Method for counting arcing position and arcing time of low-clutter antenna by using machine learning

Abstract

The invention discloses a method for counting the arcing position and time of a low-clutter antenna by using machine learning, relates to the field of magnetic confinement fusion, and realizes continuous recording of the change process of the arcing state of the low-clutter antenna by observing bright spots generated by arcing through a visible camera when a magnetic confinement fusion device runs. A large number of different arc states are recognized and classified, a corresponding database is established, the database serves as input in a deep convolutional neural network analysis program, different arc starting states are counted, learned and classified, the results of manual recognition and program recognition of the arc starting states are compared, and the accuracy of program calculation is verified while calibration is conducted. And based on a program identification result, a machine learning-DBSCAN algorithm is further adopted to carry out normalized clustering processing on the classified images for different arcing states, and statistics is carried out on the duration time of different arcing positions on the low-clutter antenna. According to the invention, real-time identification of the arcing state of the low-clutter antenna can be realized, and the arcing position and duration of the low-clutter antenna can be counted at the same time.

Description

Technical Field

[0001] The present invention belongs to the field of magnetic confinement nuclear fusion, and in particular relates to a method for using machine learning to count the arcing position and time of a low-clutter antenna. Background Art

[0002] In a tokamak device, the plasma toroidal current plays a vital role in plasma confinement. However, the traditional method uses ohmic induction to drive the toroidal current. The limited variable magnetic flux provided by the transformer restricts the maintenance time of the toroidal current, so an auxiliary non-inductive method is needed to drive the plasma current. Low-noise current drive (LHCD) is a highly efficient non-inductive current drive method in the current tokamak device and one of the important means to achieve long-pulse steady-state operation of the tokamak. LHCD can be used as a method to maintain its plasma current. However, the low-noise antenna of the tokamak device is prone to arcing when high power is injected. These phenomena will produce local high-density plasma on the antenna surface, reduce the low-noise coupling efficiency, and affect the stable operation of the low-noise system. The arcing phenomenon will also produce a large number of high-atomic-number metal impurities, which will contaminate the plasma, seriously affect the stability of the plasma discharge, and may even cause the plasma to rupture, affecting the safe operation of the tokamak device. Therefore, understanding the mechanism of arcing on the surface of the low-noise antenna and realizing real-time monitoring and identification is crucial for the stable operation of the tokamak device. Due to the different moments of change in the heating power of the low-clutter antenna and the different plasma current stages during discharge, the low-clutter antenna will inevitably produce arcing states of varying degrees. The different discharge conditions of each shot will also result in different changes in the arcing state. At present, low-clutter antennas will have arcing states of varying degrees, which will seriously affect the steady-state operation of the tokamak device. Manual statistics of these arcing states, positions, and times will consume a lot of human resources and will be difficult to achieve real-time statistics during the operation of the device in the future. Therefore, there is an urgent need to obtain a convenient, fast, and accurate method for counting the arcing position and time of the low-clutter antenna. Summary of the invention

[0003] In order to solve the above technical problems and more accurately count the influence of different arcing states of low-clutter antennas, the present invention provides a method for using machine learning to count the arcing position and time of low-clutter antennas, and collects these different arcing states through high-speed CCD visible camera observation, and the output results of deep convolutional neural network program training and learning are combined with the existing different arcing states that are not recorded in the database and manually determined to achieve accurate identification of different arcing states. Based on the results output by the program, the machine learning-DBSCAN algorithm is further used to realize statistical analysis of arcing position and duration. The trained machine learning program is applied to real-time statistics of different arcing states, arcing positions, durations, etc. of low-clutter antennas when the device is discharged.

[0004] In order to achieve the above object, the present invention adopts the following technical scheme:

[0005] A method for using machine learning to count the arcing position and time of a low-clutter antenna comprises the following steps:

[0006] Step 1: Before the magnetic confinement nuclear fusion device is operated, the high-speed CCD visible camera and the endoscope system at the front end of the visible light path are adjusted, and the exposure time and acquisition frame rate of the high-speed CCD visible camera are set according to the experimental conditions to realize the observation of the arcing phenomenon of the low-clutter antenna;

[0007] Step 2: When the magnetic confinement nuclear fusion device is in operation, a high-speed CCD visible camera is used to record the entire process of the image of the change of the bright spot state generated when the low-clutter antenna arcs;

[0008] Step 3, combining the changes in the operating parameters of the magnetic confinement nuclear fusion device and using the change process images, an image database including three different arc starting states is established; the three different arc starting states are no arc starting state, single point arc starting state, and multi-point arc starting state;

[0009] Step 4: Use the image database established in step 3 as the input of the deep convolutional neural network model, train the deep convolutional neural network model, combine the image output by the deep convolutional neural network model, compare and analyze the artificially determined arc starting states not recorded in the image database, accurately identify different arc starting states, and obtain the output result;

[0010] Step 5: For different arcing states in the output results of step 4, a machine learning-DBSCAN algorithm is used to count all arcing positions on the low-clutter antenna and the arc duration at the arcing position during each shot discharge process;

[0011] Step 6: When the low-clutter antenna arcs, the images collected during one or two discharge processes are used as inputs of the machine learning-DBSCAN algorithm, and the output calculation results are compared with the manual statistical results to achieve accurate statistics of the arcing position and duration by the machine learning-DBSCAN algorithm;

[0012] Step 7. Apply the deep convolutional neural network model and machine learning-DBSCAN algorithm to the discharge process of the magnetic confinement fusion device. Combined with the experimental control conditions during the operation of the magnetic confinement fusion device, accurate identification of different arcing states is achieved. Then, the machine learning-DBSCAN algorithm is further used to realize the statistics of arcing position and duration.

[0013] When a magnetic confinement fusion device is running, the present invention uses a high-speed CCD visible camera to observe the bright spots generated by arcing, so as to realize continuous recording of the changing process of the arcing state of a low-clutter antenna, identify and classify a large number of different arcing states, and establish a corresponding image database. The image database is used as the input of a deep convolutional neural network, statistical learning is performed to classify different arcing states, and the results of the arcing states identified by manual recognition and the deep convolutional neural network are compared. The accuracy of the deep convolutional neural network calculation is verified while calibrating. Based on the results of the deep convolutional neural network recognition, a machine learning-DBSCAN algorithm is further used to normalize and cluster the classified images for different arcing states, and different arcing positions and durations on the low-clutter antenna are statistically counted, which is finally applied in the real-time discharge process of the magnetic confinement fusion device.

[0014] Beneficial effects:

[0015] The present invention can accurately and in real time determine the different arcing states in each shot discharge, the arcing position under different arcing states, and the arc duration at that position, providing more accurate data analysis for studying the coupling performance of low-clutter antennas and the physical operation of the device. The present invention is of great significance for further studying its impact on plasma operation and developing methods for controlling arc triggering, ensuring the steady-state operation of the low-clutter current drive system, and promoting the development of long-pulse steady-state operation of Tokamaks. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A flow chart of a method for using machine learning to count the arcing position and time of a low-clutter antenna in real time according to the present invention;

[0017] Figure 2 This is a schematic diagram of the low-clutter antenna in the non-arcing state;

[0018] Figure 3 This is a schematic diagram of the single-point arc starting state of the low-clutter antenna;

[0019] Figure 4 This is a schematic diagram of the multi-point arcing state of the low-clutter antenna;

[0020] Figure 5 This is a statistical diagram of the number of images of different arcing states under a complete discharge of a gun that has not been recorded in the database and has been recognized by the output of the deep convolutional neural network program;

[0021] Figure 6 The image of the arcing position and time of the low-clutter antenna is statistically analyzed in real time by a deep convolutional neural network program and machine learning-DBSCAN during the steady-state operation of the magnetic confinement fusion device. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0023] In order to better illustrate the method of how to use a deep convolutional neural network to identify the arcing state of a low-clutter antenna in real time, the present invention is described in detail below with reference to the accompanying drawings and implementation cases.

[0024] like Figure 1 As shown, the arcing of the low-clutter antenna is a local discharge phenomenon with high current density and short duration that occurs on the surface of the plasma facing the component material when the magnetic confinement fusion device discharges. A method for using machine learning to count the arcing position and time of the low-clutter antenna in an embodiment of the present invention includes the following steps:

[0025] Step 1: Before the magnetic confinement nuclear fusion device experiment, adjust the high-speed CCD visible camera and the endoscope system at the front end of the visible light path to achieve position observation of the low-clutter antenna;

[0026] Step 2: The high-speed CCD visible camera records the arcing phenomenon of the low-clutter antenna in the magnetic confinement fusion device and establishes a database. According to the different operating conditions of the magnetic confinement fusion device, a frame rate of 1000fps or above and an exposure time adjustment of 1-10us can be achieved. After collecting images in different states, the different arcing states of the low-clutter antenna can be statistically analyzed (i.e. Figure 1 The camera in the image captures images under different arc starting states);

[0027] Step 3: Combine the low noise phase and different moments of power injection during long pulse discharge to artificially establish different arc starting states ( Figure 2-Figure 4 ) database; it can be roughly divided into three categories, one is the state without arcing (no power injection - state "Ⅰ"), after power injection, another is the state of single-point arcing - state "Ⅱ", and the last one is the state of multi-point arcing - state "Ⅲ";

[0028] Step 4: Figure 1 As shown in the figure, the database of different arc starting states is used as input, and the training model is established by using the deep convolutional neural network learning method. The output results are compared and analyzed with images of other arc starting states that are not recorded in the database, so that the deep convolutional neural network can accurately count the different arc starting states;

[0029] The input of the deep convolutional neural network learning is a database of manually classified different states. Its fixed-size JPG (pixel 768 736) image, convolution kernels are used to convolve the input image to extract local features in the image. These convolution kernels slide over the entire image to generate feature maps. The convolutional feature maps are nonlinearly transformed through the ReLU activation function, and then pooling operations are performed to reduce the size of the feature maps. Local response normalization is performed to improve training stability. Convolution, activation, pooling and normalization operations are repeated to gradually extract higher-level image features of different scales. After the convolution and pooling layers, the neural network will contain several fully connected layers. The fully connected layers integrate the features extracted from the convolution and pooling layers and classify them through one or more fully connected layers. In the last layer of the model, a Softmax layer is usually used to select the category with the highest probability and directly obtain the final classification result in combination with the argmax operation. At this point, the image features of all different arcing states have been learned by the neural network. The number of images of each category is counted using a dictionary and each category is output. When the different arcing state images of the complete stage of the discharge of another gun are introduced, the output will give the calculated different arcing states. Combine the results of different arcing states output by the program with the data in the input port database for verification and analysis. The manually classified state in step 3 needs to be combined with other parameters, such as low-clutter antenna power injection and phase change time, antenna port probe diagnosis to analyze and remove some data with large deviations (such as multiple bright spots generated by multi-point arcing may overlap and be judged as single-point arcing), thereby more accurately improving the accuracy of program recognition;

[0030] Step 5: Combine the results of the deep convolutional neural network output to identify and classify and count the number of various arcing states. For the classified single-point arcing state and multi-point arcing state images, the machine learning-DBSCAN algorithm without a given classification number is used to normalize the overall brightness of the image. The image sets a brightness threshold (preferably 150) to normalize its output. DBSCAN is a density-based clustering algorithm. Since the location of the bright spot is actually more than the size of one pixel, the distance between the two pixels needs to be given. That is, if the distance between the two points is less than this value, they will be classified into one category. After using DBSCAN to obtain the classification of the bright spot, the mean is obtained within each cluster (that is, Figure 1 The number of frames and position coordinates of the arc are detected, and the brighter area, i.e., the surrounding pixels of the bright spot, are merged and marked in the image as the location where the arc is generated (i.e. Figure 1 Finally, the results are saved, including the time and number of arcs detected, the frame rate of the cine file, and the duration of the bright spot position in each shot discharge (i.e. Figure 1 ) of the recorded arc duration in .

[0031] Step 6: When the low-clutter antenna arcs, try to use the images collected during one or two discharge processes as the input of the machine learning program, and compare the output calculation results with the manual statistical results to achieve accurate statistics of the arcing position and duration by the machine learning program;

[0032] Step 7: Apply the deep convolutional neural network model and machine learning-DBSCAN algorithm to the discharge process of the magnetic confinement fusion device. Combined with the experimental control conditions during the operation of the magnetic confinement fusion device, the different arcing states can be accurately identified. Then, the machine learning-DBSCAN algorithm is further used to realize the statistics of arcing position and duration.

[0033] Example:

[0034] After applying the machine program to the magnetic confinement fusion device, the low-noise antenna in the device was observed to accurately calculate the arc starting position and duration under different discharge conditions, such as Figure 5 , Figure 6 As shown in the figure, the arc starting position and duration of the low-clutter antenna in all stages of operation are identified using the machine learning-DBSCAN algorithm, including:

[0035] S1: Before the magnetic confinement nuclear fusion device experiment, adjust the high-speed CCD visible camera and the endoscope system at the front end of the visible light path to achieve position observation of the low-clutter antenna;

[0036] S2: When the magnetic confinement nuclear fusion device is in stable operation, a high-speed CCD visible camera is used to collect images of the low-clutter antenna in the non-arcing state and in the multi-point and single-point arcing states at different positions;

[0037] S3: Combined with the different moments of power injection and phase change during long pulse discharge, a large number of images of different arcing states and their corresponding databases are manually classified; the manually classified states need to be combined with other parameters, such as low-clutter antenna power injection and phase change time, antenna port probe diagnostic information, etc. to analyze and remove some data with large deviations (such as multiple bright spots generated by multi-point arcing may overlap and be judged as single-point arcing). The image information collected during the entire stage of a certain shot of arcing is used as the input of the deep convolutional neural network program, and the output arcing states "Ⅰ", "Ⅱ", and "Ⅲ" are obtained. Figure 5 It can be displayed on the Internet (each moment only corresponds to one state). After manual verification, it is found that its statistical accuracy is high and the situation analysis of different states is accurate.

[0038] S4: using the classified images of different arcing states and their database as the input of the deep convolutional neural network, training the deep convolutional neural network model, and comparing and analyzing the existing determined arcing states that are not recorded in the database at the output, and combining the results of different arcing states output by the deep convolutional neural network with the data of the input port for verification and analysis;

[0039] S5: Combined with the results of different arcing states output by the deep convolutional neural network learning program, the machine learning-DBSCAN program is further used to normalize the image and then cluster the pixels with bright spots to obtain the average value. The number of frames and position coordinates of the arc are detected, and the brighter area, that is, the surrounding pixels where the bright spot is located, are merged and marked in the image, which is regarded as the location where the arc is generated. The number of frames that the bright spot at this position lasts during the discharge process is recorded, and the duration of the bright spot is obtained by combining the parameters set by the camera; Figure 6 The image shows the location and duration of arcing after merging 10 pixels of the original image captured by a high-speed CCD visible camera of a certain gun;

[0040] S6: When the low-clutter antenna arcs, try to use the image collected during a discharge process as the input of the machine learning-DBSCAN algorithm, and compare and verify the output calculation results with the manually counted arc position and duration, so as to achieve accurate statistics of the arc position and duration by the machine learning program;

[0041] S7: Combined with S2, the machine learning-DBSCAN algorithm is applied to the discharge process of the magnetic confinement fusion device, taking into account the experimental control conditions during operation, to achieve real-time identification of the arcing state and accurate statistics of the arcing position and duration, to provide data support for precise control of the arcing state of the low-noise antenna and promote the steady-state operation of the low-noise current drive system.

[0042] The present invention does not elaborate on some of the common technologies of those skilled in the art. The above-described embodiments are only descriptions of preferred implementations of the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementations described. Without departing from the design spirit of the present invention, various modifications and improvements made by ordinary technicians in this field to the technical solution of the present invention should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A method for using machine learning to calculate the arcing position and time of a low-clutter antenna, characterized in that: The following steps are involved: Step 1: Before the magnetic confinement nuclear fusion device is operated, the high-speed CCD visible camera and the endoscope system at the front end of the visible light path are adjusted, and the exposure time and acquisition frame rate of the high-speed CCD visible camera are set according to the experimental conditions to realize the observation of the arcing phenomenon of the low-clutter antenna; Step 2: When the magnetic confinement nuclear fusion device is in operation, a high-speed CCD visible camera is used to record the entire process of the image of the change of the bright spot state generated when the low-clutter antenna arcs; Step 3, combining the changes in the operating parameters of the magnetic confinement nuclear fusion device and using the change process images, an image database including three different arc starting states is established; the three different arc starting states are no arc starting state, single point arc starting state, and multi-point arc starting state; Step 4: Use the image database established in step 3 as the input of the deep convolutional neural network model, train the deep convolutional neural network model, combine the image output by the deep convolutional neural network model, compare and analyze the artificially determined arc starting states not recorded in the image database, accurately identify different arc starting states, and obtain the output result; Step 5: For different arcing states in the output results of step 4, a machine learning-DBSCAN algorithm is used to count all arcing positions on the low-clutter antenna and the arc duration at the arcing position during each shot discharge process; Step 6: When the low-clutter antenna arcs, the images collected during one or two discharge processes are used as inputs of the machine learning-DBSCAN algorithm, and the output calculation results are compared with the manual statistical results to achieve accurate statistics of the arcing position and duration by the machine learning-DBSCAN algorithm; Step 7. Apply the deep convolutional neural network model and machine learning-DBSCAN algorithm to the discharge process of the magnetic confinement fusion device. Combined with the experimental control conditions during the operation of the magnetic confinement fusion device, accurate identification of different arcing states is achieved. Then, the machine learning-DBSCAN algorithm is further used to realize the statistics of arcing position and duration.

2. The method for calculating the arcing position and time of a low-clutter antenna by using machine learning according to claim 1, characterized in that: In step 1, the experimental conditions indicate that the duration of discharge of the magnetic confinement fusion device is different; when the magnetic confinement fusion device is in operation, the phenomenon of plasma discharge rupture is excluded from the record.

3. The method of using machine learning to calculate the arcing position and time of a low-clutter antenna according to claim 1, characterized in that: In step 2, a high-speed CCD visible camera is used to record the different arcing states of the low-clutter antenna during each shot discharge process, and each frame of image is used for subsequent statistics of the arcing position and corresponding duration of the shot.

4. The method of using machine learning to calculate the arcing position and time of a low-clutter antenna according to claim 1, characterized in that: In step 4, the results of deep convolutional neural network recognition and manual classification are compared in combination with other parameters to analyze and remove deviation data. The other parameters include the time of low-clutter antenna injection power and phase change, and the time of the steady-state stage of plasma current.

5. The method of using machine learning to calculate the arcing position and time of a low-clutter antenna according to claim 1, characterized in that: The deep convolutional neural network in step 4 uses a large amount of identified data of different arc starting states as input; the different arc starting states include one, two or three of the three arc starting states.

6. The method of using machine learning to calculate the arcing position and time of a low-clutter antenna according to claim 1, characterized in that: In step 5, the arc duration at a certain arcing position statistically calculated by the machine learning-DBSCAN algorithm is calculated by the collection of frames in which the arc appears discontinuously during the entire discharge process.

Images