution twice continuously can achieve the feature extraction ability of a 5×5 convolutional kernel in one convolution, that is, the results obtained from these two convolution processes both reflect the features within the same pixel size in the original image. The convolution kernel substitution process is as Figure 3 shown. A 5×5 convolutional kernel has 25 weights to be trained, while using two 3×3 convolutional kernels with a stride of 1, the number of weights is reduced to 18 (3×3×2), saving 28% of the computational cost. Since the activation of the ReLU function is required in both consecutive convolution processes, more non-linear feature expressions are obtained compared to before the convolution kernel substitution, which is beneficial for more comprehensive and accurate extraction of arc current features, thereby improving the recognition accuracy of the fault arc diagnosis model. Therefore, the 5×5 convolutional kernel in the second convolutional layer of the AlexNet network model is replaced with two 3×3 convolutional kernels.

[0085] The fault arc diagnosis model based on the classic AlexNet model consists of 6 convolutional layers, 3 pooling layers, and 3 fully connected layers. The structural parameters are as Figure 2 shown. The input of the model is the fault arc data set, and the output is the line state. The structure diagram of the fault arc diagnosis model is as Figure 3 shown.

[0086] In addition to building the network structure model, it is also necessary to configure the parameters during the training process. The ReLU function is selected as the activation function, dropout regularization is used, the dropout rate is set to 0.5, the optimization algorithm is the SGD algorithm, and the initial learning rate is 0.01. If the accuracy of the validation set no longer changes as the number of iterations increases, the learning rate is reduced to one-tenth of the previous value. The loss function is the cross-entropy loss function applicable to binary classification problems. Other parameters such as the batch size and the number of iterations are set as Figure 4 shown.

[0087] The batch size is the size of the random sample used for each gradient optimization, which is related to the training duration and recognition accuracy of the network. When the batch size is large, the time to train a complete dataset is reduced and the convergence speed is accelerated, but the number of parameter updates will become smaller, which is not conducive to improving the arc recognition accuracy. On the contrary, it will increase the time cost of training. Therefore, it is necessary to find the optimal batch size to balance the relationship between the training duration and the fault arc recognition rate. Seven batch sizes of 20, 40, 60, 80, 100, 120, and 140 are selected respectively to analyze the changing trends with the training duration and recognition accuracy, as Figure 6 shown. When the batch size increases, the fault arc recognition accuracy shows a trend of first rising and then falling, and the training duration shows a trend of rapidly shortening to slowly shortening. When the batch size is 100, the recognition accuracy is the highest, and the training time is about 1.7 hours. Compared with when the batch size is 120, although the training duration increases by about 0.1 hour, the accuracy increases by about 3%. During the process of increasing the batch size from 120 to 140, the fault arc recognition accuracy decreases significantly. When the fluctuation range of the training duration is small, the recognition accuracy of the model should be ensured. Considering the time cost and the fault arc recognition accuracy comprehensively, the batch size adopted by the fault arc diagnosis model is 100.

[0088] The specific steps of the fault arc diagnosis process are as follows:

[0089] (1) Preprocess the collected current signal to obtain the fault arc dataset. Divide the data into three groups according to a certain ratio, which are used as the training set, validation set, and test set respectively;

[0090] (2) Initialize the network parameters. The network model randomly extracts a sample from the training set, calculates the output sequentially from front to back, and adjusts the network parameters and updates the network model through backpropagation of the error. Then input the next sample until the error reaches the minimum, and then output the fault arc diagnosis model;

[0091] (3) Input the test set into the fault arc diagnosis network model to obtain the recognition accuracy rate, that is, the recognition result of the model can be evaluated.

[0092] To verify the performance of the improved AlexNet fault arc diagnosis model, when the computer RAM is 16.00GB and equipped with an NVIDIA GeForce GTX 1060 graphics card, train the AlexNet models before and after improvement respectively and analyze and compare the fault arc recognition results. The training set determines what features the fault arc diagnosis model learns. The accuracy rate of the validation set is used to preliminarily evaluate the recognition ability of the model and determine the parameters such as the learning rate and the number of iterations that cannot be automatically optimized through network learning and can only be determined manually according to experience. That is, both the training set and the validation set are related to the parameter optimization process and determine the feature learning ability of the network. The accuracy rate of the test set is the index for evaluating the precision of the model. The division of the data set in different proportions is as Figure 5 shown.

[0093] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation. An element defined by the statement "comprising a..." does not exclude the existence of additional identical elements in the process, method, article or device comprising the element.

[0094] The above technical solutions only reflect the preferred technical solutions of the technical solutions of the present invention. Some changes that those skilled in the art of this technology may make to some parts thereof all reflect the principles of the present invention and are within the protection scope of the present invention.

Claims

1. Fault arc protection method based on AlexNet, characterized in that, the method comprises the following steps: Step 1, generating a fault arc data set, the process is as follows: (1) Data cleaning There is diversity in fault arc tests. The test data is collected from multiple laboratories and at multiple time points. These objective reasons may cause abnormal points in the arc test data. In order to eliminate the inconsistency of similar data and improve data quality, it is necessary to "clean" the test data through manual intervention before training, delete the abnormal data points in the data. The data types that need to be cleaned mainly include the data with zero sampling caused by improper setting of the sampling device or other data that is obviously irrelevant to the arc current signal; (2) Data segmentation In order to increase the number of training samples and improve the generalization ability of the fault arc diagnosis model, the data is segmented by means of sliding window sampling. At the same time, in order to ensure the consistency of the samples, it is stipulated that all samples are.csv files with a fixed number of sampling points. Taking every 4 power frequency cycle current data as a sample, the sliding offset is 1 power frequency cycle, the test sampling rate is 100 Ks / s, and each sample contains 8000 sampling points; (3) Data annotation After the data cleaning and data segmentation processes, there are 5000 groups of single-load fault arc data and normal state data respectively; before model training, the data set is divided into a test set, a training set and a validation set according to a certain proportion; All normal state data is stored in the folder labeled "normal", and the fault arc state data is stored in the folder labeled "fault". At this time, the data is in.csv format; convert all.csv format data into an image format of 227×227×3. 227×227 represents the image pixels, and 3 represents the RGB three-channel image input; Step 2, establishing the fault arc diagnosis model structure and training parameters, the process is as follows: The fault arc diagnosis model is established based on the optimization of the classical AlexNet model. The fault arc diagnosis model optimized based on AlexNet is composed of 6 convolutional layers, 3 pooling layers and 3 fully connected layers. The input of the model is the fault arc data set, and the output is the line state; The parameter configuration used in the training process is as follows: the activation function selects the ReLU function, uses dropout regularization, and the dropout rate is set to 0.5; the optimization algorithm is the SGD algorithm, the initial learning rate is 0.01, if the accuracy of the validation set no longer changes with the increase of the number of iterations, the learning rate is reduced to one-tenth of the previous value; the loss function is the cross-entropy loss function applicable to binary classification problems; Step 3, diagnosing the fault arc. The diagnosis process of the fault arc diagnosis model based on AlexNet is as follows: (1) Preprocess the collected current signal to obtain a fault arc data set, and divide the data into three groups according to a certain proportion, which are used as the training set, the validation set and the test set respectively; (2) Initialize network parameters. The network model randomly extracts a sample from the training set, calculates the output sequentially from front to back, adjusts the network parameters and updates the network model through backpropagation of errors, then inputs the next sample until the error reaches the minimum, and then outputs the fault arc diagnosis model; (3) Input the test set into the fault arc diagnosis network model to obtain the recognition accuracy rate, that is, the recognition result of the model can be evaluated.

2. The fault arc protection method based on AlexNet according to claim 1, characterized in that, the classical AlexNet model includes 5 convolutional layers and 3 fully connected layers.

3. The fault arc protection method based on AlexNet according to claim 2, characterized in that, the first convolutional layer of the classical AlexNet model uses 96 convolutional kernels of size 11×11 to extract image features with a pixel size of 227×227; the second convolutional layer contains 256 convolutional kernels of size 5×5; the third and fourth convolutional layers each include 384 kernels of size 3×3, and the fifth convolutional layer has 256 kernels of size 3×3.

4. The fault arc protection method based on AlexNet according to claim 1, characterized in that, the fault arc diagnosis model optimized based on AlexNet replaces the 5×5 convolutional kernels in the second convolutional layer of the classical AlexNet model with two 3×3 convolutional kernels.

Images