Anti-interference intelligent identification method and system for magnetic fluid instability of fusion device

Abstract

The invention discloses an anti-interference intelligent identification method and system for magnetic fluid instability of a fusion device, and the method comprises the following steps: carrying out the secondary processing of original experimental diagnosis data, and forming a secondary database containing MHD physical mode information through the calculation of a physical formula; processing the physical label through a discontinuous data transformation algorithm to suppress interference, and obtaining a preliminarily corrected secondary database; constructing a neural network framework with high feature extraction capability; taking the original experimental diagnosis data as input data, taking the preliminarily corrected secondary data as label proxy learning, and obtaining an anti-interference neural network MHD recognition algorithm model; noise interference in mode signals is removed through the mapping characteristics of the model, and a secondary database with more remarkable physical characteristics is obtained. The anti-interference MHD mode recognition free of manual labeling is achieved, meanwhile, the problem of noise interference of a physical formula label is solved, and the method has the advantages of being high in recognition accuracy, good in objective and simple performance and the like.

Description

Technical Field

[0001] This application relates to the field of instability identification and control technology of tokamak plasma, specifically to an anti-interference intelligent identification method and system for magnetohydrodynamic instability of fusion devices. Background Technology

[0002] Magnetohydrodynamic (MHD) instability in plasmas within fusion devices is a macroscopic instability driven by factors such as current gradients, pressure gradients, and the curvature of the bad magnetic field. It has a large spatial scale and can cause significant changes in the plasma magnetic surface, leading to high energy loss and reduced confinement effectiveness. If not effectively controlled, it can cause rupture or even discharge termination, severely jeopardizing the device's safety. To effectively control and avoid MHD instability, accurate monitoring and spatial location of the MHD are necessary; this is a prerequisite for effective control.

[0003] Traditional MHD pattern recognition involves calculating physical formulas or performing MHD profile inversion based on diagnostic data.

[0004] Another approach to artificial intelligence is to use data / physical information-driven neural networks. The neural network model is trained to perceive and identify patterns by labeling the time periods in which MHD instability occurs with manually marked tags (2025PlasmaPhys.Control.Fusion67025023).

[0005] However, both methods have their own problems:

[0006] 1. Traditional methods based on physical calculations, such as time-frequency transformation, are designed only for the physical definition of instability modes, without considering electronic noise, non-MHD type plasma disturbances, etc., present in experiments. The measured diagnostic interference signals and mode disturbance signals may overlap, making it impossible to effectively distinguish them using fixed procedural rules such as threshold discrimination, which interferes with the accuracy of real-time identification. Taking the calculation of the n=1 mode magnetic disturbance amplitude by spatial Fourier decomposition as an example, due to measurement differences caused by factors such as installation errors among the Mirnov coils on the China Tokamak 3 (HL-3) device, and the nonlinear changes of these differences due to factors such as eddy currents, the low-modulus n=1 mode magnetic disturbance amplitude obtained by spatial Fourier decomposition usually has a component related to the plasma current distribution superimposed. This component causes a large overlap between the amplitudes calculated by spatial Fourier decomposition when there is n=1 mode magnetic disturbance and when there is no n=1 mode magnetic disturbance, making it difficult to automatically distinguish them using simple threshold methods.

[0007] 2. Recent studies have also shown many cases of using artificial intelligence algorithms to identify MHD instability patterns, which can solve the speed and interference problems of physical methods in real-time computing. However, this method requires manual annotation of the time periods of pattern occurrence in the data, which will bring a huge workload. At the same time, manual annotation inevitably has a large degree of subjectivity, making it difficult to widely promote to the identification of multiple instabilities. Summary of the Invention

[0008] The technical problem to be solved by this application is that the physical feature formula method is easily affected by noise in the traditional magnetohydrodynamic instability identification method, and the manual labeling method of the traditional artificial intelligence identification method is inefficient and highly subjective. The purpose is to provide an anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices, which solves the problem that the existing technology cannot achieve high reliability and real-time pattern recognition in complex plasma environments.

[0009] In a first aspect, this application provides an anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices, comprising the following steps:

[0010] Physical calculations are performed on the original experimental diagnostic data to form a secondary database containing information on magnetohydrodynamic physical models;

[0011] The secondary database is preliminarily processed using a discontinuous data transformation algorithm to objectively and selectively suppress interference factors, thereby obtaining a preliminarily corrected secondary database.

[0012] A neural network model was built, using the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels. The neural network model was then used for surrogate learning to obtain an anti-interference neural network recognition algorithm model.

[0013] By utilizing the characteristics of the mapping function fitted by the anti-interference neural network recognition algorithm model, noise correction is applied to the labels input to the neural network, and anti-interference magnetohydrodynamic instability recognition results are obtained based on the noise correction results.

[0014] A further optimized approach involves performing physical calculations on the original experimental diagnostic data to form a secondary database containing magnetohydrodynamic physical model information. This specifically includes the following steps:

[0015] Acquire raw magnetic diagnostic signal data to obtain the circumferential phase angle and real-time frequency parameters of each coil;

[0016] The original magnetic diagnostic signal is processed using a pattern feature extraction algorithm to extract the feature components of the corresponding pattern;

[0017] Based on the aforementioned feature components, a secondary database guided by physical features is constructed.

[0018] A further optimization scheme involves the application mode feature extraction algorithm processing the original magnetic diagnostic signal to extract feature components corresponding to the mode, specifically including the following steps:

[0019] Based on the spatial phase distribution of each detector, the original magnetic diagnostic signal is expanded by spectral analysis to obtain the mode feature expansion formula;

[0020] Calculate the amplitude of the characteristic coefficients corresponding to the target mode in the spectral analysis expansion;

[0021] The characteristic components of the corresponding mode are determined based on the amplitude of the characteristic coefficients.

[0022] A further optimization scheme involves using a discontinuous data transformation algorithm to perform preliminary processing on the secondary database, selectively suppressing interference factors, and obtaining a preliminarily corrected secondary database. This specifically includes the following steps:

[0023] The physical labels in the secondary database contain electromagnetic noise components, which include noise items and noise items coupled with magnetohydrodynamics.

[0024] Based on the electromagnetic noise components described above, the physical tag is preliminarily processed using a discontinuous data transformation algorithm to suppress the correlation mapping between interference signals and the tag, and to obtain the data transformation processing result.

[0025] Based on the data transformation and processing results, a preliminary corrected secondary database after interference suppression is generated.

[0026] A further optimization scheme involves processing the physical tags using a discontinuous data transformation algorithm, specifically including the following steps:

[0027] Slope calculation is performed on pattern feature signals in the secondary database to identify intervals with abnormal rates of change;

[0028] Amplitude limiting or truncation is applied to the amplitude of the identified abnormal intervals to obtain the pre-corrected pattern feature signal data.

[0029] Based on the preliminarily corrected mode characteristic signal data, data filtering is performed on the intervals where the plasma current change rate exceeds the threshold and is highly correlated with MHD instability, and physical tag data after filtering out the interference intervals is obtained.

[0030] A further optimized solution involves building a neural network model, using the original experimental diagnostic database as input data and a preliminarily corrected secondary database as labels. The neural network model is then used for surrogate learning to obtain an anti-interference neural network recognition algorithm model. This specifically includes the following steps:

[0031] Configure feature extraction network layers to build a neural network model;

[0032] Using forward and backward propagation algorithms, the original experimental diagnostic database is used as the input data for the neural network model, and the preliminarily corrected secondary database is used as the label signal for supervised learning. The neural network model is trained, and during the training process, irregular interference signals in the label signal are suppressed based on the fitting characteristics of the neural network model to regular mapping, thereby obtaining an anti-interference neural network recognition algorithm model.

[0033] A further optimization scheme is that the configuration feature extraction network layer specifically includes:

[0034] Configure convolutional neural network layers to extract local spatiotemporal features of magnetohydrodynamic diagnostic signals;

[0035] Configure residual network layers to increase network depth and maintain gradient stability;

[0036] Configure a global average pooling layer to replace the fully connected layer for feature dimensionality reduction;

[0037] Configure a fully connected layer for amplitude estimation of the output magnetohydrodynamic instability modes.

[0038] A further optimization scheme involves using the mapping function characteristics fitted by the anti-interference neural network recognition algorithm model to correct noise in the labels input to the neural network, and obtaining the anti-interference magnetohydrodynamic instability recognition result based on the noise correction result. This specifically includes the following steps:

[0039] Based on the inherent regularity mapping relationship learned by the neural network, the original experimental diagnostic data of magnetohydrodynamics is processed by pattern feature extraction to obtain feature data that highlights the key features of the magnetohydrodynamic mode.

[0040] Based on the characteristic that the mapping function can only identify signals with regularity and inherent connection, interference signal filtering is performed on the feature data to obtain purified magnetohydrodynamic signals and obtain a secondary database.

[0041] Based on the purified magnetohydrodynamic signal, the plasma state is monitored in real time through the pattern recognition results output by the neural network, and the magnetohydrodynamic instability identification results with anti-interference capabilities are obtained.

[0042] Secondly, this application provides an anti-interference intelligent identification system for magnetohydrodynamic instability of fusion devices, characterized in that it is used to implement the anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices as described in any of the preceding claims; the system includes:

[0043] The secondary database acquisition module is used to perform physical calculations on the raw experimental diagnostic data to form a secondary database containing information on magnetohydrodynamic physical modes.

[0044] The secondary database preliminary correction module is communicatively connected to the secondary database acquisition module. It is used to perform preliminary processing on the secondary database using a discontinuous data transformation algorithm, selectively suppress interference factors, and obtain a preliminary corrected secondary database.

[0045] The framework construction module communicates with the secondary database preliminary correction module. The original experimental diagnostic data secondary database uses a neural network layer consisting of convolutional layers, residual layers, global average pooling layers, and fully connected layers to construct a neural network anti-interference identification algorithm model for magnetohydrodynamic instability.

[0046] The model training module is communicatively connected to the framework building module. It is used to take the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels, and use the neural network model to perform surrogate learning to obtain an anti-interference neural network recognition algorithm model.

[0047] An anti-interference identification module, which is communicatively connected to the model training module, is used to correct the noise of the labels input to the neural network by using the characteristics of the mapping function fitted by the anti-interference neural network identification algorithm model, and to obtain the anti-interference magnetohydrodynamic instability identification result based on the noise correction result.

[0048] Secondly, this application provides a computer-readable storage medium storing an anti-interference intelligent identification program for magnetohydrodynamic instability of a fusion device, wherein when the anti-interference intelligent identification program for magnetohydrodynamic instability of a fusion device is executed by a processor, the program implements the steps of the anti-interference intelligent identification method for magnetohydrodynamic instability of a fusion device as described above.

[0049] Compared with the prior art, this application has the following advantages and beneficial effects:

[0050] By using the results of physical formula calculations and preliminary noise reduction as neural network labels, and by utilizing the fitting characteristics of neural networks to regular mappings, interference signals such as electromagnetic noise and non-MHD disturbances are effectively filtered out, significantly improving the accuracy and reliability of pattern recognition.

[0051] The method uses physical formulas to automatically generate training labels, which overcomes the problems of strong subjectivity, large workload and poor consistency caused by the reliance on manual annotation in traditional artificial intelligence methods, making label acquisition more objective and efficient.

[0052] This method uses the results of physical formula calculations as label signals for supervised learning, ensuring that the recognition model can retain the physical characteristics and evolutionary laws of MHD instability.

[0053] The trained neural network model can correct noise in the label signals, making the secondary database cleaner. This improves the accuracy of real-time MHD instability mode recognition and provides a foundation for end-to-end real-time recognition and control of MHD instability modes in fusion devices.

[0054] This method is not only applicable to the identification of n=1 modes, but can also be extended to the identification of other MHD instability modes, demonstrating strong versatility and scalability. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the exemplary embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0056] Figure 1 A flowchart illustrating the anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices provided in this application embodiment;

[0057] Figure 2 A flowchart of physical label correction processing based on discontinuous mapping data transformation provided in this application embodiment;

[0058] Figure 3 The effect diagram of the magnetic disturbance amplitude of the n=1 mode obtained by spatial Fourier decomposition calculation in the discharge experiment of the HL-3 device 2773 gun;

[0059] Figure 4 The image shows a comparison of the effects of applying the neural network anti-interference recognition algorithm described in this application to correct the same 2773 shot discharge data. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0061] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0062] MHD: Magnetohydrodynamics;

[0063] The HL-3 device, also known as the HL-3 tokamak, or by its project name "China Circulation 3", is China's newest and most advanced medium-sized tokamak nuclear fusion experimental device.

[0064] To address the limitations of traditional methods in removing interference and the difficulties in data labeling using artificial intelligence methods, this application combines these two approaches. Building upon artificial intelligence, it proposes a method that uses the results calculated and pre-processed using traditional physical formulas as labels for the artificial intelligence neural network. This method incorporates labels containing physical information, correcting for interference caused by diagnostic interference, plasma current state, and other factors while preserving the physical characteristics of the MHD instability modes. Furthermore, it avoids tedious and subjective manual labeling, making label acquisition more objective and convenient. This is a real-time, interference-resistant instability analysis and calculation method, providing a new technical means for the identification and control of HL-3 instability.

[0065] Firstly, such as Figure 1 As shown, this application provides an anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices, including the following steps:

[0066] Step S1: Construct a diagnostic database based on historical experimental data from the HL-3 device;

[0067] The original experimental diagnostic data were processed a second time, and a secondary database containing information on magnetohydrodynamic physical modes was calculated using the physical formula A=F(X).

[0068] The physical formula A=F(X) is an MHD feature formula based on physical principles. Taking n=1 mode recognition as an example, it is specifically a spatial Fourier decomposition formula:

[0069]

[0070] in, Let i be the signal value measured by the i-th Mirnov coil. Let i be the circumferential phase angle of the Mirnov coils. This represents the real-time frequency of the probe. The perturbation amplitude for the n=1 mode;

[0071] Specifically, it includes the following sub-steps:

[0072] (1) Obtain various relevant diagnostic system data required for MHD unstable pattern recognition from historical experimental data, and use them as input data X for traditional physical calculation and neural network.

[0073] For example, to identify n=1 mode magnetic disturbances, it is necessary to apply 12 channels of data from both the strong and weak field sides of the circumferential Mirnov magnetic disturbance coil;

[0074] (2) To ensure the consistency of measurement data across channels of the same diagnostic system, the original data needs to be calibrated.

[0075] For example, using the channel data X1 of the Mirnov coil without the MHD stage as a benchmark, linear regression calibration is performed on the data of the other 11 coils on the same side, that is, for the other 11 data X2~12=BX1+C, where B and C are constants;

[0076] (3) Use the physical characteristic formula of MHD instability to perform secondary processing on the original diagnostic data to form a secondary database (symbol A) containing MHD physical pattern information, that is, use the physical formula to calculate A=F(X);

[0077] Step S2: Based on the physical signals in the secondary database, perform preliminary processing on the discontinuous data transformation algorithm to selectively suppress interference factors while maintaining the integrity of the magnetohydrodynamic information structure, and obtain a preliminarily corrected secondary database;

[0078] Among these factors, due to uneven plasma current distribution, abrupt changes in plasma state, and local probe disturbances, interference terms will be superimposed in physical label A, namely... ;

[0079] By introducing a discontinuous data transformation algorithm to process label A, selective suppression of interference factors is implemented while maintaining the integrity of the target domain MHD information structure: the MHD pattern characteristics are maintained at the feature mapping level, while the correlation mapping between interference items and labels is broken through the data transformation strategy.

[0080] Step S3: Construct a neural network model. Use the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels. Use the neural network model for surrogate learning to obtain an anti-interference neural network recognition algorithm model. Specifically, use highly feature-extracting neural network layers and auxiliary layers, such as convolutional neural networks, residual networks, global average pooling networks, and fully connected networks, to construct the neural network model. Use the original experimental diagnostic database of the device as the learning signal of the neural network. Use the secondary database calculated by physical formulas for MHD unstable modes and preliminarily corrected as labels. Use the neural network model for surrogate learning to obtain an anti-interference neural network recognition algorithm model.

[0081] Step S4: Using the mapping function characteristics fitted by the anti-interference neural network recognition algorithm model, noise correction is applied to the labels input to the neural network. This eliminates interference signals in the initially corrected secondary database data, making the physical features of the secondary database more significant and the data purer. This improves the accuracy of real-time MHD unstable pattern recognition and provides a foundation for end-to-end recognition and control of MHD unstable patterns throughout the entire process.

[0082] like Figure 2 The diagram shows a flowchart of the physical label correction process based on a neural network provided in this application embodiment. The specific process is as follows: ① A feature label dataset is generated by analyzing historical experimental data through physical formulas, thus obtaining a coarse secondary database; ② The pre-processed MHD instability feature database is used as a label for the neural network model and combined with the input data to train the anti-interference recognition model of the neural network; ③ The trained model can correct the labels, eliminate noise interference in the pattern features, and obtain a purer secondary database of physical features. It can also be applied to the real-time control loop, receiving data from the fusion device diagnostic system, using the neural network MHD instability anti-interference recognition algorithm model to perceive the occurrence of instability modes, providing a decision-making basis for the central control system, thereby driving the auxiliary heating system to control the MHD instability modes and avoid the occurrence of large plasma fractures.

[0083] In one embodiment, step S1: performing secondary processing on the original experimental diagnostic data to calculate and form a secondary database containing magnetohydrodynamic physical mode information, specifically includes the following steps:

[0084] Step S11: Collect raw circumferential Mirnov coil data and obtain the circumferential phase angle and real-time frequency parameters of each coil;

[0085] The real-time frequency is indirectly obtained by calculating the wavelength that changes in real time between the peaks of the circumferential Mirnov coil signal.

[0086] Step S12: Apply the spatial Fourier decomposition formula to process the original circumferential Mirnov coil data and extract the first harmonic component as the magnetic disturbance amplitude of the n=1 mode;

[0087] Step S13: Based on the first harmonic component, construct a secondary database guided by physical characteristics.

[0088] In one embodiment, step S12, which involves processing the original circumferential Mirnov coil data using the spatial Fourier decomposition formula to extract the first harmonic component, specifically includes the following steps:

[0089] Step S121: Based on the circumferential phase angle of each Mirnov coil, perform a spatial Fourier expansion on the original circumferential Mirnov coil signal to obtain the spatial Fourier expansion.

[0090] Step S122: Calculate the amplitude of the coefficient corresponding to the first harmonic in the spatial Fourier expansion;

[0091] Step S123: Determine the first harmonic component based on the amplitude of the coefficient.

[0092] In one embodiment, step S2: The secondary database is preliminarily processed using a discontinuous data transformation algorithm to selectively suppress interference factors while maintaining the integrity of the magnetohydrodynamic information structure, thereby obtaining a preliminarily corrected secondary database. This specifically includes the following steps:

[0093] Step S21: Identify the electromagnetic noise components contained in the physical tags in the secondary database, the electromagnetic noise components including noise items and noise items coupled with magnetohydrodynamics;

[0094] Step S22: Based on the identified interference components, the physical feature signals of the secondary database are preliminarily processed using a discontinuous data transformation algorithm strategy. While maintaining the integrity of the magnetohydrodynamic mode features, the correlation mapping between the interference signals and the labels is suppressed, and the data transformation processing results are obtained to facilitate further noise filtering by the subsequent neural network.

[0095] Step S23: Based on the data transformation processing results, generate a preliminary corrected secondary database.

[0096] In one embodiment, the processing of physical tags using a discontinuous data transformation algorithm strategy in step S22 specifically includes:

[0097] Step S221: Calculate the slope of the first harmonic component of the pattern feature signal in the secondary database and identify the intervals with abnormal rate of change.

[0098] Specifically targeting non-n=1 mode spike interference during the plasma breakdown ramp-up and plasma breakup phases;

[0099] Step S222: Apply amplitude limiting or truncation processing to the identified abnormal interval amplitude to obtain the first harmonic component data after preliminary correction;

[0100] Step S223: Data filtering is performed on the intervals where the plasma current change rate exceeds the threshold and is highly correlated with MHD instability, and physical tag data after filtering out the interference intervals is obtained.

[0101] In one embodiment, step S3: building a neural network model, using the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels, and using the neural network model for surrogate learning to obtain an anti-interference neural network recognition algorithm model, specifically includes the following steps:

[0102] Step S31: Standardize the preliminary corrected secondary database, including taking the logarithm of the data, uniform standardization and smoothing, so that the dataset is suitable for neural network training in a relatively stable gradient environment, and obtain the standardized preliminary corrected secondary database.

[0103] Because traditional AI pattern recognition labels (label range 0~1) are not used, but instead the amplitude of the MHD instability mode calculated by physical formula is used as label A, the numerical size of the dataset covers an excessively wide range. This excessively wide distribution will bring greater instability to the training of the neural network. Therefore, the data is standardized and smoothed after taking the logarithm.

[0104] These processing steps enable the data from each shot to be used in a relatively stable gradient environment;

[0105] Step S32: Construct a feature extraction network layer that includes a convolutional neural network, a residual network, a global average pooling network, and a fully connected network. Use the original experimental diagnostic database as input data and the standardized, preliminarily corrected secondary database as labels to form a neural network model.

[0106] Step S33: Based on the original experimental diagnostic data and the neural network model, train the framework to fit the nonlinear mapping relationship between the input data and the output label, and finally obtain an anti-interference neural network recognition algorithm model that can describe the characteristics of magnetohydrodynamic instability modes.

[0107] In one embodiment, the construction of the feature extraction network layer in step S32 specifically includes: step S321: configuring a convolutional neural network layer to extract the local spatiotemporal features of the circumferential Mirnov coil signal;

[0108] Step S322: Configure residual network layers to increase network depth and maintain gradient stability;

[0109] Step S323: Configure a global average pooling layer to replace the fully connected layer for feature dimensionality reduction;

[0110] Step S324: Configure the fully connected layer for amplitude estimation of the output magnetohydrodynamic instability mode.

[0111] In one embodiment, step S33: Based on the original experimental diagnostic data and the neural network model, the framework is trained to fit the nonlinear mapping relationship between the input data and the output label, and finally obtains an anti-interference neural network recognition algorithm model that can describe the characteristics of magnetohydrodynamic instability modes. This specifically includes the following steps:

[0112] Based on the established neural network model, forward propagation and backward propagation algorithms are used to train the neural network model with the input data and label signals. During the training process, irregular interference signals in the label signals are suppressed according to the fitting characteristics of the neural network model to regular mapping, thereby obtaining an anti-interference neural network recognition algorithm model.

[0113] In one embodiment, step S4 involves using the mapping function characteristics fitted by the anti-interference neural network recognition algorithm model to perform noise correction on the labels input to the neural network, i.e., eliminating interference signals in the initially corrected secondary database data to obtain a secondary database with more significant physical features and purer data. Specifically, this includes the following steps:

[0114] Step S41: Based on the inherent regularity mapping relationship learned by the neural network, perform pattern feature extraction processing on the original experimental diagnostic data characterizing MHD instability to obtain feature-enhanced data that highlights the key features of the magnetohydrodynamic mode.

[0115] Step S42: Based on the characteristic that the mapping function can only identify signals with regularity and inherent connection, interference signal filtering is performed on the feature enhancement data to obtain purified magnetohydrodynamic signals, thereby obtaining a secondary database with more significant physical characteristics and purer data.

[0116] Step S43: Based on the purified magnetohydrodynamic signal, the accuracy of real-time MHD instability mode recognition can be improved. Real-time plasma state monitoring is performed using the pattern recognition results output by the neural network, providing a decision-making basis for subsequent real-time control and laying the foundation for end-to-end real-time recognition and control of MHD instability modes throughout the entire process. This invention patent takes the n=1 mode intelligent anti-interference recognition of the China Circulator No. 3 (HL-3) device as an example, and the specific steps are as follows:

[0117] (1) Collect 2178 effective discharge experiments of the HL-3 device to form a historical experimental database, and collect diagnostic data X from 12 probes on each of the strong and weak field sides of the circumferential Mirnov device.

[0118] (2) By performing spatial Fourier decomposition on the collected original dataset X, the dataset of the n=1 mode magnetic disturbance amplitude A after secondary processing can be obtained.

[0119] by Figure 3Taking the discharge of the 2773 gun as an example, it can be seen that a frequency fluctuation band appeared in the 1-4kHz frequency band of the spectrum after about 180ms. In the original probe signal NDBPL during the corresponding time period, the changes in frequency and amplitude can be clearly seen. The amplitude curves of the strong field side (A_H) and weak field side (A_L) n=1 mode obtained by spatial Fourier decomposition can intuitively show that the n=1 mode tearing mode appeared after 180ms and continued to develop until the plasma lock mode broke up at 680ms. However, according to the image, there are many intermittent high-amplitude spike interferences before the tearing mode appears (0~180ms) and during the breakup (680~710ms). This makes it difficult for the real-time control stage to determine the n=1 mode and control accuracy through threshold setting, etc. Therefore, measures need to be taken to correct the interference.

[0120] A series of feature extraction neural networks, including convolutional layers, were employed. Twelve data points collected from the circumferential Mirnov coil were used as the learning data input to the neural network. Strong and weak field side n=1 mode magnetic perturbation data were obtained using spatial Fourier decomposition. Thresholds were set for the amplitude and growth rate, and regions with rapid changes in plasma current were filtered out. Standardized data were used as neural network labels. Data from the same 2773 shot discharge was selected. Figure 4 The diagram shows the effect after neural network correction. It can be seen that the neural network correction (NN) curve in the figure is well corrected for the interference peaks on the strong and weak fields during the early stage of discharge (0~180ms) and the rupture period (680~710ms). Furthermore, it has a good fitting effect on the mode disturbance trend during the n=1 mode, reflecting the amplitude change of mode development. This allows the real-time control stage to determine the n=1 mode and improve control accuracy through threshold setting and other methods.

[0121] Secondly, this application provides an anti-interference intelligent identification system for magnetohydrodynamic instability of fusion devices to implement the anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices as described above; the system includes:

[0122] The secondary database acquisition module is used to perform physical calculations on the raw experimental diagnostic data to form a secondary database containing information on magnetohydrodynamic physical modes.

[0123] The secondary database preliminary correction module is communicatively connected to the secondary database acquisition module. It is used to perform preliminary processing on the secondary database using a discontinuous data transformation algorithm. While maintaining the integrity of the magnetohydrodynamic information structure, it selectively suppresses interference factors to obtain a preliminary corrected secondary database.

[0124] The framework construction module communicates with the secondary database preliminary correction module and uses a neural network layer consisting of convolutional layers, residual layers, global average pooling layers, and fully connected layers to construct a neural network anti-interference identification algorithm model for magnetohydrodynamic instability from the original experimental diagnostic data secondary database.

[0125] The model training module is communicatively connected to the framework building module. It is used to take the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels, and use the neural network model to perform surrogate learning to obtain an anti-interference neural network recognition algorithm model.

[0126] The anti-interference identification module, communicatively connected to the model training module, is used to correct noise in the labels input to the neural network by utilizing the mapping function characteristics fitted by the anti-interference neural network identification algorithm model. This involves eliminating interference signals in the initially corrected secondary database data, resulting in a secondary database with more significant physical features and cleaner data. This improves the accuracy of real-time MHD unstable pattern recognition and provides a foundation for end-to-end real-time identification and control of MHD unstable patterns throughout the entire process.

[0127] The beneficial effects of this invention are as follows:

[0128] 1. The result of the MHD feature formula based on physical principles is used as label A. A deep learning neural network is used to learn it by proxy. At the same time, the characteristic that the mapping function fitted by the neural network must be regular and have an inherent connection is taken into account, so that information such as interference signals cannot be learned by the neural network, thereby achieving the effect of anti-interference calculation.

[0129] This approach is more convenient and efficient than manual annotation and can preserve physical information. It does not introduce bias into the model due to human subjective factors and can support the training of a more reliable instability recognition model.

[0130] 2. This method uses some intermediate processing to make the interference labels change discontinuously, thus suppressing the consistency of interference information in the function mapping.

[0131] Neural networks can eliminate interfering data due to inconsistencies in the labels of interfering signals.

[0132] If only the secondary data results of physical calculations are used as labels, the neural network, as a proxy model for the physical calculation process, only learns the mapping of physical formulas and cannot autonomously eliminate interference.

[0133] This method can effectively filter out interfering labels during model learning and fit the correct mode perturbation function;

[0134] 3. Use the original experimental diagnostic data and the data calculated and preliminarily processed by the physical formula as input and labels for the artificial intelligence algorithm model, so that the model fitted by the neural network algorithm contains physical information;

[0135] 4. No manual data labeling is required for each shot in the early stage. The acquisition of labels is more objective and convenient, realizing a real-time and interference-resistant instability analysis and calculation method, providing a new technical means for the identification and control of instability of the HL-3 device;

[0136] 5. Based on artificial intelligence, we further use neural networks to learn from the results of physical formula calculations and preliminary processing. This method can not only preserve the perturbation mode of MHD instability, but also correct the interference caused by diagnostic interference, plasma current state and other factors, and obtain a secondary database with more significant physical characteristics of MHD instability.

[0137] The functions of each module in the above-mentioned intelligent identification system for anti-interference of magnetohydrodynamic instability of fusion device correspond to the steps in the above-mentioned intelligent identification method for anti-interference of magnetohydrodynamic instability of fusion device, and their functions and implementation processes will not be described in detail here.

[0138] Thirdly, embodiments of this application provide an anti-interference intelligent identification device for the magnetohydrodynamic instability of a fusion device. The anti-interference intelligent identification device for the magnetohydrodynamic instability of a fusion device can be a personal computer (PC), a laptop computer, a server, or other device with data processing capabilities.

[0139] In this embodiment, the anti-interference intelligent identification device for magnetohydrodynamic instability of the fusion device may include a processor, a memory, a communication interface, and a communication bus.

[0140] The communication bus can be of any type and is used to interconnect the processor, memory, and communication interface.

[0141] The communication interface includes input / output (I / O) interfaces, physical interfaces, and logical interfaces. These interfaces are used for interconnecting internal components of the anti-interference intelligent identification device for magnetohydrodynamic instability in fusion devices, as well as for interconnecting the anti-interference intelligent identification device for magnetohydrodynamic instability in fusion devices with other devices (such as other computing devices or user equipment). Physical interfaces can be Ethernet interfaces, fiber optic interfaces, ATM interfaces, etc.; user equipment can be displays, keyboards, etc.

[0142] Memory can be various types of storage media, such as random access memory (RAM), read-only memory (ROM), non-volatile RAM (NVRAM), flash memory, optical storage, hard disk, programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), etc.

[0143] The processor can be a general-purpose processor, which can call the anti-interference intelligent identification program for the magnetohydrodynamic instability of the fusion device stored in the memory and execute the anti-interference intelligent identification method for the magnetohydrodynamic instability of the fusion device provided in the embodiments of this application. For example, the general-purpose processor can be a central processing unit (CPU). The method executed when the anti-interference intelligent identification program for the magnetohydrodynamic instability of the fusion device is called can refer to the various embodiments of the anti-interference intelligent identification method for the magnetohydrodynamic instability of the fusion device in this application, and will not be repeated here.

[0144] Fourthly, embodiments of this application also provide a readable storage medium.

[0145] The present application stores an anti-interference intelligent identification program for the magnetohydrodynamic instability of a fusion device on a readable storage medium. When the anti-interference intelligent identification program for the magnetohydrodynamic instability of a fusion device is executed by a processor, it implements the steps of the anti-interference intelligent identification method for the magnetohydrodynamic instability of a fusion device as described above.

[0146] The method implemented when the anti-interference intelligent identification program for magnetohydrodynamic instability of fusion device is executed can be referred to in various embodiments of the anti-interference intelligent identification method for magnetohydrodynamic instability of fusion device in this application, and will not be repeated here.

[0147] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent identification of magnetohydrodynamic instability in a fusion device, characterized in that, Includes the following steps: Physical calculations are performed on the original experimental diagnostic data to form a secondary database containing information on magnetohydrodynamic physical models; The secondary database is preliminarily processed using a discontinuous data transformation algorithm to selectively suppress interference factors and obtain a preliminarily corrected secondary database. A neural network model was built, using the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels. The neural network model was then used for surrogate learning to obtain an anti-interference neural network recognition algorithm model. By utilizing the characteristics of the mapping function fitted by the anti-interference neural network recognition algorithm model, noise correction is applied to the labels input to the neural network, and anti-interference magnetohydrodynamic instability recognition results are obtained based on the noise correction results.

2. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 1, characterized in that, The process of performing physical calculations on the original experimental diagnostic data to form a secondary database containing information on magnetohydrodynamic physical models specifically includes the following steps: Acquire raw magnetic diagnostic signal data to obtain the circumferential phase angle and real-time frequency parameters of each coil; The original magnetic diagnostic signal is processed using a pattern feature extraction algorithm to extract the feature components of the corresponding pattern; Based on the aforementioned feature components, a secondary database guided by physical features is constructed.

3. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 2, characterized in that, The application mode feature extraction algorithm processes the original magnetic diagnostic signal to extract the feature components of the corresponding mode, specifically including the following steps: Based on the spatial phase distribution of each detector, the original magnetic diagnostic signal is expanded by spectral analysis to obtain the mode feature expansion formula; Calculate the amplitude of the characteristic coefficients corresponding to the target mode in the spectral analysis expansion; The characteristic components of the corresponding mode are determined based on the amplitude of the characteristic coefficients.

4. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 1, characterized in that, The secondary database is preliminarily processed using a discontinuous data transformation algorithm to selectively suppress interference factors, thereby obtaining a preliminarily corrected secondary database. This process specifically includes the following steps: The physical labels in the secondary database contain electromagnetic noise components, which include noise items and noise items coupled with magnetohydrodynamics. Based on the electromagnetic noise components described above, the physical tag is preliminarily processed using a discontinuous data transformation algorithm to suppress the correlation mapping between interference signals and the tag, and to obtain the data transformation processing result. Based on the data transformation and processing results, a preliminary corrected secondary database after interference suppression is generated.

5. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 4, characterized in that, The process of processing physical tags using a discontinuous data transformation algorithm specifically includes the following steps: Slope calculation is performed on pattern feature signals in the secondary database to identify intervals with abnormal rates of change; Amplitude limiting or truncation is applied to the amplitude of the identified abnormal intervals to obtain the pre-corrected pattern feature signal data. Based on the preliminarily corrected mode characteristic signal data, data filtering is performed on the intervals where the plasma current change rate exceeds the threshold and is highly correlated with MHD instability, and physical tag data after filtering out the interference intervals is obtained.

6. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 1, characterized in that, The construction of the neural network model, using the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels, and employing the neural network model for surrogate learning to obtain an anti-interference neural network recognition algorithm model, specifically includes the following steps: Configure feature extraction network layers to build a neural network model; Using forward and backward propagation algorithms, the original experimental diagnostic database is used as the input data for the neural network model, and the preliminarily corrected secondary database is used as the label signal for supervised learning. The neural network model is trained, and during the training process, irregular interference signals in the label signal are suppressed based on the fitting characteristics of the neural network model to regular mapping, thereby obtaining an anti-interference neural network recognition algorithm model.

7. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 6, characterized in that, The configuration feature extraction network layer specifically includes: Configure convolutional neural network layers to extract local spatiotemporal features of magnetohydrodynamic diagnostic signals; Configure residual network layers to increase network depth and maintain gradient stability; Configure a global average pooling layer to replace the fully connected layer for feature dimensionality reduction; Configure a fully connected layer for amplitude estimation of the output magnetohydrodynamic instability modes.

8. The anti-interference intelligent identification method for magnetohydrodynamic instability of fusion devices according to claim 1, characterized in that, The step of using the mapping function characteristics fitted by the anti-interference neural network identification algorithm model to correct noise in the labels input to the neural network, and obtaining the anti-interference magnetohydrodynamic instability identification result based on the noise correction result, specifically includes the following steps: Based on the inherent regularity mapping relationship learned by the neural network, the original experimental diagnostic data of magnetohydrodynamics is processed by pattern feature extraction to obtain feature data that highlights the key features of the magnetohydrodynamic mode. Based on the characteristic that the mapping function can only identify signals with regularity and inherent connection, interference signal filtering is performed on the feature data to obtain purified magnetohydrodynamic signals and obtain a secondary database. Based on the purified magnetohydrodynamic signal, the plasma state is monitored in real time through the pattern recognition results output by the neural network, and the magnetohydrodynamic instability identification results with anti-interference capabilities are obtained.

9. An anti-interference intelligent identification system for magnetohydrodynamic instability of a fusion device, characterized in that, A method for implementing anti-interference intelligent identification of magnetohydrodynamic instability in fusion devices according to any one of claims 1-8; the system comprises: The secondary database acquisition module is used to perform physical calculations on the raw experimental diagnostic data to form a secondary database containing information on magnetohydrodynamic physical modes. The secondary database preliminary correction module is communicatively connected to the secondary database acquisition module. It is used to perform preliminary processing on the secondary database using a discontinuous data transformation algorithm, selectively suppress interference factors, and obtain a preliminary corrected secondary database. The framework construction module communicates with the secondary database preliminary correction module. The original experimental diagnostic data secondary database uses a neural network layer consisting of convolutional layers, residual layers, global average pooling layers, and fully connected layers to construct a neural network anti-interference identification algorithm model for magnetohydrodynamic instability. The model training module is communicatively connected to the framework building module. It is used to take the original experimental diagnostic database as input data and the preliminarily corrected secondary database as labels, and use the neural network model to perform surrogate learning to obtain an anti-interference neural network recognition algorithm model. An anti-interference identification module, which is communicatively connected to the model training module, is used to correct the noise of the labels input to the neural network by using the characteristics of the mapping function fitted by the anti-interference neural network identification algorithm model, and to obtain the anti-interference magnetohydrodynamic instability identification result based on the noise correction result.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an anti-interference intelligent identification program for the magnetohydrodynamic instability of a fusion device, wherein when the anti-interference intelligent identification program for the magnetohydrodynamic instability of a fusion device is executed by a processor, the steps of the anti-interference intelligent identification method for the magnetohydrodynamic instability of a fusion device as described in any one of claims 1 to 8 are implemented.

Images