Vacuum arc arcing mode identification method based on deep learning

Abstract

The invention discloses a method for identifying an arcing mode of a vacuum arc based on deep learning. The method comprises the following steps: S1, acquiring an arcing mode image sequence; s2, establishing a deep learning network and obtaining a network training model; s3, inputting an arc image needing to be identified into the training model for identification; the deep learning network selects an improved GoogleNet network in a convolutional neural network, and the improved Googlenet network structure comprises an input layer, an initial convolutional layer, a pooling layer, an inception module, an auxiliary classifier, a global average pooling layer, a Dropout layer, a full connection layer, a softmax layer and an output layer which are connected in sequence; the vacuum arc identification method based on deep learning mainly aims at solving a series of problems that in the prior art, arcing experiments have a large amount of data including arc dynamic images shot by a high-speed camera, and manual identification is long in time consumption, low in efficiency, high in subjectivity and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical image recognition, and in particular to a vacuum arc burning pattern recognition method based on deep learning. Background Art

[0002] At present, the arc extinguishing medium of high-voltage circuit breakers is mainly SF6, which is a strong greenhouse gas with a global warming potential of 23,500 times that of CO2. More than 10,000 tons of SF6 are used worldwide each year (corresponding to the emissions of 50 million cars). At present, international conventions have clearly restricted the use of SF6.

[0003] Developing environmentally friendly vacuum circuit breakers is an effective measure to reduce SF6 emissions, and the vacuum interrupter is an important component of the vacuum circuit breaker. Whether the vacuum interrupter can successfully interrupt the current is closely related to the evolution of the arcing mode of the vacuum arc.

[0004] At present, arc pattern recognition mainly relies on manual recognition, which has the following two disadvantages: First, arc experiments have a large amount of experimental data, including dynamic images of arcs taken by high-speed cameras. Second, manual recognition of arc images takes a lot of time, is inefficient, and is highly subjective. In addition, vacuum arc discharge is random and uncertain, and has no obvious characteristic features, which increases the difficulty of vacuum arc recognition.

[0005] Convolutional neural network is a kind of artificial neural network, which has become a research hotspot in the current field of speech analysis and image recognition. Its weight-sharing network structure makes it more similar to biological neural network, reducing the complexity of network model and the number of weights. This advantage is more obvious when the network input is a multi-dimensional image, so that the image can be directly used as the input of the network, avoiding the complex feature extraction and data reconstruction process in traditional recognition algorithms. Convolutional neural network (CNN) is the most commonly used type of network for recognizing visual images. Pre-trained convolutional neural network has the ability of transfer learning, which can apply the learned image recognition skills to new vacuum arc recognition scenarios, but the prior art has not seen the use of convolutional neural network for vacuum arc recognition scenarios. Summary of the invention

[0006] In view of the defects existing in the prior art, the purpose of the present invention is to provide a vacuum arc burning pattern recognition method based on deep learning, which is used to solve the problems that manual identification of vacuum arc burning pattern is time-consuming, inefficient and highly subjective.

[0007] In order to achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0008] A vacuum arc burning pattern recognition method based on deep learning comprises the following steps:

[0009] S1. Acquire vacuum arc burning mode image sequence;

[0010] S2. Establish a deep learning network and obtain a training model;

[0011] The deep learning network uses an improved GoogleNet network in a convolutional neural network. The improved GoogleNet network structure includes an input layer, an initial convolution layer, a pooling layer, an inception module, an auxiliary classifier, a global average pooling layer, a Dropout layer, a fully connected layer, a softmax layer, and an output layer connected in sequence.

[0012] S3. Input the vacuum arc burning pattern image to be identified into the training model for identification.

[0013] The training model is used to identify the arcing mode image of the vacuum arc, generate the arcing mode result confidence map, and determine the mode with the highest confidence at a certain moment as the arcing mode.

[0014] The step S3 is specifically as follows: after the vacuum arc burning mode image to be identified is sent to the training model, it is sequentially subjected to resizing and normalization preprocessing, initial feature extraction of the initial convolution layer and the pooling layer, deep feature extraction of the inception module, feature aggregation of the auxiliary classifier and the full average pooling layer, and mapped to the classification space through the Dropout layer and the fully connected layer, and the probability distribution of the four burning modes is output by the softmax layer.

[0015] The step S2 comprises the following steps:

[0016] S11. Establish and initialize a deep learning network;

[0017] S12. Dividing the arcing mode image sequence into training samples and verification samples;

[0018] S13. Setting the training sample and verification sample data, including image cropping, rotation, and translation;

[0019] S14. Set up the training of the deep learning network, including setting the solver to use the CPU for solving, and all other settings are default settings;

[0020] S15. Input training samples to train the deep learning network and perform iterative verification through verification samples to determine whether the deep learning network has reached the set maximum number of iterations. If so, execute S16, otherwise continue to execute S15;

[0021] S16. End the training and output the training model.

[0022] A random rotation angle of -90° to +90° is introduced when setting the training sample and validation sample data.

[0023] Different image brightness is introduced when setting the training sample and validation sample data.

[0024] The specific implementation method of step S15 is: the training sample is input into the deep learning network for forward transmission, the gap between the predicted value and the true label is calculated through the loss function, and then the gradient of the loss relative to each parameter in the network is calculated through the back propagation algorithm, and the parameters of the network are updated according to the gradient until the set maximum number of iterations is reached.

[0025] The 142nd layer of the GoogleNet network is a fully connected layer with an output of 4; the 144th layer replaces the output layer with an output of 4 corresponding to the fully connected layer.

[0026] Step S1 of obtaining the vacuum arc burning mode image sequence is to export the arc burning video captured by the high-speed camera frame by frame into an image sequence.

[0027] The obtained vacuum arc burning mode image sequence is manually divided into four categories, namely, diffusion state mode, point spot mode, anode plate point mode and strong arc mode, and stored separately, with an equal number of images in each category.

[0028] Compared with the prior art, the advantages of the present invention are: the pre-processed deep learning network has the ability of transfer learning, and the learned image recognition ability can be used for new vacuum arc recognition scenarios. At the same time, the brightness and deflection angle of the arc image will not interfere with the results of arc recognition, providing an efficient and accurate solution for vacuum arc recognition. Deep learning technology has excellent recognition and classification capabilities for text and image information, avoiding the problems of time-consuming, low efficiency and strong subjectivity in manual recognition of arc images. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0030] Figure 1 is a flow chart of the method of the present invention;

[0031] Figure 2 This is the GoogleNet network structure diagram;

[0032] Figure 3 This is the structure diagram of the Inception module;

[0033] Figure 4It shows that after 80 iterations, the validation accuracy of the trained deep learning network reached 91.25%;

[0034] Figure 5 Confidence map for single arcing image recognition. DETAILED DESCRIPTION

[0035] 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 only 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.

[0036] The purpose of the present invention is to provide a vacuum arc burning pattern recognition method, which will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0037] like Figure 1 As shown, a vacuum arc burning pattern recognition method based on deep learning is applied to arc burning recognition, and includes the following steps:

[0038] S1. Acquire vacuum arc burning mode image sequence;

[0039] S2. Establish a deep learning network and obtain a training model;

[0040] S3. Sending the vacuum arc burning pattern image to be identified into the training model for identification;

[0041] The deep learning network uses the improved GoogleNet network in the convolutional neural network. Figure 2 As shown, the improved GoogleNet network structure includes an input layer, an initial convolution layer and a pooling layer, an inception module, an auxiliary classifier, a global average pooling layer, a Dropout layer, and a softmax layer connected in sequence.

[0042] Furthermore, step S2 includes:

[0043] S11. Establish and initialize the GoogleNet network;

[0044] S12. Dividing the vacuum arc burning mode image sequence into a training sample and a verification sample;

[0045] S13. Setting the training sample and verification sample data, including image cropping, rotation, translation, etc.;

[0046] S14. Perform training settings for the deep learning network, including image solver settings, normalization, and hardware settings;

[0047] S15. Input training samples to train the deep learning network and perform iterative verification through verification samples to determine whether the deep learning network has reached the set maximum number of iterations. If so, execute S16, otherwise continue to execute S15;

[0048] S16. End the training and output the training model.

[0049] Furthermore, the initial convolution layer and pooling layers conv1, pool1, norm1, conv2, norm2, pool2, inception modules inception(3a), inception(3b), pool3, inception(4a), inception(4b), inception(4c), inception(4d), inception(4e), Auxiliary Classifier 1, pool4, inception(5a), inception(5b), global average pooling layer GlobalAverage Pooling, Dropout, and fully connected layer are arranged between the input layer and the softmax layer.

[0050] Furthermore, conv1 is composed of a 7×7 convolution kernel, conv2 is composed of a 3×3 convolution kernel, the pooling layer is a 3×3 maximum pooling, the format of the input image is 224×224, and the input image is normalized and averaged before entering the initial convolution layer and pooling layer.

[0051] Further, such as Figure 3 As shown in the figure, the inception module of the selected GoogleNet network consists of a 1×1 convolution branch, a 1×1 convolution and 3×3 convolution branch, a 1×1 convolution and 5×5 convolution branch, and a 3×3 pooling and 1×1 convolution branch.

[0052] Furthermore, pool1, pool2, pool3, and pool4 adopt the maximum pooling method.

[0053] Furthermore, the 142th fully connected layer with an output of 1000 of the GoogleNet network is replaced by a fully connected layer with an output of 4; the 144th layer replaces the output layer corresponding to the fully connected layer.

[0054] The first convolution layer uses a 7×7 convolution kernel, 64 channels, and an output of 112×112×64. After convolution, a ReLU operation is performed, and after 3×3 max pooling, the output is 56×56×64, and then a ReLU operation is performed.

[0055] The second convolution layer uses a 3×3 convolution kernel, 192 channels, and an output of 56×56×192. After convolution, a ReLU operation is performed and a 3x3 max pooling is performed. The output is 28×28×192, and then a ReLU operation is performed.

[0056] The third layer is divided into Inception 3a layer and 3b layer. The 3a layer is divided into four branches, using convolution kernels of different scales for processing: (1) 64 1×1 convolution kernels, followed by RuLU, output 28×28×64; (2) 96 1×1 convolution kernels, as dimensionality reduction before 3×3 convolution kernel, become 28×28×96, then perform ReLU calculation, and then perform 128 3×3 convolutions, output 28×28×128; (3) 16 1×1 convolution kernels, as dimensionality reduction before 5×5 convolution kernel, become 28×28×16, after ReLU calculation, perform 32 5×5 convolutions, output 28×28×32; (4) Pool layer, using 3×3 kernel, output 28×28×192, and then perform 32 1×1 convolutions, output 28×28×32. The four results are connected, and the third dimension of the four parts of the output results is connected in parallel, and the final output is 28×28×256; the 3b layer has (1) 128 1×1 convolution kernels, followed by RuLU, and the output is 28×28×128 (2) 128 1×1 convolution kernels, as the dimension reduction before the 3×3 convolution kernel, become 28×28×128, perform ReLU, and then perform 192 3×3 convolutions, and output 28×28×192 (3) 32 1×1 convolution kernels, as the dimension reduction before the 5×5 convolution kernel, become 28×28×32, after ReLU calculation, perform 96 5×5 convolutions, and output 28×28×96 (4) pool layer, using a 3×3 kernel, output 28×28×256, and then perform 64 1×1 convolutions, output 28×28×64. The four results are connected, and the third dimension of the four parts of the output results is connected in parallel, and the final output is 28×28×480.

[0057] The fourth layer (4a, 4b), fifth layer and subsequent layers of the network are similar to the third layer. At the end of the fifth layer, the final output is 7×7×1024.

[0058] The global average pooling layer reduces the spatial size of each feature map to 1×1, reduces the output parameters and prevents overfitting. The Dropout layer randomly discards the activation values ​​of some neurons with a certain probability to prevent overfitting. The fully connected layer flattens all tensors of 1×1×1024 into 1024-dimensional vectors, connects them to a fully connected layer, and finally passes through the softmax layer. The output of the fully connected layer is passed through the function to obtain the probability distribution of the four arcing modes.

[0059] Furthermore, in actual experiments, the shooting angles of arc images are not completely consistent, which can easily lead to misidentification. In order to simulate the actual situation, random rotation angles from -90° to +90° and different image brightnesses of 40%-100% are introduced when setting the sample data. The confidence levels of the recognition results generated when rotating 90°, +45°, and -45° are 99.95%, 99.58%, and 99.85%, respectively. The confidence levels generated when recognizing arc images with brightness of 100%, 70%, and 40% are 99.96%, 99.98%, and 99.95%, respectively.

[0060] Furthermore, during iterative training, the training samples are input into the deep learning network for forward propagation to obtain the predicted output of each image. The difference between the predicted value and the true label is calculated through the loss function, and the gradient of the loss relative to each parameter in the network is calculated through the back-propagation algorithm. The parameters of the network are updated according to the gradient until the maximum number of iterations is reached. After every 10 steps of iteration, the verification samples are input into the training network for forward propagation to verify the network performance; the parameters of the model are continuously updated through forward propagation and back-propagation, so that the output results of the training model are closer and closer to the actual labels.

[0061] Furthermore, the specific implementation method of the forward pass is as follows: The key module Inception module in the GoogleNet network contains multiple convolutional layers and pooling layers. The input arc image is converted into an array input convolutional layer through the input layer, and the convolution operation is performed with the convolution kernel to calculate the feature map, and then the convolution result is nonlinearly mapped through the ReLU function. The pooling layer is inserted between specific layers to pool the feature map of the convolution output, reduce the dimension and reduce the amount of calculation, and extract the most significant features. The feature maps obtained by all different convolution and pooling operations are spliced ​​together in the depth dimension to form the output of the module. After undergoing the convolution and pooling operations of multiple Inception modules, it is connected to a fully connected layer to integrate the extracted feature information. Finally, in the output layer, the softmax function is used to convert the output into the probability distribution of the four arcing modes.

[0062] Furthermore, the specific implementation of the reverse transfer is as follows: the output layer gradient is transmitted back to the convolution layer using the chain rule. In the Inception module, the gradient of each path is calculated separately, and for each convolution and pooling operation, the chain rule is used to calculate the gradient of its output to the loss, and the gradient of the previous layer is propagated to each convolution layer and pooling layer, and the contribution of the convolution kernel and pooling kernel to the loss is calculated respectively. After calculating the corresponding gradient, the parameters of all layers are updated according to the determined learning rate.

[0063] Furthermore, Figure 4The figure shows the accuracy of the network after training. After 40 iterations, the network's verification accuracy reaches more than 90%, which is a high accuracy.

[0064] Furthermore, Figure 5 This is the confidence map generated for the image recognition result. The confidence levels of the four arcing modes in this image are shown in the figure. The diffusion mode has the highest confidence level, so the recognition result is the diffusion mode.

[0065] In summary, the present invention provides a vacuum arc burning pattern recognition method based on deep learning, which can quickly identify the duration and evolution process of the arc burning pattern, and generate the corresponding relationship between the corresponding current and voltage parameters and the arc burning pattern at any time. The efficiency is much higher than manual recognition, and the probability of misjudgment is lower.

[0066] It should be noted that, for the above-mentioned various method embodiments, for the sake of simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, some steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and units involved are not necessarily required by this application.

Claims

1. A vacuum arc burning pattern recognition method based on deep learning, characterized in that: The following steps are involved: S1. Acquire vacuum arc burning mode image sequence; S2. Establish a deep learning network and obtain a training model; The deep learning network uses an improved GoogleNet network in a convolutional neural network. The improved GoogleNet network structure includes an input layer, an initial convolution layer, a pooling layer, an inception module, an auxiliary classifier, a global average pooling layer, a Dropout layer, a fully connected layer, a softmax layer, and an output layer connected in sequence. S3. Input the vacuum arc burning pattern image to be identified into the training model for identification; The training model is used to identify the arcing mode images of vacuum arcs, generate the arcing mode result confidence map, and determine the mode with the highest confidence at a certain moment as the arcing mode.

2. A vacuum arc burning pattern recognition method based on deep learning as claimed in claim 1, characterized in that: The step S3 is specifically as follows: after the vacuum arc burning mode image to be identified is sent to the training model, it is sequentially subjected to resizing and normalization preprocessing, initial feature extraction of the initial convolution layer and the pooling layer, deep feature extraction of the inception module, feature aggregation of the auxiliary classifier and the full average pooling layer, and mapped to the classification space through the Dropout layer and the fully connected layer, and the probability distribution of the four burning modes is output by the softmax layer.

3. The vacuum arc burning pattern recognition method based on deep learning according to claim 1, characterized in that: The step S2 comprises the following steps: S11. Establish and initialize a deep learning network; S12. Dividing the arcing mode image sequence into training samples and verification samples; S13. Setting the training sample and verification sample data, including image cropping, rotation, and translation; S14. Set up the training of the deep learning network, including setting the solver to use the CPU for solving, and all other settings are default settings; S15. Input training samples to train the deep learning network and perform iterative verification through verification samples to determine whether the deep learning network has reached the set maximum number of iterations. If so, execute S16, otherwise continue to execute S15; S16. End the training and output the training model.

4. A vacuum arc burning pattern recognition method based on deep learning as claimed in claim 3, characterized in that: A random rotation angle of -90° to +90° is introduced when setting the training sample and validation sample data.

5. A vacuum arc burning pattern recognition method based on deep learning as claimed in claim 3, characterized in that: Different image brightness is introduced when setting the training sample and validation sample data.

6. A vacuum arc burning pattern recognition method based on deep learning as claimed in claim 3, characterized in that: The specific implementation method of step S15 is: the training sample is input into the deep learning network for forward transmission, the gap between the predicted value and the true label is calculated through the loss function, and then the gradient of the loss relative to each parameter in the network is calculated through the back propagation algorithm, and the parameters of the network are updated according to the gradient until the set maximum number of iterations is reached.

7. A vacuum arc burning pattern recognition method based on deep learning as claimed in claim 1, characterized in that: The 142nd layer of the GoogleNet network is a fully connected layer with an output of 4; the 144th layer replaces the output layer with an output of 4 corresponding to the fully connected layer.

8. The vacuum arc burning pattern recognition method based on deep learning according to claim 1, characterized in that: Step S1 of obtaining the vacuum arc burning mode image sequence is to export the arc burning video captured by the high-speed camera frame by frame into an image sequence.

9. A vacuum arc burning pattern recognition method based on deep learning as claimed in claim 8, characterized in that: The obtained vacuum arc burning mode image sequence is manually divided into four categories, namely, diffusion state mode, point spot mode, anode plate point mode and strong arc mode, and stored separately, with an equal number of images in each category.

Images