Cross-device tokamak plasma disruption prediction method based on domain generalization
By constructing a joint discharge dataset from multiple tokamaks and using adversarial learning to train the domain discriminator and classifier, the problem of poor transferability of tokamak plasma fracture prediction algorithms across different devices was solved, achieving accurate fracture prediction on future large-scale commercial fusion reactors and reducing device damage.
Patent Information
- Application Number
- CN202310090150.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-02-09
- Publication Date
- 2025-11-21
- Estimated Expiration
- 2043-02-09
AI Technical Summary
Existing tokamak plasma breakup prediction algorithms have poor portability across different devices, and future large-scale commercial fusion reactors lack plasma data under high-parameter discharge conditions, making it difficult to train reliable breakup prediction methods.
A joint discharge dataset from multiple tokamaks was constructed. A domain-generalized deep learning method was adopted. Through feature extractor, domain discriminator and classifier, adversarial learning was used to train the domain discriminator to be unable to distinguish the device source, and the classifier accurately predicted the rupture precursor. A domain-generalized cross-device rupture prediction model was designed.
It improves the transferability of fracture prediction, enabling the model to be applied to future large-scale commercial fusion reactors, achieving accurate fracture prediction across different tokamak devices and reducing device damage.
Smart Images
Figure CN116304859B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of magnetic confinement nuclear fusion, and more particularly relates to a cross-device tokamak plasma disruption prediction method based on domain generalization. BACKGROUND
[0002] A tokamak is a toroidal vessel that uses magnetic confinement to achieve controlled nuclear fusion. Plasma major disruption is one of the events that poses the greatest threat to the safe and long-term stable operation of a tokamak device. Disruption will directly cause the discharge of the device to stop, and will produce heat load, electromagnetic load and escaped electrons, causing serious damage to the structure of the device itself, thereby causing great economic loss and safety hazard. Disruption prediction and mitigation is a key link in the research of magnetic confinement nuclear fusion. The disruption mitigation system can dissipate the energy confined in the tokamak in a safe manner within a certain period of time before the disruption occurs, thereby minimizing the damage to the device due to the possible disruption. The disruption mitigation system needs a disruption prediction algorithm to provide an early warning so as to trigger it. Future large commercial fusion reactors cannot withstand plasma major disruption under high parameter discharge, and therefore a safe and reliable disruption prediction method is needed to predict possible disruptions as much as possible.
[0003] Existing tokamak plasma disruption prediction algorithms mainly include traditional machine learning methods based on physical principles or deep learning methods. The traditional machine learning method based on physical principles requires a deep understanding of the physical principles, and feature engineering is performed manually based on the understanding of the principles. Since there are differences in diagnostic systems, device parameters, and configurations between different tokamak devices, this method needs to be tailored for different devices, and has poor transferability, making it difficult to extrapolate existing disruption prediction algorithms to other devices. The deep learning method requires a large amount of data for training, but since future large commercial fusion reactors cannot withstand plasma major disruption under high parameter discharge, sufficient data cannot be provided to train a reliable disruption prediction method. SUMMARY
[0004] In view of the defects of the prior art and the need for improvement, the present application provides a cross-device tokamak plasma disruption prediction method based on domain generalization, which aims to improve the transferability of disruption prediction and enable its application to large commercial fusion reactors.
[0005] To achieve the above-mentioned purpose, according to one aspect of the present application, a cross-device tokamak plasma disruption prediction method based on domain generalization is provided, comprising:
[0006] Training phase:
[0007] Constructing a joint discharge data set of multiple tokamaks;
[0008] training a domain generalization based disruption prediction model with the dataset as training samples, and with a label of a sample source from which the sample is derived and whether the sample represents a disruption precursor or not;
[0009] The domain generalization based disruption prediction model comprises a feature extractor, a domain discriminator and a classifier; a feature vector output by the feature extractor is an input of the domain discriminator and the classifier.
[0010] In the training process, the domain discriminator is trained by adopting adversarial learning to maximize loss, so that the domain discriminator cannot distinguish whether the feature vector is from the same tokamak device or not; the classifier is trained to minimize classification loss, so that the classifier predicts information of whether the feature vector has a disruption precursor or not.
[0011] Application stage:
[0012] Discharge data of a target tokamak device is input into the trained feature extractor, and a trained classifier outputs information of whether there is a disruption precursor or not.
[0013] Further, the discharge data set of the plurality of tokamaks comprises diagnostic system data and control system signals related to plasma discharge parameters, structural parameters of each tokamak device and diagnostic system field of view arrangement information.
[0014] Further, it further comprises selecting the diagnostic system data for normalization based on the requirements of the diagnostic system and the disruption prediction task required on the target tokamak device.
[0015] Further, for diagnostic system data of the same physical quantity measured by different tokamaks, normalization processing is performed according to the measurement principle, combined with the diagnostic field of view and the number of diagnostic arrays.
[0016] Further, for diagnostic system data of physical quantities related to parameters of the tokamak device itself, normalization is performed by statistics.
[0017] Further, the domain discriminator comprises one of a generative adversarial network or a domain adversarial neural network.
[0018] Further, the domain discriminator is a domain adversarial neural network, and the domain adversarial neural network comprises a gradient reversal layer and a plurality of fully connected layers connected in series.
[0019] Further, the classifier comprises a plurality of fully connected layers.
[0020] Further, in the application stage, further comprising parameter adjustment of the trained classifier by using the labeled data on the target tokamak device, to obtain an optimized classifier;
[0021] Discharge data of the target tokamak device is input into the trained feature extractor, and the optimized classifier outputs information about whether there is a precursor of a disruption.
[0022] According to a second aspect of the present application, a cross-device tokamak plasma disruption prediction system based on domain generalization is provided, comprising a computer readable storage medium and a processor;
[0023] The computer readable storage medium is used to store executable instructions;
[0024] The processor is used to read the executable instructions stored in the computer readable storage medium, and execute the prediction method according to any one of the first aspect.
[0025] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:
[0026] (1) The cross-device tokamak plasma disruption prediction method based on domain generalization of the present application inputs the discharge data of multiple tokamaks into the feature extractor, and the inputs of the domain discriminator and the classifier come from the feature extractor, but the target of the domain discriminator is to maximize the domain classification loss, and the domain discriminator is trained by using adversarial learning to confuse multiple domain data, so that the domain discriminator cannot distinguish the specific source of the data, and in the parameter reverse iteration, the feature vector extracted by the feature extractor is domain-independent, while the target of the classifier is to minimize the classification loss to achieve accurate classification, so that the classifier can accurately judge whether the feature vector has a disruption precursor, and in the parameter reverse iteration, the feature extracted by the feature extractor is related to the disruption precursor, and finally when the loss converges or reaches the predetermined training round, the feature extracted by the feature extractor is both domain-independent and contains the disruption precursor feature, which can be extrapolated to other tokamak devices for disruption prediction, thereby improving the transferability.
[0027] Meanwhile, the present application uses historical data of multiple tokamak devices to construct a training data set, rather than directly using data of the target device, which can be applied to future large commercial fusion reactors, so that the method of the present application improves the transferability while solving the problem that future large commercial fusion reactors cannot provide sufficient data.
[0028] (2)The designed multi-tokamak joint discharge data set includes not only the diagnostic system data related to the plasma discharge parameters and the control system signals, but also the structural parameters of each tokamak device and the diagnostic system field layout information, and the difference information of different tokamak devices is introduced, so as to train the field generalization-based disruption prediction model, so that the feature extractor extracts the bottom layer features irrelevant to the field.
[0029] (3) Further, in order to avoid the great difference of the diagnostic system, structural parameters and position of different tokamak devices, in the construction of the database of the application, the diagnostic system data is normalized; the different diagnostic system data of the same physical quantity measured on different devices is normalized according to the measurement principle and diagnostic field, and the diagnostic system data of the physical quantity measured related to the device itself is normalized in a statistical manner.
[0030] In summary, the method of the application can train a field-independent magnetic confinement fusion feature extractor using the existing tokamak device disruption data, and can fine-tune and extrapolate the pre-trained feature extractor to the target tokamak through transfer learning, and can be applied to future large commercial fusion reactors. BRIEF DESCRIPTION OF DRAWINGS
[0031] Figure 1 The application provides a cross-device plasma disruption prediction method based on field generalization.
[0032] Figure 2 The application provides a cross-device plasma disruption prediction method based on field generalization. DETAILED DESCRIPTION
[0033] In order to make the purpose, technical scheme and advantages of the application clearer, the application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.
[0034] As shown in the drawings, the application provides a cross-device tokamak plasma disruption prediction method based on field generalization, which mainly includes: Figure 1
[0035] (1) constructing a multi-tokamak joint discharge data set;
[0036] Specifically, historical experimental data of at least two and more tokamak devices are acquired, the historical experimental data including: diagnostic system data and control system signals related to key parameters of plasma discharge, structural parameters of each tokamak device, and diagnostic system field of view arrangement information. Wherein, by introducing the structural parameters and diagnostic system field of view arrangement information of each tokamak device, the difference information of different tokamak devices can be introduced. On this basis, the historical experimental data is labeled and cleaned to obtain a large amount of labeled data on multiple devices, and the label is divided into two parts, one part is which device the data comes from, and the other part is whether the data represents a disruption precursor.
[0037] (2) Normalizing the diagnostic system data of multiple tokamak devices;
[0038] Specifically, considering that the diagnostic systems, structural parameters and shape of different tokamak devices are quite different, the obtained diagnostic system data needs to be normalized. The diagnostic system data is normalized based on the requirements of the diagnostic signals of the target tokamak device and the disruption prediction task. For different tokamak devices measuring the same physical quantity, different devices have different diagnostic systems. For such diagnostic system data, according to the measurement principle, combined with the diagnostic field of view and the number of diagnostic arrays, the normalization processing is carried out; for the diagnostic system data of the physical quantity related to the parameters of the device itself, the normalization processing is carried out respectively according to the statistics.
[0039] (3) Design and train a deep learning disruption prediction model based on domain generalization;
[0040] The disruption prediction model includes: a transferable feature extractor, a domain discriminator and a classifier;
[0041] The transferable feature extractor is used to extract bottom layer features irrelevant to each tokamak device with the constructed data set as input, and obtain a feature vector;
[0042] The domain discriminator is used to predict whether the sample comes from the same device. Specifically, the domain discriminator takes the feature vector as input, takes which device the sample comes from as label, and maximizes the loss as target, and is trained by using adversarial learning, so that the domain discriminator cannot distinguish whether the feature vector from the feature extractor comes from the same device, and when the parameters are iteratively reversed, the feature vector extracted by the feature extractor is irrelevant to the domain;
[0043] The classifier is used to predict whether the sample is a disruption, specifically, the feature vector is input, the sample is labeled whether it represents a disruption precursor, and the classifier is trained to minimize the classification loss, so that the classifier can accurately determine whether the feature vector has a disruption precursor, and when the parameters are iterated in reverse, the parameters of the feature extractor are affected, so that the feature vector output by the feature extractor contains information about the disruption precursor.
[0044] As preferred, the domain discriminator in the application comprises one of a generative adversarial network (GAN) or a domain-adversarial neural network (DANN).
[0045] Specifically, in this embodiment, the domain discriminator adopts a domain-adversarial neural network (DANN) and comprises a gradient reversal layer (GRL) and a plurality of fully connected layers connected in series. The gradient reversal layer is used to reverse the parameters in the process of back propagation, and the plurality of fully connected layers are used to distinguish whether the sample data comes from which domain, i.e., which tokamak device. When updating the gradient, the gradient direction is reversed and fed back to the feature extractor, so that the features extracted by the feature extractor are domain-independent.
[0046] The classifier comprises a plurality of fully connected layers, which are used to predict whether the current sample represents a disruption precursor.
[0047] The domain discriminator and the classifier are spliced after the feature extractor, i.e., the feature vector output by the feature extractor is the input of the domain discriminator and the classifier. The domain discriminator determines whether the current sample comes from different domains, i.e., whether it comes from different tokamak devices, and the prediction target is that the domain discriminator cannot distinguish the domain from which the sample comes as much as possible; the classifier determines whether the current sample represents a disruption precursor, and the prediction target is to accurately determine whether the sample represents a disruption precursor as much as possible. The gradient reversal layer (GRL) in the domain discriminator reverses the neural network parameters in the process of back propagation, so that when the parameters are iterated, the update direction is opposite to that of the classifier, i.e., during training, the domain discriminator cannot distinguish the domain from which the sample comes as much as possible, while the classifier accurately determines whether the sample represents a disruption precursor as much as possible. Finally, when the loss converges, a feature extractor that is domain-independent and contains a disruption precursor is obtained.
[0048] In this embodiment, Adam is used as the optimizer to optimize the parameters of the feature extractor, domain discriminator and classifier. The domain discriminator uses CategoricalCrossEntropy as the loss function, and the classifier uses BinaryCrossEntropy as the loss function to train the above prediction targets. It should be noted that the loss of the above domain discriminator is reversed, so the training target is to make the accuracy of the domain discriminator as low as possible, and the accuracy of the classifier as high as possible. After several iterations or loss convergence, a set of optimal neural network parameters is obtained as the final determined neural network parameters.
[0049] (4) Application stage
[0050] The target domain data is processed into the same data format as the training data set and input into the trained feature extractor. The trained classifier outputs information on whether it has a cracking precursor.
[0051] The trained feature extractor and classifier are saved locally. The target tokamak device discharge data is processed into the same data format as the training set and input into the model for feedforward calculation to obtain the result. The domain discriminator is not required in the feedforward calculation process. After the target domain data set enters the feature extractor to obtain the feature vector, it only enters the classifier and does not need to enter the domain discriminator.
[0052] When there is a small amount of labeled data on the target tokamak device, the transfer learning method can also be used. The parameters of the feature extractor are frozen, and the labeled data on the target device are used to fine-tune the parameters of the classifier to obtain an optimized classifier to better adapt to the data distribution on the target tokamak device. The data on the target tokamak device is processed into the same data format as the training data set and input into the trained feature extractor. The optimized classifier outputs early warning information on whether it has a cracking precursor. The early warning information is connected to the triggering of the cracking mitigation system to trigger the cracking mitigation system and take appropriate cracking mitigation measures to reduce or avoid the harm caused by the plasma major disruption, such as Figure 2 as shown.
[0053] The domain generalization based cross-device tokamak plasma disruption prediction method of the application, by constructing a plurality of tokamak joint discharge data set input into the feature extractor, the domain discriminator takes the feature vector output by the feature extractor as the input, takes which device the sample comes from as the label, takes maximizing the loss of the domain discriminator as the goal, uses the adversarial learning to train the domain discriminator, so that the domain discriminator cannot distinguish which device the data comes from, when the parameter is updated reversely, let the domain discriminator and the feature extractor update the parameters in the direction of being more indistinguishable, finally make the features extracted by the feature extractor irrelevant to the domain; the classifier takes the feature vector output by the feature extractor as the input, takes whether the sample represents the disruption precursor as the label, takes minimizing the classification loss as the goal to train the classifier, when the parameter is back propagated, make the performance of the classifier as good as possible, at the same time, influence the parameters of the feature extractor, make the features extracted by the feature extractor related to the disruption precursor, finally the features extracted by the feature extractor are not only domain independent, but also contain the disruption precursor features, which is conducive to extrapolation or migration to the target device to realize the disruption prediction task.
[0054] And, the application uses the historical data of multiple tokamak devices to construct the training data set, rather than directly using the data of the target device, so that the method of the application can be applied to future large commercial fusion reactors.
[0055] The embodiment of the application also provides a domain generalization based cross-device tokamak plasma disruption prediction system, comprising: a computer readable storage medium and a processor.
[0056] The computer readable storage medium is used for storing executable instructions.
[0057] The processor is used for reading the executable instructions stored in the computer readable storage medium, and executing the domain generalization based cross-device tokamak plasma disruption prediction method as described in the above embodiment.
[0058] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the application, and is not used to limit the application, any modification, equivalent replacement and improvement made within the spirit and principle of the application should be included in the protection scope of the application.
Claims
1. A cross-device tokamak plasma breakage prediction method based on domain generalization, characterized in that, include: Training phase: Construct a joint discharge dataset from multiple tokamaks; Using the dataset as training samples, the domain-generalized rupture prediction model is trained by labeling which tokamak device the sample comes from and whether the sample represents a rupture precursor. The domain-generalization-based rupture prediction model includes a feature extractor, a domain discriminator, and a classifier; the feature vector output by the feature extractor is the input to the domain discriminator and the classifier. During training, adversarial learning is used to train the domain discriminator with the goal of maximizing the loss, making the domain discriminator unable to distinguish whether the feature vectors come from the same tokamak device; the classifier is trained with the goal of minimizing the classification loss, making the classifier predict whether the feature vectors have information about the precursors of rupture. Application phase: The discharge data of the target tokamak device is input into a trained feature extractor, and the trained classifier outputs information on whether there are signs of impending rupture.
2. The prediction method according to claim 1, characterized in that, The combined discharge dataset of multiple tokamaks includes: diagnostic system data and control system signals related to plasma discharge parameters, structural parameters of each tokamak device, and diagnostic system field-of-view layout information.
3. The prediction method according to claim 2, characterized in that, It also includes selecting diagnostic system data for normalization based on the diagnostic signal requirements of the diagnostic system and rupture prediction task on the target tokamak device.
4. The prediction method according to claim 3, characterized in that, For diagnostic system data measuring the same physical quantity using different tokamak devices, normalization is performed based on the measurement principle and the diagnostic field of view and the number of channels in the diagnostic array.
5. The prediction method according to claim 3, characterized in that, For diagnostic system data on physical quantities related to the parameters of the tokamak device itself, statistical normalization is performed.
6. The prediction method according to claim 1, characterized in that, The domain discriminator includes either a generative adversarial network or a domain adversarial neural network.
7. The prediction method according to claim 6, characterized in that, The domain discriminator is a domain adversarial neural network, which includes a gradient inversion layer and multiple fully connected layers connected in series.
8. The prediction method according to claim 1, characterized in that, The classifier comprises multiple fully connected layers.
9. The prediction method according to claim 1, characterized in that, In the application phase, it also includes using labeled data from the target tokamak device to adjust the parameters of the trained classifier and obtain an optimized classifier. The discharge data of the target tokamak device is input into a trained feature extractor, and the optimized classifier outputs information on whether there are signs of impending rupture.
10. A cross-device tokamak plasma fracture prediction system based on domain generalization, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the prediction method as described in any one of claims 1-9.