Method and apparatus for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks

Abstract

This invention belongs to the field of plasma diagnostic technology and discloses a method and apparatus for identifying and locating magnetohydrodynamic (MHD) instabilities based on convolutional neural networks. The technical solution includes the following steps: S1, obtaining the propagation delay time at different radial positions based on microwave reflection diagnostic data; S2, obtaining the spatial gradient of the propagation delay time and inputting it into the constructed convolutional neural network; S3, analyzing the results of S2, and if periodic disturbances exist, performing correlation analysis between the delay time and the magnetic signal; if the maximum correlation value is greater than twice the average value, proceeding to S4; S4, converting the results of the convolutional neural network into the radial position of the periodic disturbance. This invention intelligently judges microwave reflection time delay data using a computer, eliminating the need for manually set thresholds, simplifying the workflow, and improving the efficiency and accuracy of MHD identification. This method does not rely on manually set thresholds, making it more intelligent, efficient, and accurate.

Description

Technical Field

[0001] This invention belongs to the field of plasma diagnostic technology, specifically relating to a method and apparatus for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks. Background Technology

[0002] Magnetohydrodynamic instabilities (MHDs) are common in tokamaks, such as neoclassical tearing modes (NTMs) and twisting modes. NTMs can significantly enhance radial transport of particles, limiting the fusion power of future fusion devices and even leading to discharge failure. Depositing additional heating power at the O-point of the magnetic island can mitigate the growth of NTMs. Therefore, achieving high spatial and temporal resolution for MHD identification and location detection is crucial.

[0003] Currently, the identification and localization of MHDs mainly rely on two diagnostic methods: electron cyclotron radiation diagnosis and microwave reflection diagnosis. The occurrence of MHDs may cause periodic changes in plasma density and temperature. Electron cyclotron radiation diagnosis can identify these temperature changes, thus enabling the identification and localization of MHDs. However, this method cannot be used under conditions with low-hybrid current driving. To address this limitation, a method using microwave reflection to measure MHDs was later developed. Microwave reflection measurement is more widely applicable than electron cyclotron radiation diagnosis, but this method relies on a manually set threshold as a judgment condition to determine whether MHDs have occurred. The optimal threshold may vary significantly under different plasma parameters. Summary of the Invention

[0004] The purpose of this invention is to provide a method and apparatus for identifying and locating magnetohydrodynamic instabilities (MHDs) based on convolutional neural networks. This method can automatically identify and locate MHDs in tokamaks in real time. Compared with traditional methods, this invention uses computer-intelligent judgment of microwave reflection time delay data, eliminating the need for manually setting thresholds, simplifying the workflow, and improving the efficiency and accuracy of MHD identification.

[0005] Firstly, to achieve the above objectives, this invention proposes a method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks, comprising the following method steps:

[0006] S1. Based on the microwave reflection diagnostic data, the transmission delay time at different radial positions is obtained;

[0007] S2. Obtain the spatial gradient of the transmission delay time and input it into the constructed convolutional neural network;

[0008] S3. Analyze the results of S2. If there is a periodic disturbance, perform a correlation analysis between the delay time and the magnetic signal. If the maximum value of the correlation is greater than twice the average value, then execute S4.

[0009] S4. Convert the results of the convolutional neural network into the radial positions of periodic perturbations.

[0010] Optionally, the transmission delay time at different radial positions is obtained in step S1, and the specific steps are as follows:

[0011] A series of continuously varying frequency probe microwaves are emitted into the plasma using a swept-frequency microwave source. The emitted electromagnetic waves are reflected by a cutoff layer and received by a receiving antenna. After processing the reflected signals, the propagation delay time τ of microwaves at different frequencies in the plasma can be obtained. Based on... The propagation delay at different radial positions x can be obtained, where u is the group velocity of the probe microwave, which can be obtained directly from the plasma background parameters. Finally, the propagation delay time at different radial positions is obtained through an inversion algorithm.

[0012] Optionally, the spatial gradient of the transmission delay time in S2 is obtained as follows:

[0013] The gradient distribution data is spliced ​​together according to a certain time length to obtain a series of slices, that is, each slice is the evolution data of the transmission delay time gradient of a certain time length.

[0014] Optionally, the convolutional neural network constructed from the input in S2 is specifically as follows:

[0015] When a tearing mode or a twisting mode (MHD) exists, there will be periodic perturbations in the transmission delay time gradient. The location of the perturbation is the radial position of the MHD. Through pre-training with a large amount of data, the convolutional neural network identifies these periodic perturbations and constructs the convolutional neural network structure.

[0016] Optionally, the convolutional neural network structure includes a first layer, a second layer, a third layer, a fourth layer, a fifth layer, and a sixth layer;

[0017] First layer: Input data is padded to make the output size equal to the input size, and then processed by the ReLU activation function;

[0018] The second layer: obtain the corresponding data through the max pooling layer;

[0019] The third layer: The corresponding data in the second layer is input data, and the output size is made equal to the input size by padding. Then, ReLU activation function is applied to obtain the processed data.

[0020] Fourth layer: Obtain the corresponding data through the max pooling layer;

[0021] Fifth layer: Input data is the corresponding data from the fourth layer, which is fully connected, then processed by the ReLU activation function, then dropped out, and finally convolutional data is obtained;

[0022] The sixth layer: Input convolutional data, fully connected, and processed using the sigmoid activation function to obtain feature data.

[0023] Optionally, the analysis of the results in S3 is as follows:

[0024] The convolutional neural network (CNN) outputs a probability value between 0 and 1. The first value represents whether there is a periodic perturbation in the time delay gradient, and the subsequent values ​​represent the probability distribution at different radial positions. If the maximum probability value output by the neural network is at the first value, it indicates that there is no periodic perturbation in the time delay gradient. If the maximum probability value output by the neural network appears in the subsequent values, it indicates that there is a periodic perturbation in the time delay gradient, and the position of the maximum probability value is the radial position corresponding to the perturbation. If the results of the CNN indicate the presence of a periodic perturbation, a correlation analysis is performed between the time delay data measured by microwave reflection and the magnetic signal. If there is a significant correlation, it indicates that this periodic perturbation in the time delay gradient is a micro-HD (microwave-high frequency, microwave-high frequency, microwave-high frequency).

[0025] Optionally, S4 specifically involves: the output of the convolutional neural network contains the probability distribution of the radial position of the MHD, and the radial position of the MHD can be determined based on the position of the maximum probability.

[0026] Secondly, a device for identifying and locating magnetohydrodynamic instability based on convolutional neural networks is applied to any of the above methods. The device includes a diagnostic system composed of microwave diagnostics and magnetic probe diagnostics, a real-time data acquisition system, and a real-time identification and positioning system. The signals detected by the diagnostic system are transmitted to the real-time data acquisition system composed of data acquisition modules via a coaxial transmission line. The real-time data acquisition system is responsible for acquiring real-time microwave signals and magnetic signals. The real-time identification system is connected to the real-time data acquisition system.

[0027] Optionally, the real-time recognition system includes a transmission delay extraction module, a neural network processing module, an MHD recognition module, and an MHD localization module.

[0028] The transmission delay extraction module is used to obtain the transmission delay time at different radial positions;

[0029] The neural network processing module is used to obtain probability data in the range of 0 to 1;

[0030] The MHD identification module is used to identify whether periodic disturbances exist;

[0031] The MHD positioning module is used to transform the probability distribution of the convolutional neural network to obtain the radial position of the periodic perturbation.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] This invention can automatically identify and locate microwave reflection time delay data in a tokamak in real time. Compared with traditional methods, this invention uses computer intelligence to judge microwave reflection time delay data, eliminating the need for manually setting thresholds, simplifying the workflow, and improving the efficiency and accuracy of MHD identification. This method does not rely on manually set thresholds, making it more intelligent, efficient, and accurate. Attached Figure Description

[0034] Figure 1 This is a flowchart illustrating a method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks.

[0035] Figure 2 This is a schematic diagram of a device for real-time identification and localization of magnetohydrodynamic instabilities based on neural networks.

[0036] Figure 3 A schematic diagram of the ReLU activation function and the sigmoid activation function.

[0037] Figure 4 This is a schematic diagram showing the perturbation of the transmission delay gradient caused by a periodic MHD at the position R~2m during a discharge of a tokamak device. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Please see Figure 1 This invention proposes a method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks, comprising the following steps:

[0040] S1. Obtaining the transmission delay (TOF) at different radial positions: Microwave reflection diagnostics emits a series of microwaves with varying frequencies into the plasma. After being reflected back by the plasma and processed, the transmission delay at different frequencies can be obtained. Based on the transmission delay, the corresponding radial position can be obtained, and finally, the transmission delay at different radial positions can be obtained.

[0041] A series of continuously varying frequency probe microwaves are emitted into the plasma using a swept-frequency microwave source. The emitted electromagnetic waves are reflected by a cutoff layer and received by a receiving antenna. After processing the reflected signals, the propagation delay time τ of microwaves at different frequencies in the plasma can be obtained. Based on... The propagation delay at different radial positions x can be obtained, where u is the group velocity of the probe microwave, which can be obtained directly from the plasma background parameters. Finally, the propagation delay time at different radial positions is obtained through an inversion algorithm.

[0042] S2. Take the spatial derivative of the delay time obtained in S1 to obtain the spatial gradient of the delay time. Pass the gradient data to the convolutional neural network processing module to obtain 51 values ​​between 0 and 1.

[0043] The gradient distribution data is spliced ​​together according to a certain time length to obtain a series of slices, that is, each slice is the evolution data of the transmission delay time gradient of a certain time length.

[0044] When a tearing mode or a twisting mode (MHD) exists, there will be periodic perturbations in the transmission delay time gradient. The location of the perturbation is the radial position of the MHD. Through pre-training with a large amount of data, the convolutional neural network identifies these periodic perturbations and constructs the convolutional neural network structure.

[0045] The convolutional neural network structure includes a first layer, a second layer, a third layer, a fourth layer, a fifth layer, and a sixth layer;

[0046] The first layer: The input data is 101×10×1, containing 64 convolutional kernels with a kernel size of 3×3 and a stride of 1. Padding is used to make the output size equal to the input size, and the output features are 101×10×64. Then, the ReLU activation function is used to obtain 101×10×64 data.

[0047] The second layer: a 2×2 max pooling layer with a step size of 1, yielding 50×5×64 data points;

[0048] The third layer: The input data is 50×5×64, containing 128 convolutional kernels. The kernel size is 3×3 and the stride is 1. The output size is made equal to the input size by padding. The output features are 50×5×128. Then, ReLU activation function is applied to obtain 50×5×128 data.

[0049] Fourth layer: 2×2 max pooling layer with a step size of 1, yielding 25×2×128 data;

[0050] Fifth layer: Input data is 25×2×128, using a fully connected layer to obtain 256 features, then ReLU activation function is applied, followed by dropout processing, finally obtaining 256 data points;

[0051] The sixth layer: Input data 256, using a fully connected layer, processed with the sigmoid activation function to obtain 51 feature data.

[0052] The ReLU activation function mentioned above is as follows: Figure 3 As shown, the negative half-axis being 0 creates sparsity in the network and reduces the interdependence of parameters, thus mitigating the overfitting problem; the positive half-axis being a linear region avoids the gradient vanishing problem and can significantly accelerate the convergence speed.

[0053] The above sigmoid activation function is as follows: Figure 3 As shown in the sigmoid function, the output of this function ranges from 0 to 1. This function can be used to convert the output of a neural network into a probability distribution. The 51 feature data mentioned above represent 51 probability distributions.

[0054] S3. Analyze the 51 values ​​output by the convolutional neural network in S2 to find the position of the maximum value. If the maximum value is in the first position, it indicates that no MHD has been detected. If the maximum value is not in the first position, perform correlation analysis between the time delay data and the magnetic signal. If the maximum correlation value is greater than twice the overall average, it indicates that MHD exists. Execute S4.

[0055] The convolutional neural network (CNN) outputs 51 probability values ​​between 0 and 1. The first value represents whether there is a periodic perturbation in the time delay gradient, and the following 50 values ​​represent the probability distribution at different radial positions. If the maximum probability value output by the neural network is the first value, it indicates that there is no periodic perturbation in the time delay gradient. If the maximum probability value appears in one of the following 50 values, it indicates that there is a periodic perturbation in the time delay gradient, and the position of the maximum probability value is the radial position corresponding to the perturbation. If the results of the CNN indicate the presence of a periodic perturbation, correlation analysis is performed between the microwave reflection measurement time delay data and the magnetic signal. If a significant correlation exists, it indicates that this periodic perturbation in the time delay gradient is a micro-HD (microwave-high frequency, microwave-high frequency, microwave-high frequency).

[0056] S4. Convert the results of the convolutional neural network into the radial positions of periodic perturbations.

[0057] Please see Figure 4 In a discharge of a tokamak device, a periodic MHD exists at a position R ~ 2m. Before each sawtooth collapse, this MHD causes a perturbation in the propagation delay gradient, such as... Figure 4 As shown. Following the above steps, the location of the MHD can be accurately obtained under high temporal resolution conditions, such as... Figure 4 (b) The blue data points also show that the neural network accurately eliminated the interference of boundary transport barriers; no threshold needs to be set manually in this process, as the system automatically identifies them.

[0058] In summary, this invention can automatically identify and locate MHDs in a tokamak in real time. Compared with traditional methods, this invention uses computer-intelligent judgment of microwave reflection time delay data, eliminating the need for manually setting thresholds, simplifying the workflow, and improving the efficiency and accuracy of MHD identification.

[0059] The output of a convolutional neural network contains the probability distribution of the radial position of the MHD, and the radial position of the MHD can be determined based on the position of the maximum probability.

[0060] For further information, please refer to [link / reference]. Figure 2 A device for identifying and locating magnetohydrodynamic instability based on convolutional neural networks, applied to any of the above methods, includes a diagnostic system composed of microwave diagnostics and magnetic probe diagnostics, a real-time data acquisition system, and a real-time identification and positioning system. The signals detected by the diagnostic system are transmitted to the real-time data acquisition system composed of data acquisition modules via a coaxial transmission line. The real-time data acquisition system is responsible for acquiring real-time microwave and magnetic signals. The real-time identification system is connected to the real-time data acquisition system.

[0061] Furthermore, the real-time recognition system includes a transmission delay extraction module, a neural network processing module, an MHD recognition module, and an MHD localization module:

[0062] The transmission delay extraction module is used to obtain the transmission delay time at different radial positions;

[0063] The neural network processing module is used to obtain 51 probability data points in the range of 0 to 1;

[0064] The MHD identification module is used to identify whether periodic disturbances exist;

[0065] The MHD localization module is used to transform the probability distribution of the convolutional neural network to obtain the radial position of the periodic perturbation.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks, characterized in that, The methods and steps include the following: S1. Based on the microwave reflection diagnostic data, the transmission delay time at different radial positions is obtained; S2. Obtain the spatial gradient of the transmission delay time and input it into the constructed convolutional neural network; S3. Analyze the results of S2. If there is a periodic disturbance, perform a correlation analysis between the delay time and the magnetic signal. If the maximum value of the correlation is greater than twice the average value, then execute S4. S4. Convert the results of the convolutional neural network into the radial positions of periodic perturbations.

2. The method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks according to claim 1, characterized in that, The transmission delay time at different radial positions is obtained in step S1, and the specific steps are as follows: A series of continuously varying frequency probe microwaves are emitted into the plasma using a swept-frequency microwave source. The emitted electromagnetic waves are reflected by a cutoff layer and received by a receiving antenna. The reflected signals are then processed to obtain... The propagation delay to position x, where u is the group velocity of the probe microwave, can be obtained directly from the plasma background parameters. Finally, the propagation delay time at different radial positions is obtained through an inversion algorithm.

3. The method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks according to claim 1, characterized in that, The spatial gradient of the transmission delay time in S2 is obtained specifically as follows: The gradient distribution data is spliced ​​together according to a certain time length to obtain a series of slices, that is, each slice is the evolution data of the transmission delay time gradient of a certain time length.

4. The method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks according to claim 1, characterized in that, The convolutional neural network constructed from the input in S2 is as follows: When tearing mode or twisting mode magnetohydrodynamic instability exists, there will be periodic perturbations in the transmission delay time gradient. The location of the perturbation is the radial location of the magnetohydrodynamic instability. Through pre-training with a large amount of data, the convolutional neural network identifies such periodic perturbations and constructs the convolutional neural network structure.

5. The method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks according to claim 4, characterized in that, The convolutional neural network structure includes a first layer, a second layer, a third layer, a fourth layer, a fifth layer, and a sixth layer; First layer: Input data is padded to make the output size equal to the input size, and then processed by the ReLU activation function; The second layer: obtain the corresponding data through the max pooling layer; The third layer: The corresponding data in the second layer is input data, and the output size is made equal to the input size by padding. Then, ReLU activation function is applied to obtain the processed data. Fourth layer: Obtain the corresponding data through the max pooling layer; Fifth layer: Input data is the corresponding data from the fourth layer, which is fully connected, then processed by the ReLU activation function, then dropped out, and finally convolutional data is obtained; The sixth layer: Input convolutional data, fully connected, and processed using the sigmoid activation function to obtain feature data.

6. The method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks according to claim 1, characterized in that, The specific analysis of the results in S3 is as follows: The convolutional neural network (CNN) outputs probability values ​​between 0 and 1. The first value represents whether there is a periodic perturbation in the time delay gradient, and the subsequent values ​​represent the probability distribution at different radial positions. If the maximum probability output by the neural network is at the first value, it indicates that there is no periodic perturbation in the time delay gradient. If the maximum probability output by the neural network appears in a subsequent value, it indicates that there is a periodic perturbation in the time delay gradient, and the position of the maximum probability is the radial position corresponding to the perturbation. If the results of the CNN indicate the presence of a periodic perturbation, correlation analysis is performed between the microwave reflection measurement time delay data and the magnetic signal. If there is a significant correlation, it indicates that this periodic perturbation in the time delay gradient is a magnetohydrodynamic instability.

7. The method for identifying and locating magnetohydrodynamic instabilities based on convolutional neural networks according to claim 1, characterized in that, Specifically, S4 is: the output of the convolutional neural network contains the probability distribution of the radial position of the magnetohydrodynamic instability, and the radial position of the magnetohydrodynamic instability can be determined based on the position of the maximum probability.

8. A device for identifying and locating magnetohydrodynamic instability based on a convolutional neural network, applied to the method described in any one of claims 1-7, the device comprising a diagnostic system composed of microwave diagnostics and magnetic probe diagnostics, a real-time data acquisition system, and a real-time identification and location system, characterized in that: The signals detected by the diagnostic system are transmitted to a real-time data acquisition system composed of data acquisition modules via a coaxial transmission line. The real-time data acquisition system is responsible for acquiring real-time microwave and magnetic signals. The real-time identification system is connected to the real-time data acquisition system.

9. The device for identifying and locating magnetohydrodynamic instability based on a convolutional neural network according to claim 8, characterized in that, The real-time recognition system includes a transmission delay extraction module, a neural network processing module, an MHD recognition module, and an MHD localization module. The transmission delay extraction module is used to obtain the transmission delay time at different radial positions; The neural network processing module is used to obtain probability data in the range of 0 to 1; The MHD identification module is used to determine whether periodic disturbances exist. The MHD positioning module is used to transform the probability distribution of the convolutional neural network to obtain the radial position of the periodic perturbation.

Images