Survival analysis-based fusion plasma disruption time prediction method and system

Abstract

The application discloses a fusion plasma disruption time prediction method and system based on survival analysis, and relates to the technical field of nuclear fusion energy, and comprises the following steps: obtaining historical plasma discharge data of a tokamak device, preprocessing, and then adopting a sliding window to construct a time sequence sample, wherein the time sequence sample comprises a disruption time sequence sample and a non-disruption time sequence sample; constructing a training label for each time sequence sample of the sliding window, wherein the training label comprises an observation time and a disruption label; establishing a disruption time prediction model, taking the time sequence sample and the training label as input to perform model training, and adjusting model parameters through a constructed survival analysis loss function; inputting a real-time collected plasma diagnostic signal into the trained disruption time prediction model, outputting a survival function, and calculating a disruption remaining time; and the method improves the accuracy and practicality of disruption prediction.

Description

Technical Field

[0001] This invention relates to the field of nuclear fusion energy technology, and in particular to a method and system for predicting the fusion plasma breakup time based on survival analysis. Background Technology

[0002] During the operation of the fully superconducting tokamak device (EAST), plasma breakup can lead to the termination of discharge, while generating strong thermal loads, electromagnetic forces, and escape electrons, which can damage the device structure and critical components. Therefore, accurate prediction of plasma breakup is of great significance for ensuring the safe operation of the device.

[0003] Currently, extensive research has been conducted both domestically and internationally on plasma breakup prediction. Existing methods mostly employ machine learning or deep learning models to analyze the discharge state based on plasma diagnostic signals, thereby predicting the risk of breakup. These methods typically output a risk value in the range [0, 1] to represent the probability or severity of plasma breakup at the current moment, thus providing a breakup warning. Once a breakup warning is issued, the plasma control system (PCS) triggers the breakup mitigation system to dissipate the thermal and magnetic energy of a large breakup while preventing the generation of escape electrons. However, the single breakup risk metric output by the breakup prediction only reflects the probability of breakup, completely failing to utilize the crucial information of time to breakup. The lack of time-dimensional warning information makes it impossible to provide refined decision-making basis for the plasma control system, making it difficult to adapt to the graded handling requirements in actual operation, and limiting the practical application effectiveness of breakup warning and prevention.

[0004] In recent years, a few studies have attempted to directly predict the plasma disruption time in order to obtain richer disruption information. However, these methods typically only use disruption discharge data during model training, failing to fully utilize the large amount of non-disruption discharge data available in experiments. The 2001 paper by Th. Zehetbauer et al., "Real-time disruption handling at ASDEX upgrade," directly used MSE as the loss function for training. This resulted in the model being trained solely on disruption discharge data, neglecting non-disruption discharge factors. More critically, using MSE made the model insensitive to samples about to disrupt, which is precisely the focus of the prediction, leading to a false negative rate as high as 15% and a mean squared error of 15% in the disruption survival time.

[0005] In addition, non-rupture discharge data also contains a wealth of information on stable plasma operation. If this data is not effectively utilized during model training, it will result in: insufficient use of experimental data resources, a single distribution of model training samples which may affect the model's generalization ability, and insufficient ability of the model to characterize the stable plasma operation state. Summary of the Invention

[0006] Based on the technical problems existing in the background technology, this invention proposes a method and system for predicting the fusion plasma breakup time based on survival analysis, which improves the accuracy and practicality of breakup prediction.

[0007] The method for predicting the fusion plasma breakup time and identifying the breakup risk proposed in this invention includes: Historical plasma discharge data of a tokamak device are acquired, preprocessed, and then time series samples are constructed using a sliding window. The time series samples include rupture time series samples and non-rupture time series samples. For each sliding window, a training label is constructed for the time series samples, including the observation time and the break label. The break label for the broken time series sample is 1, and the observation time is the time difference from the end of the window to the break time. The break label for the non-broken time series sample is 0, and the observation time is the time difference from the end of the window to the end of the top segment. A rupture time prediction model was established, and the model was trained using time series samples and training labels as inputs. The model parameters were adjusted by the constructed survival analysis loss function. The real-time acquired plasma diagnostic signals are input into the trained rupture time prediction model, which outputs a survival function and calculates the remaining rupture time.

[0008] Furthermore, the survival analysis loss function for: ; in, For broken labels, The total number of training samples, For training sample index, For time The probability density function at which a break occurs. The survival function, i.e., time. The probability that no further breakage will occur. The breakage weight is used to address the imbalance between broken and unbroken samples. When the fracture time prediction model increases the attention weight placed on the fracture cannon; when When the rupture time prediction model is in the standard state; when At that time, the fracture time prediction model increases the attention weight placed on non-fractured guns.

[0009] Furthermore, the probability density function and survival function Modeling using Weibull distribution: ; ; in, To represent the remaining time from the current observation time, For shape parameters, This is the scale parameter.

[0010] Furthermore, the remaining time of fracture is obtained by calculating the limited mean survival time: ; in, To limit average survival time, For the preset time window, For time, as the integration variable, This is the survival function.

[0011] Furthermore, the preprocessing includes: timeline unification, data cleaning, data standardization, and time truncation; The time truncation specifically involves: setting a filtering range for the remaining time of the rupture; for ruptured time series samples, only samples with a remaining time of rupture less than a preset filtering time are selected; for non-ruptured time series samples, samples with a censoring time greater than the preset filtering time are replaced with the preset filtering time for their remaining time.

[0012] Furthermore, the sliding window has a window length of 50ms-200ms, the sliding step size for the ruptured discharge sample is 5ms-20ms, and the sliding step size for the non-ruptured discharge sample is 50ms-200ms.

[0013] Furthermore, the rupture time prediction model includes a feature extraction module and a distribution parameter prediction module; The feature extraction module includes a one-dimensional convolutional layer, a temporal attention mechanism layer, and a fully connected layer connected in sequence. The temporal attention mechanism layer consists of two sequentially connected one-dimensional convolutional layers.

[0014] Furthermore, using survival functions, a hierarchical risk assessment mechanism based on survival probability over a time interval is constructed, specifically as follows: Let the first preset time threshold be The second preset time threshold is ,in ; Calculate plasma in time interval The first risk probability of internal rupture ; Calculate plasma in time interval The second risk probability of internal rupture ; Calculate plasma in time The third risk probability that it will continue to operate stably. ; Constructing risk vectors This yields the component with the highest probability in the risk vector; when When the value is at its maximum, the current plasma is determined to be in a high-risk rupture state, and the system enters the rupture mitigation mode; when When the value is at its maximum, the current plasma is determined to be in a state of manageable risk, and the system enters the rupture avoidance mode; when When the value is at its maximum, it is determined that the current plasma is in a stable operating state, and the system maintains normal control mode.

[0015] The fusion plasma breakup time prediction system based on survival analysis includes a data acquisition module, a label building module, a model building and training module, and a model execution module. The data acquisition module is used to acquire historical plasma discharge data of the tokamak device. After preprocessing, a time series sample is constructed using a sliding window. The time series sample includes rupture time series sample and non-rupture time series sample. The label building module is used to build training labels for each time series sample of the sliding window, including observation time and break label. The break label of the broken time series sample is 1 and the observation time is the time difference from the end of the window to the break time. The break label of the non-broken time series sample is 0 and the observation time is the time difference from the end of the window to the end of the top segment. The model building and training module is used to build a rupture time prediction model. It uses time series samples and training labels as inputs to train the model and adjusts the model parameters through a constructed survival analysis loss function. The model execution module is used to input the real-time acquired plasma diagnostic signals into the trained rupture time prediction model, output the survival function, and calculate the remaining rupture time.

[0016] Furthermore, the survival analysis loss function for: ; in, For broken labels, The total number of training samples, For training sample index, For time The probability density function at which a break occurs. The survival function, i.e., time. The probability that no further breakage will occur. This is the breakage weight.

[0017] Furthermore, in the data acquisition module, the preprocessing includes: timeline unification, missing value interpolation repair, outlier truncation, and data standardization.

[0018] The advantages of the fusion plasma breakup time prediction method and system based on survival analysis provided by this invention are as follows: it solves the problems that most existing breakup prediction methods only output breakup risk values ​​and cannot directly predict the remaining breakup time, a few existing breakup time prediction methods do not make full use of non-breakup discharge data, and traditional MSE regression methods have a large impact on samples far from breakup in the loss function. This method uses a variety of diagnostic signals from the tokamak device to construct time series samples, and predicts the remaining plasma breakup time based on the breakup time prediction model. It also makes full use of breakup discharge data and non-breakup discharge data for model training, thereby improving the accuracy and practicality of breakup prediction. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a diagram showing the remaining time before rupture. Figure 3A The signal curve for gun number 53888; Figure 3B A schematic diagram showing the labeling of fracture time sequence samples under blast number 53888; Figure 4A The signal curve for gun number 87608; Figure 4B A schematic diagram showing the labeling of non-fracture time series samples under blast number 87608; Figure 5A The image shows the prediction results of the time series sample for the rupture of the 0.2s time window under gun number 90202; Figure 5B for Figure 5A A graph showing the variation trend of typical plasma signals; Figure 6A The image shows the prediction results for the non-rupture time series sample within a 0.2s time window for gun number 93648. Figure 6B for Figure 6A A graph showing the variation trend of typical plasma signals; Figure 7A The image shows the predicted performance of the time series sample for the rupture of the 0.3s time window under gun number 90202. Figure 7B for Figure 7A A graph showing the variation trend of typical plasma signals; Figure 8A The image shows the prediction results for a non-rupture time series sample within a 0.3s time window for gun number 93648. Figure 8B for Figure 8A A graph showing the variation trend of typical plasma signals; Figure 9 A chart of the IBS index; Figure 10 AUC indicator chart; Figure 11 This is a chart of the MAE indicator. Detailed Implementation

[0020] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] like Figures 1 to 11 As shown, the fusion plasma breakup time prediction and breakup risk identification method proposed in this invention includes: Step 1: Obtain historical plasma discharge data of the tokamak device, preprocess it, and construct a time series sample using a sliding window. The plasma discharge data includes fracture discharge data and non-fracture discharge data, and the time series sample includes fracture time series sample and non-fracture time series sample. Step 2: Construct training labels for the time series samples of each sliding window, including observation time and break label. The break label for the broken time series sample is 1, and the observation time is the time difference from the end of the window to the break time. The break label for the non-broken time series sample is 0, and the observation time is the time difference from the end of the window to the end of the top segment. Step 3: Establish a rupture time prediction model, train the model using time series samples and training labels as input, and adjust the model parameters through the constructed survival analysis loss function; Step 4: Input the real-time acquired plasma diagnostic signals into the trained rupture time prediction model, output the survival function, and calculate the remaining rupture time.

[0022] This embodiment addresses the problems that most existing fracture prediction methods only output fracture risk values ​​and cannot directly predict the remaining fracture time, as well as the problem that a few existing fracture time prediction methods do not make full use of non-fracture discharge data. This method uses various diagnostic signals from the tokamak device to construct time series samples, and predicts the remaining plasma fracture time based on the fracture time prediction model. It also makes full use of fracture discharge data and non-fracture discharge data for model training, thereby improving the accuracy and practicality of fracture prediction.

[0023] Specifically, this embodiment constructs a rupture time prediction model, enabling the model to directly output the remaining time or survival probability distribution of the plasma before rupture, thus avoiding the problem of manually setting a fixed warning time window in traditional rupture risk prediction methods. When rupture warning information at different time scales is needed, it can be achieved without retraining the model, improving the model's flexibility and applicability. In addition, this embodiment integrates rupture time information, which can make up for the time dimension shortcomings of existing models, providing more comprehensive predictive data for the plasma control system; it enables the control system to formulate rupture avoidance, soft landing, and rupture mitigation strategies in stages based on the remaining time before rupture, grasping the best handling opportunity and providing safety assurance for plasma discharge.

[0024] Meanwhile, this embodiment utilizes both rupture discharge data and non-rupture discharge data during model training, and introduces non-rupture discharge data as censored samples into the model training. This allows for more full utilization of experimental data, improves the model's ability to characterize the plasma operating state, and enhances the model's generalization performance and prediction stability. Censored samples refer to samples where the event has not yet occurred at the end of the observation period. In the plasma rupture scenario, this means that the shot ultimately did not rupture.

[0025] In one embodiment, step one involves acquiring historical plasma discharge data of the tokamak device, preprocessing it, and then constructing a time series sample using a sliding window. Specifically: (a1) Acquire plasma discharge data; Historical plasma discharge data, including both ruptured and non-ruptured discharge data, was obtained from the tokamak device experimental database. This historical plasma discharge data includes multiple diagnostic signals reflecting the plasma's operational status, such as: plasma current signal (ip), plasma vertical displacement error signal (Z_error_lmsz), loop voltage signal (v_loop), safety factor signal (q95), plasma intrinsic inductance parameter signal (li), chord integral radiation signal (pxuv), and other physical quantities reflecting plasma stability. These signals constitute multidimensional time-series data.

[0026] For each discharge gun in the partial historical plasma discharge data, record its complete time series and the corresponding end time of the flat-top segment; when a rupture occurs during the discharge process, also record the time of the rupture.

[0027] (a2) Data preprocessing; Before constructing the sliding window time series samples, the acquired historical plasma discharge data needs to be preprocessed to ensure temporal consistency between different signals and improve the stability of the rupture time prediction model training. The data preprocessing steps include time axis unification, data cleaning, data standardization, and time truncation.

[0028] (b1) Missing value repair and timeline consistency; Since some feature signals may have missing values, and the sampling frequencies of different diagnostic systems may differ, it is necessary to perform missing value repair and time alignment on each feature signal. The specific processing method is as follows: Based on the discharge gun number and processing mode (including training / test mode), the corresponding HDF5 format data file is located and read. Here, discharge gun is short for plasma discharge experiment (plasma shot / discharge). On the EAST tokamak, each experiment is called a shot, and each shot lasts from a few seconds to more than ten seconds.

[0029] A globally unified time axis (with equal time intervals) is generated based on a preset sampling rate, serving as a benchmark for aligning all feature data.

[0030] Local missing value preprocessing: For missing values ​​(missing values / infinitesimals) in the original historical plasma discharge data, local data missingness is repaired.

[0031] Unified time axis interpolation: Extract the effective time array and effective feature array of the preprocessed features of local missing values, and map the original historical plasma discharge data to the global unified time axis through linear interpolation; if there are less than 2 effective data, fill in all missing values ​​to ensure that all feature arrays are consistent with the length of the unified time axis and that the time points correspond one-to-one. Add the shot number (shot) and unified time axis (time) fields to the dataset after the above processing to form a structured data framework; integrate all preprocessed features, shot numbers and unified time axis into a DataFrame format (a two-dimensional tabular data structure) for output, and return the data loading and processing status (success / failure); if the file does not exist, return an error flag.

[0032] (b2) Data cleaning; To improve data quality, data cleaning is required after (b1) processing. Specifically, this involves removing discharge guns with excessively high NaN values, small radii, safety factors, electron densities, and Greenwald fractions that do not meet the required range.

[0033] (b3) Construct a sliding window time series sample; like Figure 2 As shown, the time series data of each shot discharge is processed by a sliding window to construct time series samples for training the rupture time prediction model.

[0034] Specifically: Set the window length to And set the sliding step size as .

[0035] This embodiment and Configure according to actual needs, for example Take 100ms. Take 10ms (fracture cannon). =150ms (non-fracture cannon).

[0036] By sliding a window across the time series, the discharge data of a single shot can be divided into multiple time-series samples. Each time-series sample contains continuous values ​​of multiple diagnostic signals within the oscillation window time range. This allows the raw historical plasma discharge data to be converted into a set of time-series samples for training a breakup time prediction model, which includes both breakup and non-breakup time-series samples.

[0037] (b4) Time truncation; This embodiment additionally truncates the time (e.g., the preset selection time is uniformly set to 0.8-1.4s, preferably 1s), using only samples with a remaining time of rupture within the set time range (within 1s) for training. This solves the problem that samples with a long time remaining after rupture cannot output discriminative distribution parameters due to similar signal values, and may even cause gradient explosion, thus resolving these problems that interfere with training. For rupture time series samples, the remaining time selection range can be set to 0.8-1.4s; for example, only samples with a remaining time of rupture less than 1s (configurable) are selected. For non-rupture time series samples, samples with a censoring time greater than 1s (configurable) have their remaining time replaced with 1s (configurable).

[0038] In one embodiment, step two, constructing the training labels for survival analysis, specifically involves: Before training, for each time series sample of the sliding window obtained by (b3), the time difference between the end of the window and the break / end of the flat top segment is calculated and used as the observation time of the sample.

[0039] For rupture time series samples, the observation time is defined as the time difference between the end of the window and the time when the rupture occurs, i.e., the remaining time of the rupture.

[0040] For non-fractured time series samples, the observation time is defined as the time difference between the end of the window and the end of the flat-top segment, i.e., the censoring time. Censoring time is used in survival analysis to measure the time at which observations terminate, referring to the point in time when observations are forced to end if the individual has not experienced the target event (such as death, failure, default, or fracture).

[0041] Meanwhile, a breakage label is set for each sample: the breakage label for broken time series samples is 1, and the breakage label for non-broken time series samples is 0.

[0042] Since excessively long observation times can lead to gradient explosion during subsequent training, and this embodiment is more concerned with the prediction effect of samples closer to the break, it is necessary to process the observation time of the time series samples. For broken time series samples, the remaining time of the break can be set to 0.8-1.4s. For example, only samples with a remaining time of less than 1s (configurable) are taken. For non-broken samples, samples with a censoring time greater than 1s (configurable) have their remaining time replaced with 1s (configurable).

[0043] This embodiment imposes a limitation on the observation time because: for broken time series samples, it is difficult to predict sliding windows that are too far from the break. If these sliding windows are added to the training, it will reduce the training effect or even directly destroy the training. For non-broken time series samples, sliding windows with a long time to the plateau may have different censoring times but similar sliding window features. In this case, if they are added to the training at the same time, these sliding windows will be input into the break time prediction model and obtain similar distribution parameters, but their remaining times are different, so the calculated survival function values ​​are different. However, the target value of the learned survival function is always 1, which will eventually lead to the loss function becoming non-numerical and gradient explosion.

[0044] Using the above method, a training dataset containing time-series samples, observation times, and break labels can be constructed.

[0045] During training, the sliding window slides from the complete timeline. The sliding window size for both broken and unbroken time series samples is 100ms (configurable), but the step size is different: 10ms (configurable) for broken time series samples and 150ms (configurable) for unbroken time series samples.

[0046] Figure 3A The study shows the trend of typical signals from the start of discharge to the breakup of the ruptured cannon with the cannon number 53888. These signals include ip (current), v_loop (loop voltage), and li (internal inductance), where each point represents a sampling point. Figure 3BShowing Figure 3A The current change trend 1 second before rupture, and the corresponding sample labels (remaining time before rupture and rupture label). Each point represents a sliding window sample. The horizontal axis represents the last moment of a sliding window, and the vertical axis represents the corresponding value.

[0047] Figure 4A The changes in typical signals from the start to the end of discharge in the non-fractured gun number 87608 are shown. These signals include ip (current), v_loop (loop voltage), and li (internal inductance), where each point represents a sampling point. Figure 4B Showing Figure 4A The current change trend from the start of discharge to the end of the flat-top segment, and the corresponding sample labels (censoring time and breakage label), where each point represents a sliding window sample, each point on the horizontal axis represents the end time of a sliding window, and each point on the vertical axis represents the corresponding value.

[0048] In one embodiment, in order to predict the remaining time of plasma breakup, this embodiment establishes a breakup time prediction model based on survival analysis.

[0049] Survival analysis is a type of statistical method used to analyze time-to-event data, with its core objective being to describe the probability distribution of an event occurring in the future. In this embodiment, the event is a plasma rupture event.

[0050] Within the framework of survival analysis, a survival function can be used to describe the time interval [of time]. The probability that the plasma will continue to operate stably is defined as follows: ; in, Indicates the time when the rupture occurred. Indicates the plasma at time The probability that rupture will not occur afterward. By modeling the survival function, the probability distribution of plasma rupture time can be obtained, thereby enabling the prediction of the remaining rupture time.

[0051] Another important function in the survival analysis framework is the probability density function, which describes the event at time t. The probability density of events occurring within an infinitesimal time interval in the vicinity.

[0052] This embodiment uses the Weibull distribution to model the rupture time, and its survival function... and probability density function for: ; ; in, To represent the remaining time from the current observation time, For shape parameters, This is the scale parameter.

[0053] In this embodiment, a rupture time prediction model is constructed to predict the above-mentioned... The prediction model comprises two parts: a feature extraction module and a distribution parameter module. Specifically, time-series samples from a sliding window are input into the rupture time prediction model. The feature extraction module extracts plasma operational state features, and the distribution parameter prediction module outputs the shape parameters of the Weibull distribution. Scale parameters The weight parameters refer to the parameter weights of each Weibull distribution when multiple Weibull distributions are selected. In this embodiment, a Weibull distribution is selected.

[0054] Model training: The training dataset containing time-series samples, observation times, and breakage labels is input into the established breakage time prediction model, and the model's loss function is minimized. Train the model parameters.

[0055] ; in, For broken labels, The total number of training samples, For training sample index, For time The probability density function at which a break occurs. The survival function, i.e., time. The probability that no further breakage will occur. The breakage weight is used to address the imbalance between broken and unbroken time series samples. When the fracture time prediction model increases the attention weight placed on the fracture cannon; when When the rupture time prediction model is in the standard state; when At that time, the fracture time prediction model increases the attention weight placed on non-fractured shots; in this example, it is preferable to use... .

[0056] During training: rupture time series samples are used as event occurrence samples for training, while non-rupture time series samples are used as censored samples for training. By simultaneously utilizing both rupture and non-rupture time series samples, experimental data can be used more fully, improving the ability of the rupture time prediction model to characterize the plasma operating state and enhancing the model's generalization performance.

[0057] In one embodiment, in step four, the trained rupture time prediction model is used to predict the remaining rupture time.

[0058] After the rupture time prediction model is trained, the plasma diagnostic signals collected in real time are input into the trained rupture time prediction model (in the form of a sliding window).

[0059] The rupture time prediction model outputs the survival function distribution of the remaining rupture time. For example, if a fixed time window t=0.2s is defined, the model outputs a survival probability every 0.01s from 0 to 0.2s, and calculates the remaining time of plasma rupture within 0.2s (with a maximum value of 0.2s) by integrating the results, thereby achieving the prediction of the remaining time of plasma rupture.

[0060] The remaining time to fracture is obtained by calculating the confined mean survival time: ; in, To limit the mean survival time, it represents the time from the start of observation to [time value missing]. The lifespan of the observed object, i.e. Inside, the area under the survival curve, For the preset time window, For time, as the integration variable, This is a survival function. The prediction results can provide a basis for decision-making in the rupture early warning and control system of a tokamak device.

[0061] This embodiment obtains the plasma survival function. Subsequently, a hierarchical risk assessment mechanism based on survival probability over time intervals was constructed. According to the response requirements of the tokamak control system to rupture events at different time scales, the system was divided into multiple risk intervals, and a survival function was used to assess these intervals. The probability of rupture occurring within each interval is calculated, achieving a comprehensive modeling of the rupture time distribution. Its core advantages are: (1) Same model, any time window: The rupture time prediction model proposed in this embodiment outputs a survival function. That is, from the current moment to any point in the future. The probability that no rupture occurs afterward. Based on this, the rupture probability within any time window can be obtained simply by integrating or differencing the survival function, without retraining the model. (2) Achieve multi-level and multi-time-scale rupture risk assessment. (3) Flexibly adapt to actual control requirements: The control system can dynamically adjust the time window threshold according to the current device status and task requirements to obtain rupture risk distributions with different levels of precision.

[0062] Specifically, let the first preset time threshold be... The second preset time threshold is ,in The first risk probability PM represents the plasma within the time interval. The probability of internal rupture is calculated using the following formula: .

[0063] Second risk probability Indicates the plasma within the time interval The probability of internal rupture is calculated using the following formula: .

[0064] Third risk probability Indicates the plasma in time The probability of maintaining stable operation thereafter is calculated using the following formula: .

[0065] This leads to the construction of a risk vector. : ; in, These correspond to the probabilities of three operational states: mitigation, avoidance, and prevention. Since these three probabilities represent mutually exclusive and complete partitions of future rupture times, they satisfy the following: .

[0066] This embodiment further considers the risk vector. Determine the control decision-making model. First, calculate the component with the highest probability in the risk vector: ; when When the value is at its maximum, the current plasma is determined to be in a high-risk fracture state, and the system enters a fracture mitigation mode, triggering fracture mitigation measures such as fragmentation injection (SPI) and large gas injection (MGI); when When the value is at its maximum, the current plasma is determined to be in an intervention-risk state, and the system enters a breakage avoidance mode. Breakage avoidance control is achieved by adjusting heating power, density control parameters, magnetic configuration parameters, or current profile parameters. When the value is at its maximum, it is determined that the current plasma is in a stable operating state, and the system maintains normal control mode.

[0067] Existing methods define the fracture state as within a certain time (e.g., 100ms) before the fracture cannon fractures before training. After training, when a sliding window sample is input, it can only provide the fracture risk value (e.g., 0.7) at the current moment (the last moment of the sliding window). If this risk value exceeds the set threshold, it means that the probability of fracture within the next 100ms is very high. This means that once the model is trained, its output prediction is fixed in the time dimension and cannot be flexibly adjusted according to the control system's requirements for different response times. If the warning time needs to be changed (e.g., from 100ms to 200ms), the labels need to be reconstructed and the model retrained, which greatly reduces the flexibility and engineering practicality of the method.

[0068] Compared to existing binary classification fracture prediction methods that only output a single risk value, this embodiment utilizes a survival function to construct a time-interval probability distribution, extending the prediction from whether fracture will occur to when and how to control it. This provides multi-level risk assessment and hierarchical control decision-making basis for plasma control systems, improving the targeting and real-time performance of control strategies. Example: In actual tokamak device operation, the plasma control system needs to adopt different strategies at different time scales: for instance, fracture avoidance measures (such as adjusting heating power and magnetic configuration) are taken a considerable amount of time before fracture (e.g., 50ms-300ms), while fracture mitigation systems (such as pellet injection) need to be triggered in a very short time (e.g., within 50ms). Traditional methods, unable to provide probability assessments across multiple time windows, struggle to support such hierarchical and dynamic control strategies.

[0069] Example: Taking a certain EAST discharge cannon as an example, the time period is taken The sliding window acquires diagnostic data such as PXUV radiation signals, Dα signals, heating power signals, and plasma equilibrium parameters, and inputs them into the trained rupture time prediction model for inference calculation to obtain the corresponding survival function. .

[0070] Set the first preset time threshold The second preset time threshold is 50ms. The time is 300ms. The model output is as follows: ; .

[0071] According to the time interval probability calculation method proposed in this embodiment, we can obtain: ; ; .

[0072] Therefore, a risk vector is constructed: .

[0073] in, This indicates an 8% probability of rupture occurring within the next 50ms; This indicates a 67% probability of rupture occurring within the next 50-300 ms; This indicates a 25% probability that the system will remain stable after 300ms.

[0074] Further calculate the component with the highest probability in the risk vector: .

[0075] Therefore, the system determines that the current plasma is in a fracture avoidance state, meaning there is a high risk of fracture, but a certain control margin is still available. At this time, the control system enters the fracture avoidance control mode, reducing the risk of fracture by adjusting auxiliary heating power, density control parameters, magnetic configuration parameters, and current profile parameters.

[0076] In subsequent runs, if the model outputs: , .

[0077] Then we can obtain: , , .

[0078] Corresponding risk vector: .

[0079] because Since this value represents the largest component in the risk vector, the system determines that the current plasma has entered a fracture mitigation state and is expected to fracture within a very short time. At this point, the control system immediately activates the fracture mitigation control mode, triggering pellet injection (SPI), large gas injection (MGI), or other fracture mitigation measures to reduce the damage caused by the fracture to the device structure and internal components.

[0080] The above embodiments demonstrate that this embodiment can not only predict the remaining plasma breakup time, but also construct the breakup probability distribution in different time intervals in the future using the survival function, thereby realizing automatic switching between breakup mitigation control, breakup avoidance control and stable operation control, and thus providing a multi-level risk assessment and hierarchical control decision basis for the plasma control system.

[0081] This embodiment can make full use of non-rupture discharge data, and the training can directly predict the remaining time of plasma rupture. By modeling the rupture time distribution, there is no need to manually set a fixed warning time window to construct labels, avoiding the problem of retraining the model every time the warning time is changed, thus improving the flexibility and applicability of the method. In addition, it has good real-time prediction capability. By inputting the plasma diagnostic signal sliding window data at the current moment, it can output the prediction result of the remaining time of rupture in real time. It has high computational efficiency and is suitable for the real-time rupture early warning and control system of tokamak device. This embodiment provides a plasma breakup remaining time prediction system based on survival analysis, which mainly includes a data acquisition module, a label building module, a model building and training module, and a model execution module.

[0082] (d1) Data acquisition module; First, various diagnostic signals (i.e., historical plasma discharge data) during plasma operation are obtained from the diagnostic system of the tokamak device, including: q95, li, ip, v_loop, Z_error_lmsz and ten radiation signals (pxuv9, pxuv52, pxuv13, pxuv46, pxuv17, pxuv42, pxuv22, pxuv36, pxuv26 and pxuv32, changed to ascending order 9, 13, 17, 22, 26 ....).

[0083] The above signals are preprocessed as follows: Time axis unification: Interpolating signals acquired by different diagnostic systems onto a time axis with a unified time resolution; Data cleaning: removing outliers and missing data; Data standardization: Normalize each signal to eliminate the impact of dimensional differences on model training.

[0084] Subsequently, a sliding window method was used to construct time-series samples, with a window length between 50 and 200 ms. The step size for broken samples was between 5 and 20 ms, and the step size for unbroken samples was between 50 and 200 ms. Each sample consisted of multidimensional diagnostic signals within a continuous time window.

[0085] (d2) Tag building module; For each time series sample of a sliding window, the following labels are constructed: Observation time: the time interval from the end of the window to the rupture time or the end of the flat-top segment; Event indicator variable: if the discharge ruptures, it is recorded as 1, otherwise it is recorded as 0.

[0086] For non-ruptured time series samples, the observation time is the time from the end of the window to the end of the top segment; for ruptured time series samples, the observation time is the time from the end of the window to the rupture time.

[0087] (d3) Model building and training module; The rupture time prediction model includes a feature extraction module and a distribution parameter prediction module.

[0088] The temporal samples from the sliding window are input into the feature extraction module for feature extraction. This feature extraction module consists of two convolutional blocks and a temporal attention mechanism layer. The first convolutional block (Block 1) contains two one-dimensional convolutional layers with kernel sizes of 7 and 5 respectively. Each convolutional layer is followed by batch normalization (BatchNorm) and ReLU activation function, and the output shape is... The second convolutional block (Block 2) also contains two one-dimensional convolutional layers with kernel sizes of 5 and 3 respectively. Each layer is followed by batch normalization and ReLU, and the output shape is... .

[0089] For the temporal attention mechanism layer, this layer consists of two one-dimensional convolutional layers with a kernel size of 1. The first one-dimensional convolution is followed by a batch normalization layer and a Tanh activation function, and then the second one-dimensional convolution compresses the number of channels to 1, resulting in a shape... The attention score is calculated. A softmax operation is performed on the last dimension of this score to obtain the temporal attention weight vector for each sample. This vector is then compared with the output of the second convolutional block (Block 2). Element-wise multiplication and summation in the time dimension yields a shape of The feature representation is that each sample is adaptively weighted in the time dimension for each feature channel.

[0090] This feature extraction module models historical plasma discharge data and extracts high-level features that reflect the plasma's operating state and dynamic changes, thereby obtaining a feature representation that can characterize the plasma behavior before rupture.

[0091] In this embodiment, the Weibull distribution is used to model the rupture time, and the neural network outputs the parameters of the distribution, including: scale parameter, shape parameter and weight parameter.

[0092] After obtaining the temporal features, the distribution parameters of the rupture time prediction model are output through the distribution parameter prediction module. This module consists of three independent fully connected layers, each followed by a batch normalization layer, a ReLU activation function, and Dropout regularization. The features are sequentially compressed in dimensionality through these fully connected layers and then fed into three output heads: scale parameters... Activated via Softplus to ensure positive values, shape parameters through Ensure the weight parameter is greater than 1. The parameters of the Weibull mixture distribution are finally output after ensuring that the normalized summation is 1 by using softmax, and the parameters are used for survival analysis.

[0093] The number of each of these parameters (scale, shape, and weight parameters) is called the number of components, typically 1-3 (optional). The feature vectors output by the feature extraction module are input into the three fully connected layers of the distribution parameter prediction module to obtain the scale, shape, and weight parameters. Then, each set of scale and shape parameters is substituted into the survival function. Finally, by summing the calculated survival function values ​​and weight parameters, a sliding window can be obtained at different times. Survival functions under [the specified conditions].

[0094] Figure 5A and Figure 5B This test focuses on the predicted results and typical plasma signal variation trends of the fracture shot with shot number 90202. In the fracture shot, the observation time is the remaining fracture time, which is the time from the end of the sliding window to the fracture time. The sliding window length is 100 ms, and the step size is 10 ms. Figure 5A In the diagram, each point on the horizontal axis represents the end time of a sliding window. The blue line represents the survival probability within 0.2 seconds, starting from the end of a single sliding window. The survival probability is close to 1 in the early stages and drops sharply in the short period before rupture. The integral of the survival function for each sample from 0 to 0.2 seconds is calculated using the restricted average survival time formula to obtain the predicted remaining survival time (with an upper limit of 0.2 seconds). The green line represents the predicted remaining survival time, and the orange line is the remaining time before rupture. Here, to make the change in the predicted remaining time before rupture more obvious, the time of the sliding window with an actual remaining time greater than 1 second is marked as 1 second.

[0095] Figure 5B The variation trends of typical plasma signals are shown, including plasma current signal (ip), safety factor signal (q95), plasma intrinsic inductance parameter signal (li), loop voltage signal (v_loop), and plasma vertical displacement error signal (Z_error_lmsz). It is important to note that... Figure 5BThe signals are taken from the beginning of the flat-top section to the end of the discharge in a single shot; the horizontal axis represents the corresponding time. Figure 5A The x-coordinate is taken at the end of the sliding window, so Figure 5B Compare Figure 5A High sampling frequency, ( Figure 5B It samples once every 1ms. Figure 5A (Equivalent to sampling once every 10ms).

[0096] Figure 6A and Figure 6B This test focuses on the prediction results and typical plasma signal variation trends of the non-fractured shot (shot number 93648) in the test group. In the non-fractured shot, the observation time is the censoring time, which is the time from the end of the sliding window to the end of the flat-top segment. The sliding window length is 100 ms, and the step size is 150 ms. Figure 6A In the model, the survival function curve (blue line) is always close to 1, and the remaining survival time curve (green line) is always close to 0.2s (the upper limit of prediction is 0.2s). Here, the sliding window time with a censoring time greater than 1s is marked as 1s.

[0097] Figure 6B The variation trends of typical plasma signals are shown, including plasma current signal (ip), safety factor signal (q95), plasma intrinsic inductance parameter signal (li), loop voltage signal (v_loop), and plasma vertical displacement error signal (Z_error_lmsz). It is important to note that... Figure 6B The signals are taken from the beginning of the flat-top section to the end of the discharge in a single shot; the horizontal axis represents the corresponding time. Figure 6A The x-coordinate is taken at the end of the sliding window, so Figure 6B Compare Figure 6A High sampling frequency, ( Figure 6B It samples once every 1ms. Figure 6A (Equivalent to sampling once every 150ms).

[0098] Figure 7A and Figure 7B yes Figure 5A and Figure 5B Different time window versions, Figure 8A and Figure 8B yes Figure 6A and Figure 6B Different versions of the time window changed the time window from 0.2s to 0.3s. From Figure 5A and Figure 7A This demonstrates that as the set time window increases, predicting the remaining fracture time in a fracture cannon becomes more difficult. Figure 6A and Figure 8A This demonstrates that in non-fracture cannons, the remaining time for fracture is always close to the time window, and even if the time window is increased, the effect is still good.

[0099] Evaluation metrics: To test the performance of the rupture time prediction model, some evaluation metrics are needed for verification. Three metrics are designed, namely IBS (Integrated Brier Score), AUC (Area Under the Curve), and MAE (Mean Absolute Error). Rupture samples with a remaining time of less than 1 s for 200 shots are used for testing.

[0100] The Brier Score (BS) is used to measure the deviation between the predicted value of the event occurrence probability and the actual observed result, and the IBS is the integral of the Brier Score on the time axis, reflecting the average prediction error of the model at all time points, which is a comprehensive evaluation of the full-cycle performance of the model. Figure 9 is the image of IBS. As the time window increases, IBS also increases. The evaluation criteria for IBS are that IBS < 0.05 is excellent, 0.05 < IBS < 0.1 is good, and 0.1 < IBS < 0.15 is average.

[0101] The AUC in this embodiment is different from that in binary classification rupture prediction and is used to rank samples close to rupture and far from rupture. Specifically, first, positive and negative sample labels need to be assigned. A time window is set artificially. Samples with a remaining time greater than this time window are labeled 0, that is, negative samples, and samples with a remaining time less than this time window are labeled 1, that is, positive samples. Then, the risk values of all samples in this time window are calculated. All positive and negative samples are freely combined and compared. If the risk value of a positive sample is greater than that of a negative sample, add 1; if they are equal, add 0.5; if it is smaller, add 0. Finally, the sum result is averaged to obtain the AUC under this time window. Figure 10 is the image of AUC. It can be seen that as the time window increases, AUC becomes smaller.

[0102] As Figure 11 shown, MAE is to take the absolute value of the difference between the predicted remaining time and the actual remaining time. Here, a time window also needs to be set artificially. The time window set here refers to only judging the accuracy of the predicted remaining time of samples with a remaining time within this time window. Figure 11 is the image of MAE. As time increases, MAE becomes larger.

[0103] (d4) Model execution module; According to the predicted survival function, the probability that the plasma has not ruptured at different future time points can be calculated. Further, by integrating the survival function, the predicted value of the remaining time until the plasma ruptures can be obtained, thereby achieving continuous prediction of the rupture time.

[0104] Using the above methods, the plasma state before rupture can be modeled using various plasma diagnostic signals, and the rupture time can be predicted by combining survival analysis methods.

[0105] Compared with existing prediction methods that only output rupture risk values ​​(Cannas et al. applied support vector machine (SVM) to tokamak rupture prediction in 2004, proving the feasibility of machine learning methods in rupture early warning), this example can directly give the remaining time of plasma rupture, improving the interpretability and practical value of rupture prediction results.

[0106] Compared to regression methods that only use the remaining rupture time as a label (Reimerdes et al. conducted a real-time rupture handling study on ASDEXUpgrade in 2016, demonstrating that rupture warning time and control response strategies have a significant impact on rupture handling), this method can utilize both rupture discharge data and non-rupture discharge data for training, improving the prediction accuracy and generalization ability of the rupture time prediction model. It should be noted that the sliding window starts sliding from the top segment during testing.

[0107] Based on the above description of the embodiments, those skilled in the art will understand that the fusion plasma breakup time prediction method and system based on survival analysis described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0108] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the breakup time of fusion plasma based on survival analysis, characterized in that, include: Historical plasma discharge data of a tokamak device are acquired, preprocessed, and then time series samples are constructed using a sliding window. The time series samples include rupture time series samples and non-rupture time series samples. For each sliding window time series sample, a training label is constructed, including the observation time and the break label. The break label of the broken time series sample is 1, and the observation time is the time difference from the end of the window to the break time. The break label of the non-broken time series sample is 0, and the observation time is the time difference from the end of the window to the end of the flat top segment. A rupture time prediction model was established, and the model was trained using time series samples and training labels as inputs. The model parameters were adjusted by the constructed survival analysis loss function. The real-time acquired plasma diagnostic signals are input into the trained rupture time prediction model, which outputs the survival function and calculates the remaining rupture time. In the survival analysis, the loss function is modeled using the Weibull distribution and its probability density function. and survival function : ； ； in, To represent the remaining time from the current observation time, For shape parameters, This is the scale parameter.

2. The method according to claim 1, characterized in that, The survival analysis loss function for: in, For broken labels, The total number of training samples, For training sample index, For time The probability density function at which a break occurs. The survival function, i.e., time. The probability that no further breakage will occur. For breakage weights.

3. The method according to claim 2, characterized in that, when At that time, the fracture time prediction model increases the attention weights placed on the fracture cannon, when When the rupture time prediction model is in the standard state; when At that time, the fracture time prediction model increases the attention weight placed on non-fractured guns.

4. The method according to claim 1, characterized in that, The remaining time of fracture is obtained by calculating the limited mean survival time: ； in, To limit average survival time, For the preset time window, For time, as the integration variable, This is the survival function.

5. The method according to claim 1, characterized in that, The preprocessing includes: timeline unification, data cleaning, data standardization, and time truncation; The time truncation specifically involves: setting a filtering range for the remaining time of the rupture; for ruptured time series samples, only samples with a remaining time of rupture less than a preset filtering time are selected; for non-ruptured time series samples, samples with a censoring time greater than the preset filtering time are replaced with the preset filtering time for their remaining time.

6. The method according to claim 1, characterized in that, The sliding window has a window length of 50ms-200ms, the sliding step size for ruptured discharge samples is 5ms-20ms, and the sliding step size for non-ruptured discharge samples is 50ms-200ms.

7. The method according to claim 1, characterized in that, The rupture time prediction model includes a feature extraction module and a distribution parameter prediction module; The feature extraction module includes a one-dimensional convolutional layer, a temporal attention mechanism layer, and a fully connected layer connected in sequence. The temporal attention mechanism layer consists of two sequentially connected one-dimensional convolutional layers.

8. The method according to claim 1, characterized in that, Using survival functions, a hierarchical risk assessment mechanism based on survival probability over a time interval is constructed, specifically as follows: Let the first preset time threshold be... The second preset time threshold is ,in ; Calculate plasma in time interval The first risk probability of internal rupture ; Calculate plasma in time interval The second risk probability of internal rupture ; Calculate plasma in time The third risk probability that it will continue to operate stably. ; Constructing risk vectors This yields the component with the highest probability in the risk vector; when When the value is at its maximum, the current plasma is determined to be in a high-risk rupture state, and the system enters the rupture mitigation mode; when When the value is at its maximum, the current plasma is determined to be in a state of manageable risk, and the system enters the rupture avoidance mode; when When the value is at its maximum, it is determined that the current plasma is in a stable operating state, and the system maintains normal control mode.

9. A fusion plasma breakup time prediction system based on survival analysis, characterized in that, It includes a data acquisition module, a label building module, a model building and training module, and a model execution module; The data acquisition module is used to acquire historical plasma discharge data of the tokamak device. After preprocessing, a time series sample is constructed using a sliding window. The time series sample includes rupture time series sample and non-rupture time series sample. The label building module is used to build training labels for each time series sample of the sliding window, including observation time and break label. The break label of the broken time series sample is 1 and the observation time is the time difference from the end of the window to the break time. The break label of the non-broken time series sample is 0 and the observation time is the time difference from the end of the window to the end of the flat top segment. The model building and training module is used to build a rupture time prediction model. It uses time series samples and training labels as inputs to train the model and adjusts the model parameters through a constructed survival analysis loss function. The model execution module is used to input the real-time acquired plasma diagnostic signals into the trained rupture time prediction model, output the survival function, and calculate the remaining rupture time.

10. The system according to claim 9, characterized in that, The survival analysis loss function for: in, For broken labels, The total number of training samples, For training sample index, For time The probability density function at which a break occurs. The survival function, i.e., time. The probability that no further breakage will occur. For breakage weights.

Images