A method for real-time identification of nitrogen-charged deflector off-target state using machine learning

Abstract

The application discloses a method for identifying the nitrogen filling off-target state of a divertor in real time by using machine learning, and the method comprises the following steps: during the operation of a magnetic confinement fusion experimental device, the emission spectrum of specific nitrogen is observed by combining a visible camera with a nitrogen filter, and then the radiation change process of the nitrogen filling state of the divertor is monitored; different states observed are identified and classified; a database containing a large amount of state data is established; original data accumulated are used as input; the identification learning of different off-target states is realized by using a machine learning analysis program; the actual off-target state is compared with the result of program identification; the accuracy of program identification is verified while self-consistent calibration is performed; and the method is applied to the process of real-time discharge of the device. The method for identifying the nitrogen filling off-target state of a divertor in real time by using machine learning realizes the real-time classification of the off-target state of the divertor, and provides technical support for the heat flow regulation and the steady-state operation control of the divertor of the device.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of magnetic confinement nuclear fusion, and particularly relates to a method for real-time identification of nitrogen-filled divertor off-target state by using machine learning. BACKGROUND

[0002] The divertor is one of the key components of a magnetic confinement nuclear fusion device. One of its functions is to withstand and remove extremely high heat flow from the central plasma. Currently, long-pulse fusion experimental devices represented by EAST and future ITER all use active water-cooled tungsten-copper divertors to alleviate the thermal stress and heat flow on the divertor components. Despite this, methods such as impurity gas injection are still needed to alleviate the problem of excessive heat flow on the divertor. Nitrogen injection is one of the impurity gas injection methods, and nitrogen-filled divertor will form a stable high-radiation zone near the X point. Within the radiation zone, most of the plasma energy can be uniformly released in the form of photons and other forms in all directions, off-target while relieving the problem of excessive heat load on the divertor. Therefore, real-time and accurate identification of the off-target state of the divertor after nitrogen injection has important value for accurate control of the off-target state of the divertor and stable operation control of the device.

[0003] However, there is no effective method for identifying and analyzing the different off-target states of the nitrogen-filled divertor. SUMMARY

[0004] To solve the above technical problems, the present application provides a method for real-time identification of nitrogen-filled divertor off-target state by using machine learning. During the nitrogen injection process, due to the differences in device heating power, nitrogen injection time, and plasma current steady-state phase during discharge, the divertor will inevitably produce different degrees of off-target state. Each shot discharge condition will also have different off-target state change processes. Therefore, this feature can be used to collect and establish a database of four different off-target states by using a high-speed CCD visible camera combined with a nitrogen filter. By using a large amount of data to train and learn the machine learning program, the different off-target states can be accurately identified. The accuracy of the machine learning program can be verified by comparing and contrasting the off-target state determined manually and not recorded in the database. The trained machine learning program can be applied to real-time identification of different off-target states of the nitrogen-filled divertor during device discharge.

[0005] To achieve the above purpose and more accurately identify the nitrogen-filled divertor off-target state, the present application adopts the following technical solutions:

[0006] A method for real-time identification of the nitrogen-filled divertor off-target state by using a machine learning program, comprising the following steps:

[0007] Step 1: Before the experiment of the magnetic confinement nuclear fusion device, adjust the high-speed CCD visible camera and the endoscope system at the front end of the visible light path to realize the observation of the position of the divertor;

[0008] Step 2: Before the experiment of the magnetic confinement nuclear fusion device, select a nitrogen filter under a specific waveband (NII-637.96nm) installed in front of the high-speed CCD visible camera, adjust the camera exposure time and frame rate combined with the plasma discharge conditions to realize the observation of the nitrogen emission spectrum;

[0009] Step 3: During the long pulse stable operation of the magnetic confinement nuclear fusion device (usually >5s), use the combination of high-speed CCD visible camera and nitrogen filter to collect images of the divertor under the conditions of no nitrogen filling and no off-target and different degrees of off-target by using existing software;

[0010] Step 4: Combined with the different moments of nitrogen injection during long pulse discharge, manually classify and establish a large database of images collected under different off-target states; divided into four categories, one is the non-off-target state (no nitrogen injection), after nitrogen injection, one is the partial off-target state, another is the complete off-target state, and the last one is the edge multi-faceted asymmetric radiation state;

[0011] Step 5: Use the classified database of different off-target states as the input port of machine learning, use the neural network learning method to establish a data training model, and output the self-consistent comparison and analysis of the existing off-target states not recorded in the database, to realize the precise identification of different off-target states by using machine learning;

[0012] Step 6: Verify and analyze the results of different off-target states output by machine learning and the data in the database of step 5 input port, the process of possible multiple state overlaps of the four states in step 4, combined with other parameters such as nitrogen injection time and plasma current steady state time, to analyze and remove some abnormal data, and further improve the accuracy of the program recognition;

[0013] Step 7: Combine step 3, apply machine learning in the discharge process of the device, combine the experimental control conditions during operation to realize real-time identification of the off-target state, and provide data support for accurate control of the off-target state of the divertor.

[0014] Advantages:

[0015] This invention provides a precise and reliable method for real-time identification of the nitrogen-filled off-target state of a divertor. It accurately determines changes in the off-target state during nitrogen filling, providing more accurate data analysis for divertor component off-target detection and device physical operation. During the operation of a magnetic confinement fusion experimental device, this invention uses a high-speed CCD visible camera combined with a nitrogen filter to observe the emission spectrum of specific nitrogen, thereby monitoring the radiation changes in the nitrogen-filled state of the divertor. These different states are identified, classified, and a large database is established. The accumulated raw data is used as input to a machine learning analysis program to learn and identify different off-target states. By comparing the actual off-target states with the program's identification results, self-consistent calibration is performed while verifying the accuracy of the program's identification, and it is applied to the real-time discharge process of the device. This invention utilizes machine learning to identify the nitrogen-filled off-target state of a divertor in real time, achieving real-time classification of the divertor's off-target state and providing technical support for the divertor's heat flow regulation and steady-state operation control. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a method for real-time identification of nitrogen-filled off-target state of a divertor using machine learning, according to the present invention.

[0017] Figure 2a , Figure 2b , Figure 2c , Figure 2d A schematic diagram of the nitrogen-charged and target-de-target-free state of the divertor; wherein, Figure 2a The state is "0", meaning the target has not been missed. Figure 2b State "1" indicates a partial miss. Figure 2c State "2" indicates a completely missed target state. Figure 2d This is state "3", which is the multifaceted asymmetric radiation state at the edge;

[0018] Figure 3 The images showing the identified off-target states, which are not recorded in the database, are generated by a machine learning program.

[0019] Figure 4 The image is used to identify the changes in the divertor's off-target state in real time during the steady-state operation of a magnetic confinement fusion device via a machine learning program. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] Nitrogen flare-up of the divertor occurs when nitrogen gas is injected into the divertor's charging port for a certain period of time during the discharge of a magnetic confinement fusion device, reaching the flat peak of the plasma current. Figure 1 As shown, the present invention provides a method for real-time identification of the nitrogen-filling off-target state of a divertor using machine learning, which specifically includes the following steps:

[0022] Step 1: Before the experiment of the magnetic confinement nuclear fusion device, adjust the high-speed CCD visible camera and the endoscope system at the front end of the visible light path to realize the observation of the position of the divertor;

[0023] Step 2: When the high-speed CCD visible camera (such as Phantom VEO 1310L) combined with a nitrogen filter (such as NII-637.96nm narrowband filter, FWHM 2 nm, OD6, cutoff band 300-1000nm) uses existing software to record the different degrees of nitrogen injection off-target state of the divertor in the magnetic confinement fusion device, according to the different operating conditions of the device, the frame rate and exposure time of 250-500fps and 2000-4000us are adjusted, and then the images of the divertor in different off-target states are collected;

[0024] Step 3: Combined with the different moments of nitrogen injection during long pulse discharge, manually classify the large number of image databases collected in different off-target states (N Figures 2a-2d ) can be roughly divided into four categories, one is the non-off-target state (no nitrogen injection); after nitrogen injection, one is the partial off-target state, the other is the complete off-target state; the last one is the multi-faceted asymmetric radiation state of the edge;

[0025] Step 4: combined with Figure 1 , the classified database in different off-target states is used as the input port of machine learning, and the convolutional neural network learning method is used to establish the training model. The output port compares and analyzes the existing determined off-target states not recorded in the database, realizes the precise identification of different off-target states by machine learning, including:

[0026] The convolutional neural network learning method specifically adopts the residual neural network ResNet15, which is an improved convolutional neural network structure. Its main feature is to solve the gradient disappearance problem in deep network by introducing "jump connection" or "short circuit connection". In the input port of neural network learning, the database inputs a fixed size of JPG (pixel ) image, the picture convolution operation extracts picture features such as edges and textures, after convolution operation, ReLU activation function is introduced to introduce nonlinearity, and pooling operation is used to reduce data dimension, local response normalization is performed to improve training stability. Repeat convolution, activation, pooling and normalization operations to gradually extract higher-level image features of different scales. After convolution and pooling layer, the neural network will contain several fully connected layers, which will integrate the features extracted by convolution layer and pooling layer into a fixed length vector. At this time, the image features of all different off-target states have been learned by the neural network, and when other off-target state images are introduced, the output end will give the corresponding off-target state.

[0027] Step 5: Verify and analyze the results of the combined program output under different off-target states with the data in the step 4 input port database. The process of the four possible states in step 3 may overlap, and other parameters such as nitrogen injection time and plasma current steady-state phase time need to be analyzed to remove abnormal values with deviations in artificial classification state, and then to improve the accuracy of program recognition more accurately;

[0028] Step 6: When the divertor is off-target under nitrogen filling, try to put the images collected by one or two shots into the machine learning program as input, and compare the recognition accuracy of the output recognition results;

[0029] Step 7: Combine step 2, apply residual neural network ResNet15 in the process of divertor off-target discharge under nitrogen filling, and combine the experimental control conditions during operation to realize real-time identification of off-target state and provide data support for accurate control of divertor off-target state.

[0030] Embodiment:

[0031] After applying the training model program established by residual neural network ResNet15 to the magnetic confinement fusion device, observe the recognition of two shots under different nitrogen filling conditions and off-target states in the device, as shown in FIGS. 1-3, for the off-target state of the divertor during the whole stage of operation, the machine learning program (i.e. using convolutional neural network learning) is used for recognition, and the specific steps are as follows: Figure 3 、 Figure 4

[0032] S1: Before the experiment of the magnetic confinement fusion device, adjust the high-speed CCD visible camera and the endoscope system at the front end of the visible light path to realize the position observation of the divertor; select a nitrogen filter under a specific waveband and install it in front of the high-speed CCD visible camera, and adjust the camera focal length to realize the observation of the nitrogen emission spectrum;

[0033] S2: During the long pulse stable operation of the magnetic confinement fusion device, use the combination of high-speed CCD visible camera and nitrogen filter to collect images of the divertor under the conditions of no nitrogen filling and no off-target and different degrees of off-target under nitrogen filling;

[0034] S3: Combine the different times of nitrogen injection during long pulse discharge, and manually classify to establish a large database of images under different off-target states collected;

[0035] S4: Use the database of different off-target states classified as the input port of residual neural network ResNet15 (convolutional neural network learning) to establish a training model, and compare and analyze the existing off-target states not recorded in the database in the output port, to realize the accurate recognition of different off-target states by machine learning program;

[0036] ​S5: The results of different off-target states of the output are verified and analyzed with the data of the input port database in S5, and the process of the four states in S4 may appear in various state overlaps, which needs to be combined with other parameters, such as nitrogen injection time and plasma current steady-state stage time, to analyze and remove some abnormal value data with large state deviation, and then improve the accuracy of the program recognition;

[0037] S6: The image information collected in the entire discharge stage of the two guns of the nitrogen-filled off-target is taken as the input end of the machine learning program, and the output off-target states "0", "1", "2" and "3" are all shown in Figure 3 、 Figure 4 The recognition accuracy is high, and the state analysis of each frame is accurate.

[0038] S7: Combined with S2, the machine learning is applied in the discharge process of the device, the experimental control conditions during operation are considered, the off-target state is identified in real time, and data for accurately controlling the off-target state of the partial filter is provided.

[0039] The part of the application not described in detail belongs to the known technology of those skilled in the art. The above-described embodiments are only used to describe the preferred embodiments of the application, and the preferred embodiments do not describe all the details and limit the application to the specific embodiments described. Without departing from the design spirit of the application, various modifications and improvements of the technical solutions of the application made by those skilled in the art shall fall within the protection scope of the claims of the application.

Claims

1. A method for real-time identification of nitrogen off-target state in a divertor using machine learning, characterized in that, Includes the following steps: Step 1: Before the experiment of the magnetic confinement nuclear fusion device, adjust the high-speed CCD visible camera and the endoscope system at the front end of the visible light path to observe the position of the divertor. Step 2: Before the experiment of the magnetic confinement nuclear fusion device, a nitrogen filter under a specific wavelength band is selected and installed at the front end of the high-speed CCD visible camera. The camera focal length is adjusted to observe the nitrogen emission spectrum. The nitrogen filter under the specific wavelength band is: NII-637.96nm narrowband filter, FWHM 2 nm, OD6, cutoff band 300-1000nm. Step 3: During the long-pulse stable operation of the magnetic confinement nuclear fusion device, images of the divertor under different states of nitrogen filling and nitrogen removal are acquired using a combination of a high-speed CCD visible camera and a nitrogen filter. Step 4: Combining the different times of nitrogen injection during long-pulse discharge, a large database of images under different off-target states was manually classified and established. Step 5: Introduce "skip connections" or "short-circuit connections" using ResNet15 residual neural network to solve the gradient vanishing problem in deep networks; use the database of different off-target states as the input port for machine learning, train the model with a large amount of data, and perform self-consistent comparative analysis of the existing known off-target states not recorded in the database at the output port to achieve accurate identification of different off-target states using machine learning. Step 6: Verify and analyze the results of different off-target states output by machine learning with the data in the input port database of Step 5. Combine the nitrogen injection time and the steady-state stage time of plasma current to remove outliers that may be biased by manual classification, thereby improving the accuracy of program recognition. Step 7: Combining with Step 3, apply machine learning to the device discharge process, and combine it with the experimental control conditions during operation to achieve real-time identification of the off-target state, providing data for precise control of the off-target state of the divertor.

2. The method for real-time identification of nitrogen off-target state of divertor using machine learning according to claim 1, characterized in that, In step 3, the long pulse of the magnetic confinement nuclear fusion device refers to >5s, and stable operation means that the magnetic field configuration and plasma parameters do not change.

3. The method for real-time identification of nitrogen off-target state of divertor using machine learning according to claim 1, characterized in that, In step 4, a high-speed CCD visible camera is used to record the different off-target states of the divertor after nitrogen filling, and then the off-target state change process at each moment is obtained by frame-by-frame analysis.

4. The method for real-time identification of nitrogen off-target state of divertor using machine learning according to claim 1, characterized in that, In step 6, the large deviation refers to the process in step 4 where multiple states may overlap. The data with outlier deviations are analyzed by combining the nitrogen injection time and the steady-state stage time of the plasma current.

5. The method for real-time identification of nitrogen off-target state of a divertor using machine learning according to claim 1, characterized in that, In step 5, a large amount of data on different off-target states that have been identified are used as input for machine learning, and the four off-target states may not necessarily appear in every shot of the selected nitrogen-filled off-target.

Images