Tokamak temperature diagnosis reconstruction method and device based on data-driven model

Abstract

The invention discloses a Tokamak temperature diagnosis reconstruction method and device based on a data-driven model, and belongs to the technical field of machine learning and nuclear fusion physical crossover, and the method comprises the steps: firstly, carrying out the training of a data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism through the large-scale Tokamak experiment data, and carrying out the reconstruction of the Tokamak temperature diagnosis. The complex correlation between various diagnosis signals and electronic temperature diagnosis is learned; secondly, an efficient data processing mode is provided, tokamak original signals are divided into equal-length sub-data in a window division mode, and the data are packaged and stored in an array form after being subjected to data processing so that the data can be directly used for further processing, reading, training and reasoning of subsequent machine learning; finally, the model can be seamlessly integrated into an existing Tokamak data management system by excluding a control reference signal. According to the invention, relatively accurate electronic temperature diagnosis reference can be provided.

Description

Technical Field

[0001] The present invention belongs to the cross - technical field of machine learning and fusion physics, and specifically relates to a Tokamak temperature diagnosis and reconstruction method and device based on a data - driven model. Background Art

[0002] Magnetic - confinement fusion, as an important technical route to achieve clean energy, plays a core role in the research of high - temperature plasma physics in Tokamak devices.

[0003] As the main experimental platform for magnetic - confinement fusion research, the accurate measurement of the electron temperature during the operation of a Tokamak device is the core requirement for understanding plasma energy transport, verifying physical models, and optimizing confinement performance.

[0004] Traditional electron temperature diagnosis relies on physical methods such as Thomson scattering and electron cyclotron radiation. However, these methods are limited by extreme experimental environments (such as high - energy particle irradiation and electromagnetic interference) and device geometric constraints, often resulting in the lack of temperature data or a decrease in reliability during key discharge stages.

[0005] Although the inversion method based on physical models (such as solving transport equations) can supplement diagnostic data, it has high computational complexity, poor real - time performance, and relies on simplified assumptions, resulting in significant errors during sudden plasma instabilities or abnormal diagnostic signals.

[0006] In recent years, data - driven methods have provided new ideas for real - time reconstruction by establishing the mapping relationship between diagnostic signals and temperature parameters. However, traditional machine - learning models (such as linear regression and support vector machines) are difficult to handle the high - dimensionality, strong non - linearity, and long - range spatio - temporal correlation characteristics of Tokamak data. Existing deep - learning methods still have deficiencies in long - sequence processing efficiency, model generalization ability, and system integration adaptability.

[0007] The discharge durations of Tokamak experiments are not the same, and the sampling rates of different signals also vary to a certain extent. This results in significant differences in the lengths of the time series of the entire - process Tokamak discharge experiments, and these data need to go through a cumbersome data - processing process to be used in machine - learning models. Conventional machine - learning models cannot directly train them, that is, the existing technology lacks a method for reconstructing the missing electron temperature diagnosis of Tokamak experimental data. Summary of the Invention

[0008] To solve the above - mentioned technical problems, the present invention adopts the following technical solutions:

[0009] A Tokamak temperature diagnosis and reconstruction method based on a data - driven model, which reconstructs the plasma electron temperature diagnosis in EAST by training a data - driven model that fuses a recurrent neural network and a self - attention mechanism, specifically including:

[0010] Step 1, Obtaining original data: Obtain the original actuator experiment data and the actually diagnosed experiment data of EAST;

[0011] Step 2, Data selection and processing: Select several complete real discharge experiment data suitable for machine learning training from the EAST experiment data as the training and inference data of the machine learning model;

[0012] Step 3, Conducting machine learning model training: Input the signal after data processing into the machine learning model. This machine learning model gives a reconstructed output. Construct a loss function for the machine learning reconstructed output and the actual measurement output of EAST and calculate the loss. Then, train the machine learning model through the error backpropagation method. The trained data-driven model based on the fusion of the recurrent neural network and the self-attention mechanism can directly reconstruct the missing electron temperature diagnosis;

[0013] Step 4, Evaluating the inference performance of the machine learning model: Use the data obtained from the data processing in Step 2 as the model input, input it into the data-driven model based on the fusion of the recurrent neural network and the self-attention mechanism trained in Step 3, and use this model for inference to directly reconstruct the electron temperature diagnosis. Respectively change the number of layers, batch size of the model, and whether to use a bidirectional RNN network to construct models for inference verification, and analyze the losses of its validation set and test set to evaluate the performance of the machine learning model with this hyperparameter configuration, so as to obtain the machine learning model with the optimal hyperparameters;

[0014] Step 5, Coupling the machine learning model and the EAST data management system: Use the machine learning model obtained in Step 4 to directly reconstruct the missing electron temperature diagnosis and input the reconstruction result as the missing electron temperature reference into the EAST data management system.

[0015] A Tokamak temperature diagnosis reconstruction device based on a data-driven model, which reconstructs the plasma electron temperature diagnosis in EAST by training a data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism, specifically including:

[0016] Original data acquisition module: Obtain the original actuator experiment data and the actually diagnosed experiment data of EAST;

[0017] Data selection and processing module: Select several complete real discharge experiment data suitable for machine learning training from the EAST experiment data as the training and inference data of the machine learning model;

[0018] Machine learning model training module: Input the signal after data processing into the machine learning model. The machine learning model gives a reconstructed output. Construct a loss function for the machine learning reconstructed output and the actual measured output of EAST and calculate the loss. Then, train the machine learning model through the error backpropagation method. The trained data-driven model based on the fusion of the recurrent neural network and the self-attention mechanism can directly reconstruct the missing electron temperature diagnosis;

[0019] Machine learning model inference performance evaluation module: Use the data obtained from the data selection and processing module in the data processing as the model input, input it into the data-driven model based on the fusion of the recurrent neural network and the self-attention mechanism trained by the machine learning model training module, and use this model for inference to directly reconstruct the electron temperature diagnosis. Respectively change the number of layers, batch size of the model, and whether to use a bidirectional RNN network to construct models for inference verification, analyze the losses of its validation set and test set to evaluate the performance of the machine learning model with this hyperparameter configuration, so as to obtain the machine learning model with the optimal hyperparameters;

[0020] Coupling module of the machine learning model and the EAST data management system: Use the machine learning model obtained by the machine learning model inference performance evaluation module to directly reconstruct the missing electron temperature diagnosis and input the reconstruction result as the missing electron temperature reference into the EAST data management system.

[0021] An electronic device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the steps of the Tokamak temperature diagnosis reconstruction method based on the data-driven model.

[0022] A non-transitory computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the steps of the Tokamak temperature diagnosis reconstruction method based on the data-driven model.

[0023] The present invention has the following beneficial effects:

[0024] The present invention proposes a new data-driven model based on the fusion of the recurrent neural network and the self-attention mechanism. By introducing a sliding window segmentation strategy, the discharge data of several seconds in length is decomposed into subsequence units with clear physical meanings, which not only retains the dynamic characteristics of plasma evolution but also avoids the long sequence training problem of traditional recurrent neural networks. The non-sequential parallel processing architecture of the present invention, combined with causal convolution and sparse attention mechanism, reduces the computational complexity to a linear order while ensuring time causality.

[0025] Compared with the prior art, the present invention uses a Tokamak electron temperature diagnostic reconstruction method based on a data-driven model. When reconstructing the missing electron temperature diagnosis, the model of the present invention can ensure that the reconstruction error of the entire discharge process does not exceed twice the standard deviation of the true value with a confidence level of more than 95%, meeting the actual application requirements. Through the Tokamak electron temperature diagnostic reconstruction method based on a data-driven model, the model can be seamlessly integrated into the existing Tokamak data management system. In the absence of reliable temperature diagnosis, relatively accurate electron temperature diagnostic references can be provided through model reconstruction, significantly improving data availability and research efficiency. In addition, the data processing strategy proposed by the present invention effectively solves the input problem of long-time series data through window partitioning and normalization processing, ensuring the efficient training and inference of the model. The scalability of the model also makes it possible to apply it to the reconstruction of other important physical quantities, further expanding its application scope in Tokamak experiments. Compared with traditional physical simulation methods, the present invention shows significant advantages in reconstruction accuracy, computational efficiency, and system integration, providing strong technical support for fusion physics research and device operation control. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a schematic flow chart of data acquisition, processing, and model inference of the present invention;

[0027] Figure 2 It is an example diagram of the reconstruction results of the evolution of the electron temperature diagnosis of the machine learning model of the present invention under different toroidal magnetic fields. Among them, (a) is an example diagram of the reconstruction results of the evolution of the electron temperature diagnosis of the model of the present invention under a toroidal magnetic field of 1.58 T, (b) is an example diagram of the reconstruction results of the evolution of the electron temperature diagnosis of the model of the present invention under a toroidal magnetic field of 2.41 T, (c) is an example diagram of the reconstruction results of the evolution of the electron temperature diagnosis of the model of the present invention under a toroidal magnetic field of 2.19 T, and (d) is a schematic diagram of the reconstruction results of the evolution of the electron temperature diagnosis of the model of the present invention under a toroidal magnetic field of 2.43 T;

[0028] Figure 3 It is a residual result diagram of the test set experimental data of the 374th shot in the discharge experiment of the machine learning model of the present invention in the Tokamak experiment with shot numbers in the range of #82877 - 99092;

[0029] Figure 4 It is an infinity-norm result diagram of the test set experimental data of the 374th shot in the discharge experiment of the machine learning model of the present invention in the Tokamak experiment with shot numbers in the range of #82877 - 99092;

[0030] Figure 5Coverage distribution plot of the reconstruction result error of the entire discharge process of the test set experimental data of 374 discharges in the discharge experiment of the machine learning model of the present invention in the Tokamak experimental gun number within #82877-99092 not exceeding twice the standard deviation of the true value. Detailed implementation manners

[0031] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0032] The present invention has significant advantages in preprocessing the time series data of the all-superconducting Tokamak fusion experimental device (EAST) and saving it in an array format. First of all, as a binary format, the array format supports efficient storage and fast reading. There is no need to parse text or convert types, which can significantly reduce the input / output time consumption. Secondly, the data is stored in a multi-dimensional array format and can be directly loaded as a tensor into a machine learning model (such as TensorFlow / PyTorch), avoiding the overhead of real-time decoding during training and improving the training efficiency. In addition, the array file retains meta-information such as data dimensions and types, ensuring cross-platform reading consistency and supporting memory mapping to process ultra-large datasets. The method of preprocessing the EAST time series data and saving it in an array format can achieve end-to-end data processing, simplifies the data management process, and is especially suitable for large-scale time series analysis and model deployment scenarios.

[0033] The present invention provides a data-driven model that does not rely on physical simulation codes and can accurately reconstruct the missing electron temperature diagnosis of Tokamak experimental data. This method realizes the real-time high-precision reconstruction of the missing electron temperature diagnosis in EAST through a data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism, combined with an efficient data processing strategy, providing key parameter support for plasma physics research and device operation control.

[0034] The processes of data acquisition, processing, and model inference of the present invention are as Figure 1 shown. The present invention sorts a large amount of Tokamak historical data into a data format that can be directly read and directly input into a machine learning model through an efficient data processing strategy. Specifically, as shown in the general flow chart of the data acquisition process in Figure 1 , the present invention reads the electron density diagnosis data from the hydrogen cyanide (HCN) laser interferometer in EAST respectively, obtains plasma current, toroidal magnetic field, loop voltage, heating ratio signals, etc. from magnetic diagnostics, and multi-channel diagnostic signals (such as β, κ, q, etc.) from the off-line equilibrium inversion program (EFIT) as the input of the machine learning model; as shown in Figure 1As shown in the flowchart of acquiring electron temperature data, the present invention reads, from EAST, the electron temperature diagnostic data that has been further analyzed and processed by the heterodyne radiometer system (HRS) as the input of the machine learning model; as Figure 1 As shown in the schematic diagram of data processing, encapsulation, and model inference, the present invention further smooths the target signal (i.e., the target in Figure 1 ) using the Savitzky-Golay smoothing method to eliminate the noise therein while retaining as much important information (such as inflection points) as possible. The signals of different shots and different lengths are divided into sub-data of the same length by the method of sliding window partitioning, and all the partitioned equal-length sub-data are stacked into an array for saving the dataset for machine learning training. After the model is trained, the input signal is directly passed to the time series Transformer model (i.e., a data-driven model based on the fusion of recurrent neural network and self-attention mechanism) for model inference, and then the missing electron temperature diagnosis that needs to be reconstructed can be obtained.

[0035] The method for reconstructing the Tokamak temperature diagnosis based on a data-driven model of the present invention realizes the reconstruction of the missing plasma electron temperature diagnosis in EAST by training a data-driven model (or machine learning model) based on the fusion of recurrent neural network and self-attention mechanism proposed below, and specifically includes the following steps:

[0036] Step 1, acquiring original data: Acquire the original actuator experimental data and actual diagnostic experimental data of EAST;

[0037] Step 2, data selection and processing: Select approximately 2000 (such as 1873 times, or other values) complete real discharge experimental data that can be used for machine learning training from the above experimental data of EAST as the training data of the machine learning model. Specifically, obtain the electron density signal from HCN, obtain the plasma current, toroidal magnetic field, loop voltage, and heating ratio signals from magnetic diagnostics, obtain the magnetic axis safety factor, 95% flux surface safety factor, plasma stored energy, normalized pressure ratio (normalized β), toroidal pressure ratio, poloidal pressure ratio, elongation ratio, internal inductance, top triple angle, and bottom triple angle signals from offline EFIT, and then obtain the original electron temperature data from HRS and obtain the electron temperature diagnostic data through numerical analysis. Then, these data are further processed into an array format suitable for efficient machine learning;

[0038] Step 3, performing machine learning model training: Input the signals after the above data processing into the above machine learning model. The machine learning model gives a reconstruction output, and then construct a loss function for the machine learning reconstruction output and the actual measurement output of EAST and calculate the loss. Then, train the machine learning model by the error backpropagation method;

[0039] Step 4, Inference performance evaluation of the machine learning model: Use the data obtained from the data processing in Step 2 as the model input, input it into the machine learning model trained in Step 3, and use this machine learning model for inference to directly reconstruct the electron temperature diagnosis. Respectively change the number of layers, batch size of this machine learning model, and whether to use a bidirectional RNN network to construct models for inference verification, analyze the losses of its validation set and test set to evaluate the performance of the machine learning model with this hyperparameter configuration, so as to obtain the machine learning model with the optimal hyperparameters;

[0040] Step 5, Coupling of the machine learning model and the EAST data management system: Use the machine learning model obtained in Step 4 to directly reconstruct the missing electron temperature diagnosis, and use the reconstruction result as the missing electron temperature reference and input it into the EAST data management system.

[0041] For the data selection and processing in the above Step 2, the present invention adopts an efficient data processing strategy to efficiently process the data for subsequent machine learning modeling. The data sampling rate on EAST is usually above 1 kHz, which results in the data time series length of a few seconds of discharge experiments reaching several thousand time steps, and the discharge duration of each shot is usually different, which results in the EAST data showing a long time series type with different lengths. This type of data usually cannot be directly used as the input of the machine learning model. The data processing strategy of the present invention can efficiently process these data for subsequent machine learning modeling. Step 2 specifically includes:

[0042] Step 2.1, Read the discharge experiment data from the EAST data management system and save the data in a hierarchical high-performance data format (such as HDF5 format) for storing and managing large-scale complex data, so as to facilitate subsequent data processing;

[0043] Step 2.2, Read the saved HDF5 format data, and calculate the standard deviation and variance of the discharge experiment data (training data of the machine learning model) of each channel respectively, so as to facilitate subsequent data normalization processing;

[0044] Step 2.3, Divide the discharge experiment data window of each channel, and split it into equal-length sub-data; After basic data processing of these data, pack and save them in the form of an array;

[0045] Step 2.4, Normalize the discharge experiment signals of each channel respectively to ensure that the signals of each channel are at the same magnitude for the convenience of machine learning model training; And perform Savitzky-Golay smoothing processing on the target signal (electron temperature diagnosis) with relatively large noise to eliminate noise on the premise of retaining the signal information volume, and then obtain the target of the machine learning model;

[0046] Step 2.5: After completing the above data processing flow, stack and save the processed data in an array format that can be directly read and used as the input of the machine learning model.

[0047] Specifically, the present invention includes: First, select 1873 shots from the EAST experiment as the entire data set of the experiment. 80% of the data is assigned to the training set, and 20% of the data is assigned to the test set and the validation set, with the validation set and the test set each accounting for 10%. In the electron temperature diagnostic reconstruction system, a total of 26 signals including basic diagnostic signals and actuator signals from EAST are selected as the input of the model. Among them, 14 basic diagnostic signals and 12 actuator signals are adopted. The heating ratio of neutral beam injection (NBI) and electron cyclotron resonance heating (ECRH) is calculated through these 12 actuator signals, and this heating ratio is used as the input of the machine learning model, and the total input dimension is 15.

[0048] Furthermore, the machine learning model can efficiently extract multi-scale features of the sequence and calculate quickly.

[0049] Furthermore, the training of the machine learning model is based on specialized data selection. The most important 14 diagnostic signals and 12 actuator signals are used as the input to directly reconstruct the missing electron temperature diagnosis.

[0050] Furthermore, the efficient data processing strategy reasonably processes the input and output data and encapsulates them into an array format that can be directly and efficiently read and used as the input of the machine learning model.

[0051] Furthermore, the machine learning model is trained with a large amount of EAST experimental data so that it can accurately reconstruct the missing electron temperature diagnosis during the entire discharge process.

[0052] The machine learning model of the present invention (a data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism) can provide accurate and reliable reconstruction results for the missing electron temperature diagnosis of EAST nuclear fusion. This model can also be seamlessly integrated with the EAST data management system and can provide accurate and reliable electron temperature diagnosis references for EAST diagnostic analysis.

[0053] The data-driven model based on the fusion of recurrent neural network and self-attention mechanism of the present invention realizes efficient modeling of long time series data through a hierarchical hybrid architecture, and its structural design essentially solves the inherent defects of traditional single models in long sequence processing. The front end of the model uses recurrent neural network units to recursively encode the original time series, and layer by layer extracts local temporal patterns through gating mechanisms (such as forget gates and input gates). This process is essentially to compress and abstract the features of the original sequence in the time dimension to form a hidden state sequence with semantic representation. The time series Transformer encoder (TSTEncoder) introduced at the back end constructs global perception ability on this basis. The core multi-head self-attention mechanism explicitly establishes time node dependencies across any distance by parallel computing the correlation matrix between different time steps. This two-stage processing architecture realizes the organic fusion of recurrent neural network and self-attention mechanism, making the model have both local sensitivity and global relevance.

[0054] In view of the characteristics of the EAST long time series experimental data, the model has been optimized in multiple aspects at the structural level: First, the input dimension of the subsequent attention module is reduced through the sequence compression mechanism of the recurrent neural network, and the computational complexity corresponding to the original sequence length (n) is reduced from O(n 2 ) of the pure Transformer to O(nd) (where d is the hidden layer dimension of the recurrent neural network), effectively alleviating the memory bottleneck of the ultra-long sequence of tokamak experimental data; Second, the model uses residual connection and layer normalization techniques to stack multiple layers of TSTEncoder, gradually enhancing the ability to extract deep temporal features while avoiding gradient disappearance; In addition, the introduction of a bidirectional recurrent neural network structure enables the model to capture both causal dependencies in forward propagation and potential pattern associations in backward propagation at the same time. This two-way information flow is particularly suitable for time series with periodic oscillations or delay effects. The model also decouples the feature space into multiple subspaces through the multi-head attention mechanism, enabling different attention heads to independently focus on dynamic patterns with different change rates in the sequence (such as fast-fluctuating noise components and slowly evolving trend components). This structural feature significantly improves the model's ability to distinguish multi-scale time patterns in the long time series of EAST experimental data.

[0055] Specifically, the machine learning model of the present invention is implemented as follows:

[0056] The inputs of the machine learning model include 14 plasma state signals such as plasma current, electron density, toroidal magnetic field, loop voltage, safety factor of the magnetic axis, safety factor of the 95% flux surface, plasma stored energy, normalized beta, toroidal beta, poloidal beta, elongation ratio, internal inductance, top triple angle, and bottom triple angle, and 12 actuator signals including 8-channel neutral beam injection (NBI) system signals and 4-channel electron cyclotron resonance heating (ECRH) system signals. The heating ratio of NBI and ECRH is calculated from the above 12 actuator signals. Using the above 15 input signals (14 plasma state signals and 1 heating ratio signal), the evolution of the missing electron temperature diagnostic signal during the entire discharge process can be quickly and accurately reconstructed.

[0057] Example:

[0058] The test result graphs of the present invention on EAST for shot numbers #91475, #93417, #91769, and #83820 are as Figure 2 shown. The toroidal magnetic fields of the four discharge experiments are 1.58 T, 2.41 T, 2.19 T, and 2.43 T respectively, Figure 2 which respectively show the electron temperature evolution reconstruction results of the four entire discharge experiments. Among them, Figure 2 This is an example graph of the electron temperature diagnostic evolution reconstruction results of the present invention at different toroidal magnetic fields. Among them, Figure 2 Figure (a) shows the example graph of the electron temperature diagnostic evolution reconstruction result of the model of the present invention at a toroidal magnetic field of 1.58 T, Figure 2 Figure (b) shows the example graph of the electron temperature diagnostic evolution reconstruction result of the model of the present invention at a toroidal magnetic field of 2.41 T, Figure 2 Figure (c) shows the example graph of the electron temperature diagnostic evolution reconstruction result of the model of the present invention at a toroidal magnetic field of 2.19 T, Figure 2 Figure (d) shows the schematic diagram of the electron temperature diagnostic evolution reconstruction result of the model of the present invention at a toroidal magnetic field of 2.43 T. Obtain 14 plasma state signals and 8 actuator signals of the actual EAST, and perform data preprocessing on all signals through an efficient data processing method. Load the already learned machine learning model and input data into the video memory, directly infer and calculate the electron temperature diagnostic evolution of the missing entire discharge experiment by using the machine learning model, then save the model reconstruction result and perform data visualization.

[0059] Figure 3 This is the residual result graph of the test set experimental data of 374 shots in the discharge experiment of the machine learning model of the present invention in the shot numbers within #82877 - 99092 of the tokamak experiment.

[0060] As Figure 4 and Figure 5As shown, in the Tokamak device of "Experimental Advanced Superconducting Tokamak (EAST)" in Hefei, the data within the Tokamak experiment gun number #82877 - 99092 and the experimental data of 374 guns with longer durations were used as the test set to test the performance of the model on the overall test set. The present invention will evaluate the accuracy of the model from the spatial and temporal perspectives using different model evaluation metrics based on the matrix norm and the coverage rate of the entire gun experiment under two standard deviations. Through the accuracy evaluation method based on the infinite norm of the relative error matrix ( Figure 4 ), the present invention takes the electron temperature diagnosis of the entire discharge experiment as a whole and evaluates the similarity between the reconstruction result and the true value through the infinite norm similarity, effectively representing its prediction accuracy on the entire discharge experiment. Taking the electron temperature diagnosis of the entire discharge experiment as a whole, the stability of the reconstruction result in the entire discharge experiment is evaluated by calculating the coverage rate of the predicted value and the true value under two standard deviations ( Figure 5 ). As shown in Figure 4 and Figure 5 , the average similarity of the test set based on the infinite norm in the entire discharge experiment reached 86.5%, and the coverage rate reached 95.9%.

[0061] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memory, CD-ROM, optical memory, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented in various computer languages. For example, object-oriented programming languages such as Java and interpreted scripting languages such as JavaScript.

[0062] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0063] These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means that implement the function specified in one or more of the procedures Figure 1 or procedures and / or blocks Figure 1 or blocks or more blocks.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in one or more of the procedures Figure 1 or procedures and / or blocks Figure 1 or blocks or more blocks.

[0065] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made by those skilled in the art once they learn the basic creative concept. Therefore, the appended claims are intended to be construed to cover the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.

[0066] It is obvious that those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. A Tokamak temperature diagnosis reconstruction method based on a data-driven model, characterized in that By training a data-driven model that combines a recurrent neural network and a self-attention mechanism, the plasma electron temperature diagnosis in EAST is reconstructed, specifically including: Step 1, obtaining original data: Obtain the original actuator experimental data and the actually diagnosed experimental data of EAST; Step 2, data selection and processing: Select several complete and real discharge experimental data suitable for machine learning training from the EAST experimental data as the training and inference data of the machine learning model; Step 3, perform machine learning model training: Input the processed signal into the machine learning model, and the machine learning model gives a reconstructed output. Construct a loss function for the machine learning reconstructed output and the actual measurement output of EAST and calculate the loss. Then, train the machine learning model through the error backpropagation method. The trained data-driven model that combines a recurrent neural network and a self-attention mechanism can directly reconstruct the missing electron temperature diagnosis; Step 4, evaluate the inference performance of the machine learning model: Use the data obtained from the data processing in Step 2 as the model input, input it into the data-driven model that combines a recurrent neural network and a self-attention mechanism trained in Step 3, and use this model to directly reconstruct the electron temperature diagnosis through inference. Respectively change the number of layers, batch size of the model, and whether to use a bidirectional RNN network to construct models for inference verification, and analyze the losses of its validation set and test set to evaluate the performance of the machine learning model with this hyperparameter configuration, so as to obtain the machine learning model with the optimal hyperparameters; Step 5, couple the machine learning model and the EAST data management system: Use the machine learning model obtained in Step 4 to directly reconstruct the missing electron temperature diagnosis and input the reconstructed result as the missing electron temperature reference into the EAST data management system.

2. The Tokamak temperature diagnosis reconstruction method based on a data-driven model according to claim 1, characterized in that Step 2 includes: Obtain the electron density signal from HCN, obtain the plasma current, toroidal magnetic field, loop voltage, and heating ratio signals from the magnetic diagnosis, obtain the magnetic axis safety factor, 95% flux surface safety factor, plasma stored energy, normalized specific pressure, toroidal specific pressure, poloidal specific pressure, elongation ratio, internal inductance, top triple angle, and bottom triple angle signals from the offline EFIT. Then, obtain the original electron temperature data from HRS and obtain the electron temperature diagnosis data through numerical analysis. After that, further process these data into an array format suitable for efficient machine learning.

3. The Tokamak temperature diagnosis reconstruction method based on the data-driven model according to claim 2, characterized in that Step 2 includes: Step 2.1, read the discharge experimental data from the EAST data management system and save the data as a hierarchical high-performance data format for storing and managing large-scale complex data to facilitate subsequent data processing; Step 2.2, read the saved data and calculate the standard deviation and variance of the discharge experimental data for each channel respectively to facilitate subsequent data normalization processing; Step 2.3, divide the discharge experimental data window for each channel and split it into equally long sub-data; Step 2.4: Standardize the discharge experiment signals of each channel to ensure that the signals of each channel are on the same order of magnitude for the convenience of machine learning model training; and perform Savitzky-Golay smoothing on the target signals with relatively large noise to eliminate noise while retaining the signal information content, thereby obtaining the target of the machine learning model. Step 2.5: After completing the above data processing flow, stack the processed data and save it in an array format that can be directly read and used as the input of the machine learning model.

4. The method for diagnosing and reconstructing the Tokamak temperature based on the data-driven model according to claim 1, characterized in that, In Step 2, several shots are selected from the EAST experiment as the entire data set of the experiment. Among them, 80% of the data is assigned to the training set, 20% of the data is assigned to the test set and the validation set, and the validation set and the test set each account for 10%; a total of 26 signals including 14 basic diagnostic signals from EAST and 12 actuator signals are selected as the input of the model. The heating ratio of neutral beam injection and electron cyclotron resonance heating is calculated through these 12 actuator signals, and the heating ratio and 14 basic diagnostic signals are used as the input of the machine learning model.

5. The Tokamak temperature diagnosis reconstruction method based on a data-driven model according to claim 4, characterized in that, The 14 basic diagnostic signals are 14 plasma state signals, including plasma current, electron density, toroidal magnetic field, loop voltage, safety factor of the magnetic axis, safety factor of the 95% flux surface, plasma stored energy, normalized beta, toroidal beta, poloidal beta, elongation ratio, internal inductance, top triple angle, and bottom triple angle.

6. The Tokamak temperature diagnosis reconstruction method based on a data-driven model according to claim 1, characterized in that The data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism realizes efficient modeling of long-time series data through a hierarchical hybrid architecture. The front end uses a recurrent neural network unit to recursively encode the original time series, and extracts local time series patterns layer by layer through a gating mechanism. The time series Transformer encoder introduced at the back end constructs global perception ability on this basis.

7. The method for diagnosing and reconstructing the Tokamak temperature based on the data-driven model according to claim 3, wherein In Step 2.3, after dividing the discharge experiment data window of each channel and splitting it into equally long sub-data, these data are subjected to basic data processing and then packaged and saved in an array form.

8. A Tokamak temperature diagnosis and reconstruction device based on a data-driven model, characterized in that, By training a data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism, the plasma electron temperature diagnosis in EAST is reconstructed, specifically including: Original data acquisition module: Acquire the original actuator experiment data and actual diagnostic experiment data of EAST. Data selection and processing module: Select several complete and real discharge experiment data suitable for machine learning training from the EAST experiment data as the training and inference data of the machine learning model. Machine learning model training module: Input the processed signals into the machine learning model. The machine learning model gives a reconstruction output, constructs a loss function for the machine learning reconstruction output and the actual measurement output of EAST and calculates the loss, and then trains the machine learning model through the error backpropagation method. The trained data-driven model based on the fusion of a recurrent neural network and a self-attention mechanism can directly reconstruct the missing electron temperature diagnosis. Machine learning model inference performance evaluation module: Use the data obtained from the data processing in Step 2 as the model input, input it into the data-driven model trained in Step 3 based on the fusion of recurrent neural network and self-attention mechanism, and use this model for inference to directly reconstruct the electron temperature diagnosis. Respectively change the number of layers, batch size of the model, and whether to use a bidirectional RNN network to construct models for inference verification, and analyze the losses of its validation set and test set to evaluate the performance of the machine learning model with this hyperparameter configuration, so as to obtain the machine learning model with the optimal hyperparameters; Coupling module of machine learning model and EAST data management system: Use the machine learning model obtained in Step 4 to directly reconstruct the missing electron temperature diagnosis, and input the reconstruction result as the missing electron temperature reference into the EAST data management system.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the Tokamak temperature diagnosis reconstruction method based on the data-driven model according to any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the Tokamak temperature diagnosis reconstruction method based on the data-driven model according to any one of claims 1 to 7.

Images