Vacuum breakdown type detection method and device based on convolutional neural network

Abstract

The invention discloses a vacuum breakdown category detection method and device based on a convolutional neural network, and relates to the field of vacuum breakdown detection.The method comprises the steps that pre-breakdown current waveforms of different vacuum breakdown categories are obtained to serve as training samples, the vacuum breakdown types comprise pulse current induced vacuum breakdown, field emission induced vacuum breakdown and particle induced vacuum breakdown; constructing a vacuum breakdown category diagnosis model based on the convolutional neural network, wherein the model comprises a shared convolutional layer, a multi-task processing layer, a full connection layer and a SoftMax function; inputting a training sample to train the vacuum breakdown category diagnosis model, and obtaining a weight parameter with the best training effect and loading the weight parameter to the model; and inputting to-be-classified pre-breakdown current waveforms into the vacuum breakdown category diagnosis model to obtain probabilities of three vacuum breakdown categories, and selecting the vacuum breakdown category corresponding to the maximum probability as a final diagnosis result to be output. According to the method, the diagnosis speed and efficiency of the vacuum breakdown type in engineering practice are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of vacuum breakdown detection, and in particular to a vacuum breakdown category detection method and device based on a convolutional neural network. Background Art

[0002] At present, although vacuum circuit breakers have been initially widely used in the actual production of medium and low voltage power switches, in high voltage environments, the application of vacuum circuit breakers is limited due to the weakening of the insulation performance of the electrode gap. At present, SF 6 Gas circuit breakers. However, SF 6 Gas is an environmentally unfriendly gas. Its strong greenhouse effect places extremely high demands on sealing. At the same time, the acidic compounds formed by its decomposition also increase the possibility of corrosive acid rain. In order to actively respond to my country's "dual carbon" goals, realize the application of vacuum circuit breakers in high-voltage environments, and replace SF 6 Gas circuit breakers are needed immediately.

[0003] The reason why vacuum circuit breakers are not suitable for working under high voltage environment is that vacuum circuit breakers are more likely to have vacuum breakdown under high voltage. The study of vacuum breakdown categories can reveal the mechanism of vacuum breakdown of electrode gaps. Therefore, diagnosis of vacuum breakdown categories is of great significance for improving the voltage level of vacuum circuit breakers. At present, the diagnosis of vacuum breakdown categories relies on manual judgment and computer calculation, which takes a long time and is complicated. Summary of the invention

[0004] In view of the above problems and technical requirements, the inventors have proposed a vacuum breakdown category detection method and device based on a convolutional neural network. The technical solution of the present invention is as follows:

[0005] In a first aspect, the present application provides a vacuum breakdown category detection method based on a convolutional neural network, comprising the following steps:

[0006] Obtaining pre-breakdown current waveforms of different vacuum breakdown categories as training samples, the vacuum breakdown categories including pulse current induced vacuum breakdown, field emission induced vacuum breakdown and particle induced vacuum breakdown;

[0007] Construct a vacuum breakdown classification diagnosis model based on convolutional neural network, including shared convolutional layer, multi-task processing layer, fully connected layer and SoftMax function;

[0008] The shared convolution layer is used to extract the basic features of the pre-breakdown current waveforms of the three vacuum breakdown categories and input them into the multi-task processing layer for processing;

[0009] The multi-task processing layer is used to identify the characteristics of different vacuum breakdown categories in the basic features, concatenate them to obtain a composite feature vector, and then calculate the probabilities of the three vacuum breakdown categories through the fully connected layer and the SoftMax function;

[0010] Taking the processed pre-breakdown current waveform as input and the vacuum breakdown category corresponding to the waveform as output, the vacuum breakdown category diagnosis model based on the convolutional neural network is trained, and the weight parameters with the best training effect are loaded into the vacuum breakdown category diagnosis model;

[0011] The pre-breakdown current waveform to be classified is input into the vacuum breakdown category diagnosis model to obtain the probabilities of three vacuum breakdown categories, and the vacuum breakdown category corresponding to the maximum probability is selected as the final diagnosis result output.

[0012] A further technical solution is that the shared convolution layer includes two convolution and pooling layers, where:

[0013] The first convolution pooling layer uses 32 3*3 convolution kernels, and then passes through a 2*2 pooling window; the second convolution pooling layer uses 64 3*3 convolution kernels, and then passes through a 2*2 pooling window.

[0014] Its further technical solution is that the multi-task processing layer includes a parallel feature branch network, a tail segment enhancement module, and a global pooling module;

[0015] The characteristic branch network is used to capture and amplify the rising trend of the pre-breakdown current waveform of the field emission-induced vacuum breakdown and the falling trend of the pre-breakdown current waveform of the particle-induced vacuum breakdown, and extract the rising segment features and the falling segment features;

[0016] The tail enhancement module is used to capture and amplify the changing trend of the tail segment of the waveform and extract the tail segment features;

[0017] The global pooling module is used to extract the global features of the three pre-breakdown current waveforms.

[0018] A further technical solution is that the feature branch network includes:

[0019] The ascending convolution module is used to perform convolution operations on each channel of the input basic features using a 3*3 convolution kernel, calculate the weighted sum, and then flatten it into a one-dimensional vector through a 2*2 pooling window to obtain the ascending segment features;

[0020] The descending segment convolution module is used to perform a convolution operation on each channel of the input basic features using a 3*3 convolution kernel, calculate the weighted sum, and then flatten it into a one-dimensional vector through a 2*2 pooling window to obtain the descending segment features;

[0021] Among them, the convolution kernel of the ascending convolution module represents that each row of the matrix is ​​[-1, 0, 1]; the convolution kernel of the descending convolution module represents that each row of the matrix is ​​[1, 0, -1].

[0022] The further technical solution is that the tail enhancement module includes clipping, tail convolution layer and pooling flattening layer, wherein:

[0023] By cropping the tail region that retains the basic features of the input, the tail convolution layer uses 16 3*3 convolution kernels to extract the tail region features and flattens them into a one-dimensional vector through the pooling flattening layer.

[0024] Its further technical solution is that the fully connected layer adopts a logic score calculation formula, which is expressed as: the input composite feature vector is multiplied by the weight matrix, and the result is summed with the bias matrix to obtain the logic scores of the three vacuum breakdown categories; each row element of the weight matrix corresponds to the weight parameters of the three vacuum breakdown categories;

[0025] The SoftMax function uses a probability calculation formula, which is expressed as: after the logical scores of the three vacuum breakdown categories input are indexed, they are compared with the sum of the indexed logical scores of the three vacuum breakdown categories to obtain the probability of each vacuum breakdown machine category.

[0026] Its further technical solution is that during the training of the vacuum breakdown classification diagnosis model based on the convolutional neural network, the total loss is calculated according to the predicted category and the true category, and the weight parameters of the model are optimized and adjusted through back propagation; wherein the total loss function is expressed as:

[0027]

[0028] In the formula, i is the sample category, c is the total number of categories; y i is the true category of a sample, y i ' is the predicted category of a sample; p i is the weighted value of each category, expressed as:

[0029] p 1 =N / N FEBD ,p 2 =N / N PB ,p 3 =N / N PBD ,N=N FEBD +N PB +N PBD

[0030] Where N FEBD 、N PB 、N PBDThey are the numbers of pre-breakdown current waveform samples of field emission induced vacuum breakdown, pulse current induced vacuum breakdown, and particle induced vacuum breakdown, respectively.

[0031] A further technical solution is to obtain the pre-breakdown current waveforms of different vacuum breakdown categories, including:

[0032] Conduct multiple vacuum breakdown tests of different vacuum breakdown categories and record the voltage and current data of each vacuum breakdown test;

[0033] Analyze the voltage data to locate the vacuum breakdown moment, and then extract the data between the start of the test and the vacuum breakdown moment from the current data of the same group as the pre-breakdown current waveform of the vacuum breakdown category;

[0034] The breakdown moment is the moment when the voltage data suddenly changes for the first time and drops sharply to zero volts;

[0035] The pre-breakdown current waveform of vacuum breakdown induced by pulse current shows a pulse-like steep increase in the tail section, the pre-breakdown current waveform of vacuum breakdown induced by field emission shows an upward trend in the tail section, and the pre-breakdown current waveform of vacuum breakdown induced by particles shows a downward trend in the tail section.

[0036] In a second aspect, the present application also provides a vacuum breakdown category detection device based on a convolutional neural network, comprising:

[0037] Vacuum breakdown test platform, vacuum breakdown classification diagnosis model training script and PYNQ operating system on PC, hardware device equipped with vacuum breakdown classification diagnosis model based on convolutional neural network; Among them:

[0038] Vacuum breakdown test platform, used to carry out multiple vacuum breakdown tests of different vacuum breakdown categories and obtain pre-breakdown current waveforms of different vacuum breakdown categories. Vacuum breakdown categories include pulse current induced vacuum breakdown, field emission induced vacuum breakdown and particle induced vacuum breakdown;

[0039] The vacuum breakdown category diagnosis model based on convolutional neural network includes shared convolutional layer, multi-task processing layer, fully connected layer and SoftMax function;

[0040] The shared convolution layer is used to extract the basic features of the pre-breakdown current waveforms of the three vacuum breakdown categories and input them into the multi-task processing layer for processing;

[0041] The multi-task processing layer is used to identify the characteristics of different vacuum breakdown categories in the basic features, concatenate them to obtain a composite feature vector, and then calculate the probabilities of the three vacuum breakdown categories through the fully connected layer and the SoftMax function;

[0042] The vacuum breakdown category diagnosis model training script is used to train the model and load the weight parameters with the best training effect to the vacuum breakdown category diagnosis model;

[0043] The PYNQ operating system is used to input the pre-breakdown current waveform to be classified into the hardware device, receive the probabilities of the three vacuum breakdown categories returned by the hardware device, and select the vacuum breakdown category corresponding to the maximum probability as the final diagnosis result output.

[0044] A further technical solution is that, in the hardware device:

[0045] The shared convolutional layer, multi-task processing layer, and fully connected layer of the vacuum breakdown category diagnosis model are implemented using an FPGA development board;

[0046] The weight parameters with the best training effect are stored in the BRAM built into the FPGA development board;

[0047] The logic scores of the three vacuum breakdown categories output by the fully connected layer are saved in the DMA built into the FPGA development board;

[0048] The SoftMax function of the vacuum breakdown category diagnosis model is implemented using an ARM dual-core processor, and data interaction is achieved with DMA through shared memory DDR.

[0049] The beneficial technical effects of the present invention are:

[0050] This application proposes a vacuum breakdown category detection method and device based on a convolutional neural network. By collecting vacuum breakdown test data and processing to obtain pre-breakdown current waveform training samples, a vacuum breakdown category diagnosis model based on a convolutional neural network is used to highlight the rising and falling trends of the pre-breakdown current waveform samples, and the model is trained according to the processed samples to obtain the weight parameters with the best training effect and load them into the vacuum breakdown category diagnosis model. A vacuum breakdown category diagnosis model based on a convolutional neural network is built on a hardware device, and the pre-breakdown current waveform to be classified is input. After calculation and analysis, the vacuum breakdown category diagnosis result is output. Compared with manual and computer diagnosis, this method combines convolutional neural networks and hardware device examples. The single vacuum breakdown category diagnosis speed of this method is within 1ms, and the accuracy rate is above 90%, which greatly improves the diagnosis speed and efficiency of vacuum breakdown categories in engineering practice, and is of practical help to research related to vacuum breakdown. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a vacuum breakdown category detection method based on a convolutional neural network provided in this application.

[0052] Figure 2It is a schematic diagram of the pre-breakdown current waveform samples of different vacuum breakdown categories provided in the present application, wherein: (a) is the pre-breakdown current waveform diagram of vacuum breakdown induced by pulse current, (b) is the pre-breakdown current waveform diagram of vacuum breakdown induced by field emission, and (c) is the pre-breakdown current waveform diagram of vacuum breakdown induced by particles.

[0053] Figure 3 This is a network structure diagram of the vacuum breakdown category diagnosis model based on convolutional neural network provided in this application.

[0054] Figure 4 It is a vacuum breakdown test circuit diagram provided by this application.

[0055] Figure 5 This is a flow chart of applying impulse voltage and adjusting electrode spacing in a vacuum breakdown test provided by the present application.

[0056] Figure 6 It is a schematic diagram of a vacuum breakdown category detection device based on a convolutional neural network provided in this application. DETAILED DESCRIPTION

[0057] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0058] In one embodiment of the present application, a vacuum breakdown category detection method based on a convolutional neural network is provided. Please refer to Figure 1 As shown, the method comprises the following steps:

[0059] Step 1: Obtain pre-breakdown current waveforms of different vacuum breakdown categories as training samples.

[0060] First, carry out multiple vacuum breakdown tests of different vacuum breakdown categories, and record the voltage and current data of each vacuum breakdown test. The measurement time period of voltage and current data can be defined independently, but it must include the time period before vacuum breakdown. When obtaining the pre-breakdown current waveform sample, it is necessary to pay attention to the measurement method of voltage and current data. At the beginning of each vacuum breakdown test, the current exists and is weak, and vacuum breakdown occurs after a certain period of time. The current change characteristics of different categories are different, and specific analysis can be used as the basis for category diagnosis. At the beginning of each vacuum breakdown test, the voltage exists and is obvious. After a certain period of time, vacuum breakdown occurs and the voltage drops sharply. Therefore, for each set of voltage and current data, the voltage data is first analyzed to locate the vacuum breakdown moment, that is, the moment when the voltage data first mutates and drops sharply to zero volts, and then the data between the start of the test and the vacuum breakdown moment are extracted from the same group of current data as the pre-breakdown current waveform of this vacuum breakdown category. The data in this period can retain all the characteristics of the current before breakdown. In other implementations, current data of a fixed time period can also be extracted as a pre-breakdown current waveform sample, which can reduce the amount of data processing, but at the same time, some data features will be lost.

[0061] Vacuum breakdown categories include pulse current induced vacuum breakdown, field emission induced vacuum breakdown and particle induced vacuum breakdown. The pre-breakdown current waveform of pulse current induced vacuum breakdown has obvious pulse current in the tail section, that is, it shows the characteristics of a sharp increase in current size, such as Figure 2 (a) shows. There is no big difference between the pre-breakdown current of field emission induced vacuum breakdown and particle induced vacuum breakdown in the early stage, both of which show the characteristics of rising and then falling, but they show differences in the tail section. The pre-breakdown current waveform of field emission induced vacuum breakdown shows the characteristics of falling and then rising in the tail section, as shown in Figure 2 As shown in (b), the pre-breakdown current waveform of the particle-induced vacuum breakdown shows a characteristic of dropping to near zero in the tail section, as shown in Figure 2 (c) is shown. That is, the key difference between field emission induced vacuum breakdown and particle induced vacuum breakdown lies in the tail segment of the pre-breakdown current. The tail segment of the current in field emission induced vacuum breakdown shows an upward characteristic, while the tail segment of the current in particle induced vacuum breakdown shows a downward characteristic. Based on the above principle, the present application constructs a vacuum breakdown category diagnosis model based on a convolutional neural network to diagnose the vacuum breakdown category of the vacuum pre-breakdown current waveform.

[0062] Among the above three types of pre-breakdown current waveform samples, the number of samples of particle-induced vacuum breakdown accounts for the largest proportion, the number of samples of field emission-induced vacuum breakdown accounts for the second largest proportion, and the number of samples of pulse current-induced vacuum breakdown accounts for the smallest proportion. In this embodiment, the number of samples of particle-induced vacuum breakdown is about six times the number of samples of field emission-induced vacuum breakdown, and the number of samples of field emission-induced vacuum breakdown is about twice the number of samples of pulse current-induced vacuum breakdown. Based on the proportion of the number of pre-breakdown current waveform samples of the three vacuum breakdown categories, the category weighted value of the loss function of the vacuum breakdown category diagnosis model during the training process can be set.

[0063] Step 2: Construct a vacuum breakdown classification diagnosis model based on convolutional neural network. Figure 3 As shown in the figure, the model includes a shared convolution layer, a multi-task processing layer, a fully connected layer and a SoftMax function. Among them, the shared convolution layer is used to extract the basic features of the pre-breakdown current waveform of the three vacuum breakdown categories, and input them into the multi-task processing layer for processing. The multi-task processing layer includes a parallel feature branch network, a tail enhancement module, and a global pooling module, which are used to identify the characteristics of different vacuum breakdown categories in the basic features, splice to obtain a composite feature vector, and then calculate the probabilities of the three vacuum breakdown categories through the fully connected layer and the SoftMax function.

[0064] The shared convolution layer includes two layers of convolution and pooling layers, which are specifically implemented as follows: the first convolution pooling layer uses 32 3*3 convolution kernels, and then passes through a 2*2 pooling window; the second convolution pooling layer uses 64 3*3 convolution kernels, and then passes through a 2*2 pooling window. In this embodiment, the pre-breakdown current waveform of different vacuum breakdown categories obtained in step 1 is converted into a 28*28*1 grayscale image and then input to the shared convolution layer. After the first layer of convolution, the output shape becomes 28*28*32, after the first layer of pooling, the output shape becomes 14*14*32, after the second layer of convolution, the output shape becomes 14*14*64, after the second layer of pooling, the output shape becomes 7*7*64, and the shared convolution layer finally outputs the basic features of 7*7*64.

[0065] The feature branch network is used to capture and amplify the rising trend of the pre-breakdown current waveform of field emission-induced vacuum breakdown and the falling trend of the pre-breakdown current waveform of particle-induced vacuum breakdown, and extract the rising segment features and the falling segment features. The network includes an rising segment convolution module and a falling segment convolution module. Both modules use a 3*3 convolution kernel to perform convolution operations on each channel of the input basic features, and capture the rising / falling trend of the waveform by detecting the edge; after calculating the weighted sum, it is output as a 3*3*1 feature map through a 2*2 pooling window, and then flattened into a one-dimensional vector to obtain the rising segment features and the falling segment features. Among them, the convolution kernel of the rising segment convolution module represents that each row of the matrix is ​​[-1,0,1]; the convolution kernel of the falling segment convolution module represents that each row of the matrix is ​​[1,0,-1].

[0066] Since the pre-breakdown current will show the characteristics of rising in the early stage, in order to avoid the interference of the early rising trend on the feature extraction, the tail segment enhancement module is used to further focus on the change characteristics of the current tail segment, capture and amplify the change trend of the waveform tail segment, especially amplify and capture the rising trend of the current tail segment of field emission induced vacuum breakdown and the falling trend of the current tail segment of field particle induced vacuum breakdown, and extract the tail segment features. The tail segment enhancement module includes clipping, tail segment convolution layer and pooling flattening layer, that is, the tail area of ​​the basic features of the input is retained by clipping, such as the last five rows of the basic feature 7*7*64 5*7*64, and then the tail segment convolution layer uses 16 3*3 convolution kernels to extract the tail segment area features, and flattened into a one-dimensional vector through the pooling flattening layer.

[0067] The global pooling module integrates the global information of the input 7*7*64 basic features through global average pooling, obtains global features and converts them into one-dimensional feature vectors. This module further integrates the global features of the three pre-breakdown current waveforms based on the input basic features. At the same time, the global pooling module can avoid the simplification of features caused by the feature branch network and the tail enhancement module, and avoid the model paying too much attention to the features of field emission induced vacuum breakdown and particle induced vacuum breakdown, while ignoring the features of pulse current induced vacuum breakdown.

[0068] The composite feature is to merge the ascending segment feature, descending segment feature, tail segment feature and global feature into a composite feature by splicing. The splicing processing method is adopted to ensure that the four features obtained by the multi-task processing layer fully retain the data characteristics. The final composite feature is a one-dimensional composite feature vector of 642 parameters, which is composed of 9 parameters of ascending segment features, 9 parameters of descending segment features, 560 parameters of tail segment features and 64 parameters of global features.

[0069] The fully connected layer uses a logical score calculation formula, which is expressed as: multiplying the input composite feature vector by the weight matrix, and then summing the result with the bias matrix to obtain the logical scores of the three vacuum breakdown categories. In this embodiment, the fully connected layer maps the composite feature vector (642) to the number of categories (3), and the calculation formula is:

[0070]

[0071] Where x represents the input composite feature vector, x = [x 1 , x 2 , …, x 642 ]; W represents the weight matrix of the fully connected layer, with a shape of (642, 3), and each row element corresponds to the weight parameters of three vacuum breakdown categories; b is the bias matrix; z logits,1 , z logits,2 , z logits,3 These are the logits scores of the three vacuum breakdown categories.

[0072] The SoftMax function uses a probability calculation formula, which is expressed as: after the logical scores of the three vacuum breakdown categories are inputted and indexed, they are compared with the sum of the logical scores of the three vacuum breakdown categories to obtain the probability of each vacuum breakdown machine category. The calculation formula is:

[0073]

[0074] where z logits,i is the logits score of the three vacuum breakdown categories, P i are the probabilities of the three vacuum breakdown categories.

[0075] Step 3: Take the processed pre-breakdown current waveform as input and the vacuum breakdown category corresponding to the waveform as output, train the vacuum breakdown category diagnosis model based on convolutional neural network, and obtain the weight parameter with the best training effect to load it into the vacuum breakdown category diagnosis model. In the process of training the vacuum breakdown category diagnosis model based on convolutional neural network, calculate the total loss according to the predicted category and the true category, and optimize the weight parameter W of the model through back propagation. The total loss function is expressed as:

[0076]

[0077] Where i is the sample category (i=1, 2, 3) and c is the total number of categories (c=3). i is the true category of a sample, that is, y i is 0 or 1, the true category is 1, and the false category is 0; i ' is the predicted category of a sample, ranging between 0 and 1. log(y i ') for y i 'Take the logarithm, when y i The closer it is to 1, the closer its logarithm is to 0, which means the loss between the predicted value and the true value is smaller. i is the weighted value of each category, expressed as:

[0078] p 1 =N / N FEBD ,p 2 =N / N PB ,p 3 =N / N PBD

[0079] Where N FEBD 、N PB 、N PBD They are the number of pre-breakdown current waveform samples of field emission induced vacuum breakdown, pulse current induced vacuum breakdown, and particle induced vacuum breakdown, respectively, and N is the sum of the three. As the loss function continuously adjusts the model weight parameters in the training cycle, the weight parameters with the best training effect are obtained.

[0080] Step 4: Input the pre-breakdown current waveform to be classified into the vacuum breakdown category diagnosis model to obtain the probabilities of the three vacuum breakdown categories, and select the vacuum breakdown category corresponding to the maximum probability as the final diagnosis result output. Among them, the method of obtaining the pre-breakdown current waveform to be classified is the same as the method of obtaining the pre-breakdown current waveform sample in step 1, that is, the vacuum breakdown moment is obtained according to the voltage test data of the vacuum breakdown to be classified, and then the corresponding current test data is processed according to the breakdown moment to obtain the pre-breakdown current waveform diagram to be classified, which will not be repeated here.

[0081] In this embodiment, since the waveform sample used is a pre-breakdown current sample, and the pre-breakdown current contains typical features of three vacuum breakdown categories, a vacuum breakdown category diagnosis model for identifying typical features is set up. A shared convolution layer is set in the model to extract the basic features of the pre-breakdown current waveform samples of the three vacuum breakdown categories, and the branch features, tail features, and global features obtained by the multi-task processing layer are integrated into composite features. The composite features are processed in dimension, logical scores, and probabilities are calculated in combination with the fully connected layer and the SoftMax model, and the weight parameters are accurately adjusted, which can significantly improve the diagnosis speed and efficiency of the vacuum breakdown category.

[0082] Based on the same inventive concept, another embodiment of the present application also provides a vacuum breakdown category detection device based on a convolutional neural network, including: a vacuum breakdown test platform, a vacuum breakdown category diagnosis model training script and a PYNQ operating system located on the PC side, and a hardware device equipped with a vacuum breakdown category diagnosis model based on a convolutional neural network.

[0083] The vacuum breakdown test platform is used to carry out multiple vacuum breakdown tests of different vacuum breakdown categories and obtain the pre-breakdown current waveforms of different vacuum breakdown categories. Figure 4 As shown, the platform includes a pulse voltage generating device 1, a vacuum chamber / vacuum interrupter 2, a current measuring device 3, a voltage measuring device 4 and a computer device 5. Among them, the pulse voltage generating device 1, the vacuum chamber / vacuum interrupter 2 and the current measuring device 3 are connected in series to form a test circuit, the voltage measuring device 4 is connected in parallel with both ends of the vacuum chamber / vacuum interrupter 2, the computer device 5 is connected with the voltage measuring device 4 and the current measuring device 3 by electrical signals, and the current measuring device 3 is a measuring resistor or a Rogowski coil or a Hall coil.

[0084] During the test, the pulse voltage generating device 1 and the vacuum chamber 2 are mainly used for vacuum breakdown test, and the current measuring device 3 and the voltage measuring device 4 respectively measure the current data and the voltage data. Since the difference between the voltage and current data before and after the breakdown is an order of magnitude, in order to ensure the accuracy of the voltage and current data measurement, it is necessary to adopt a voltage measuring device 4 and a current measuring device 3 of a suitable range. In the test, the voltage before the breakdown is generally in the kilovolt level, and the voltage drops sharply after the breakdown, even to about zero volts, and the weak current before the breakdown is generally 0 to 20 mA. Therefore, in this embodiment, the voltage measuring device 4 adopts a kilovolt range, and the current measuring device 3 adopts a milliampere range. During the test, since the voltage size changes suddenly after the breakdown, the difference before and after is an order of magnitude. In order to avoid damage to the voltage measuring device 4, it is necessary to build an overload protection device in the voltage measuring device 4.

[0085] On this vacuum breakdown test platform, according to Figure 5The vacuum breakdown test is performed according to the flowchart shown. The specific operation steps are: first set the initial electrode spacing of the vacuum chamber 2 and the impulse voltage peak of the pulse voltage generating device 1. In this embodiment, the initial electrode spacing is set to d = 5mm, and the impulse voltage peak is U = 80Kv. During the test, it is necessary to determine whether the gap has not been broken down for five consecutive times. If so, the impulse voltage peak needs to be increased; if not, the impulse voltage with a peak value of U is repeatedly applied for 200 breakdown tests. In this embodiment, vacuum breakdown tests are carried out on three different spherical electrodes respectively, and the test current data of each vacuum breakdown test from the beginning to the breakdown moment are obtained. After the test current data is collected, the corresponding vacuum breakdown category can be determined according to the mathematical compensation algorithm and the Fowler-Northeim formula fitting, and combined with the impulse voltage of each test, and used as a label for each group of test data. Finally, the computer device 5 locates the vacuum breakdown moment by analyzing the kilovolt voltage data to obtain various types of pre-breakdown current waveform samples and transmit them to the PYNQ operating system. Specifically, the python script can be run to find the moment when the voltage drops sharply to zero for the first time in the kilovolt voltage data. This moment is the breakdown moment.

[0086] It should be noted that the compensation algorithm and the fitting method are both implemented using existing methods in the art, and the specific implementation process will not be repeated here.

[0087] like Figure 6 As shown, the basic structure of the PC and the hardware device is shown, wherein the vacuum breakdown category diagnosis model training script is used for model training, and the weight parameters with the best training effect are loaded into the vacuum breakdown category diagnosis model. The hardware device includes a memory for storing weight parameters, a device for deploying a shared convolution layer, a multi-task processing layer, and a fully connected layer of the vacuum breakdown category diagnosis model, and a device for deploying a fully connected layer and a SoftMax function. In this embodiment, the shared convolution layer, the multi-task processing layer, and the fully connected layer are implemented using an FPGA development board, and the FPGA can perform efficient parallel data calculation to improve the operation speed. The weight parameters with the best training effect are stored in the BRAM built into the FPGA development board. The logical scores of the three vacuum breakdown categories output by the fully connected layer are saved in the DMA built into the FPGA development board, and the DMA saves them to the shared memory DDR. The SoftMax function is implemented using an ARM dual-core processor, which can process floating-point data and perform more complex division. ARM and DMA realize data interaction through shared memory DDR, that is, ARM accesses DDR memory through the AXI bus to obtain the logical scores transmitted by DMA.

[0088] The FPGA model in this embodiment is XC7Z020, and in other implementations, embedded GPU, DSP and other devices may also be used. The ARM model is Cortex-A9, and in other embodiments, DSP or ASIC and the like may also be used to implement.

[0089] The PYNQ operating system is used to process the pre-breakdown current waveform to be classified into a 28*28*1 grayscale image and then input it to the hardware device through the Ethernet line to participate in the calculation of the logical scores and probabilities of the three vacuum breakdown categories. The PYNQ operating system receives the probabilities of the three vacuum breakdown categories returned by the hardware device through the Ethernet line, and selects the vacuum breakdown category corresponding to the maximum probability as the final diagnosis result output.

[0090] The vacuum breakdown category diagnosis method and device of the present application obtain high-precision pre-breakdown current waveform samples that can characterize the effective characteristics of the current before breakdown by collecting vacuum breakdown test voltage and current data. The vacuum breakdown category diagnosis model and hardware device based on the convolutional neural network realize vacuum breakdown machine category detection with high accuracy and reliability and a high degree of automation. The entire process can be achieved by the collaboration of software and hardware.

[0091] The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

1. A vacuum breakdown category detection method based on convolutional neural network, characterized in that: The method comprises: Obtaining pre-breakdown current waveforms of different vacuum breakdown categories as training samples, the vacuum breakdown categories including pulse current induced vacuum breakdown, field emission induced vacuum breakdown and particle induced vacuum breakdown; Construct a vacuum breakdown classification diagnosis model based on convolutional neural network, including shared convolutional layer, multi-task processing layer, fully connected layer and SoftMax function; The shared convolution layer is used to extract the basic features of the pre-breakdown current waveforms of the three vacuum breakdown categories and input them into the multi-task processing layer for processing; The multi-task processing layer is used to identify the characteristics of different vacuum breakdown categories in the basic features, splice to obtain a composite feature vector, and then calculate the probabilities of the three vacuum breakdown categories through the fully connected layer and the SoftMax function; Taking the processed pre-breakdown current waveform as input and the vacuum breakdown category corresponding to the waveform as output, the vacuum breakdown category diagnosis model based on the convolutional neural network is trained, and the weight parameter with the best training effect is obtained and loaded into the vacuum breakdown category diagnosis model; The pre-breakdown current waveform to be classified is input into the vacuum breakdown category diagnosis model to obtain the probabilities of three vacuum breakdown categories, and the vacuum breakdown category corresponding to the maximum probability is selected as the final diagnosis result output.

2. The vacuum breakdown category detection method based on convolutional neural network according to claim 1 is characterized in that: The shared convolution layer includes two convolutional layers and a pooling layer, where: The first convolution pooling layer uses 32 3*3 convolution kernels, and then passes through a 2*2 pooling window; the second convolution pooling layer uses 64 3*3 convolution kernels, and then passes through a 2*2 pooling window.

3. The vacuum breakdown category detection method based on convolutional neural network according to claim 1 is characterized in that: The multi-task processing layer includes a parallel feature branch network, a tail segment enhancement module, and a global pooling module; The characteristic branch network is used to capture and amplify the rising trend of the pre-breakdown current waveform of the field emission induced vacuum breakdown and the falling trend of the pre-breakdown current waveform of the particle induced vacuum breakdown, and extract the rising section feature and the falling section feature; The tail segment enhancement module is used to capture and amplify the changing trend of the tail segment of the waveform and extract the tail segment features; The global pooling module is used to extract the global features of the three pre-breakdown current waveforms.

4. The vacuum breakdown category detection method based on convolutional neural network according to claim 3 is characterized in that: The feature branch network includes: An ascending convolution module is used to perform a convolution operation on each channel of the input basic feature using a 3*3 convolution kernel, calculate the weighted sum and then flatten it into a one-dimensional vector through a 2*2 pooling window to obtain the ascending segment feature; A descending segment convolution module is used to perform a convolution operation on each channel of the input basic feature using a 3*3 convolution kernel, calculate the weighted sum and then flatten it into a one-dimensional vector through a 2*2 pooling window to obtain the descending segment feature; Among them, the convolution kernel of the ascending convolution module represents that each row of the matrix is ​​[-1, 0, 1]; the convolution kernel of the descending convolution module represents that each row of the matrix is ​​[1, 0, -1].

5. The vacuum breakdown category detection method based on convolutional neural network according to claim 3 is characterized in that: The tail enhancement module includes clipping, tail convolution layer and pooling flattening layer, wherein: By trimming the tail region that retains the basic features of the input, the tail convolution layer uses 16 3*3 convolution kernels to extract the tail region features, and flattens them into a one-dimensional vector through the pooling flattening layer.

6. The vacuum breakdown category detection method based on convolutional neural network according to claim 1 is characterized in that: The fully connected layer adopts a logic score calculation formula, which is expressed as: the input composite feature vector is multiplied by the weight matrix, and the result is summed with the bias matrix to obtain the logic scores of the three vacuum breakdown categories; each row element of the weight matrix corresponds to the weight parameters of the three vacuum breakdown categories; The SoftMax function adopts a probability calculation formula, which is expressed as: after the logical scores of the three vacuum breakdown categories input are indexed, they are compared with the sum of the indexed logical scores of the three vacuum breakdown categories respectively to obtain the probability of each vacuum breakdown machine category.

7. The vacuum breakdown category detection method based on convolutional neural network according to claim 1 is characterized in that: During the training of the vacuum breakdown classification diagnosis model based on the convolutional neural network, the total loss is calculated according to the predicted category and the true category, and the weight parameters of the model are optimized and adjusted through back propagation; wherein the total loss function is expressed as: In the formula, i is the sample category, c is the total number of categories; y i is the true category of a sample, y i ' is the predicted category of a sample; p i is the weighted value of each category, expressed as: p1=N / N FEBD ,p2=N / N P3 ,p3=N / N PBD ,N=N FEBD +N PB +N BBD Where N FEBD 、N PB 、N PBD They are the numbers of pre-breakdown current waveform samples of field emission induced vacuum breakdown, pulse current induced vacuum breakdown, and particle induced vacuum breakdown, respectively.

8. The vacuum breakdown category detection method based on convolutional neural network according to claim 1 is characterized in that: Get pre-breakdown current waveforms for different vacuum breakdown categories, including: Conduct multiple vacuum breakdown tests of different vacuum breakdown categories and record the voltage and current data of each vacuum breakdown test; Analyze the voltage data to locate the vacuum breakdown moment, and then extract the data between the test start time and the vacuum breakdown moment from the current data of the same group as the pre-breakdown current waveform of the vacuum breakdown category; The breakdown time is the time when the voltage data suddenly changes for the first time and drops sharply to zero volts; The pre-breakdown current waveform of the pulse current induced vacuum breakdown shows a pulse-like steep increase characteristic in the tail section, the pre-breakdown current waveform of the field emission induced vacuum breakdown shows a rising characteristic in the tail section, and the pre-breakdown current waveform of the particle induced vacuum breakdown shows a falling characteristic in the tail section.

9. A vacuum breakdown category detection device based on convolutional neural network, characterized in that: include: Vacuum breakdown test platform, vacuum breakdown classification diagnosis model training script and PYNQ operating system on PC, hardware device equipped with vacuum breakdown classification diagnosis model based on convolutional neural network; Among them: The vacuum breakdown test platform is used to carry out multiple vacuum breakdown tests of different vacuum breakdown categories and obtain pre-breakdown current waveforms of different vacuum breakdown categories, and the vacuum breakdown categories include pulse current induced vacuum breakdown, field emission induced vacuum breakdown and particle induced vacuum breakdown; The vacuum breakdown category diagnosis model based on convolutional neural network includes a shared convolution layer, a multi-task processing layer, a fully connected layer and a SoftMax function; The shared convolution layer is used to extract the basic features of the pre-breakdown current waveforms of the three vacuum breakdown categories and input them into the multi-task processing layer for processing; The multi-task processing layer is used to identify the characteristics of different vacuum breakdown categories in the basic features, splice to obtain a composite feature vector, and then calculate the probabilities of the three vacuum breakdown categories through the fully connected layer and the SoftMax function; The vacuum breakdown category diagnosis model training script is used to perform model training, and load the weight parameter with the best training effect to the vacuum breakdown category diagnosis model; The PYNQ operating system is used to input the pre-breakdown current waveform to be classified into the hardware device, receive the probabilities of three vacuum breakdown categories returned by the hardware device, and select the vacuum breakdown category corresponding to the maximum probability as the final diagnosis result output.

10. The vacuum breakdown category detection device based on convolutional neural network according to claim 9, characterized in that: In the hardware device: The shared convolution layer, multi-task processing layer, and fully connected layer of the vacuum breakdown category diagnosis model are implemented using an FPGA development board; The weight parameter with the best training effect is stored in the BRAM built into the FPGA development board; The logic scores of the three vacuum breakdown categories output by the fully connected layer are stored in the DMA built into the FPGA development board; The SoftMax function of the vacuum breakdown category diagnosis model is implemented by using an ARM dual-core processor, and data interaction is achieved with the DMA through a shared memory DDR.

Images