A method for automatically discovering a precursor event chain of a magnetic confinement plasma discharge phenomenon

Abstract

The application provides a magnetic confinement plasma discharge phenomenon precursor event chain automatic discovery method, and belongs to the technical field of magnetic confinement nuclear fusion device operation and control. The method comprises the following steps: through gradual multi-round time peeling iterative calling of trigger time identifier, performing explainable attribution on each round of trigger, constructing signed split attribution embedding, performing confidence weighted constraint dynamic programming stage clustering and conditional Markov chain analysis, and automatically discovering target phenomenon precursor event chain and evolution path from a discharge database containing only shot level labels. The peeling mechanism, attribution embedding, constraint clustering and Markov analysis defined in the application are not bound to a specific trigger time identifier. As long as the trigger time identifier selected can identify the first trigger time of the target phenomenon from the diagnosis time sequence output, the framework can be reused without modifying any module structure, thereby significantly reducing the engineering cost of building a precursor knowledge base on a new device, a new phenomenon or an existing trigger of an existing operation system.

Description

Technical Field

[0001] This invention belongs to the field of operation and control technology of magnetic confinement nuclear fusion devices, and more specifically, relates to an automatic detection method for precursor event chains of magnetic confinement plasma discharge phenomena. Background Technology

[0002] Magnetic confinement fusion devices (including but not limited to tokamaks, stellarators, and antifield pinch devices) exhibit numerous physically significant but temporally sparse phenomena during a single discharge, such as plasma breakup, mode locking, tear mode growth, MARFE-type radiation-thermal collapse, low-confinement to high-confinement (L–H) transition, edge localized mode (ELM) bursts, sawtooth collapse, escape electron generation, and Greenwald density limit approximation. These phenomena are often not directly triggered by a single physical event, but rather cascaded from several precursor events in a specific temporal sequence; they may occur in different orders and at different time intervals in different discharge stages, ultimately converging into a target phenomenon with the same discharge stage label. Therefore, the automatic discovery of "precursor event chains of plasma discharge phenomena" has common and crucial value for revealing the physical mechanisms of devices under different operating parameters and establishing strategies for early mitigation and avoidance of multi-path target phenomena.

[0003] To address the fundamental requirement of providing the initial triggering time of a target phenomenon within the first-shot diagnostic timeline, various available models exist in this field, including but not limited to: hard threshold alarms based on manually selected physical criteria (e.g., mode-locked amplitude threshold, radiation-ohmic power ratio threshold, Greenwald density upper limit, etc.), gun-level and window-level classifiers based on traditional machine learning, end-to-end timeline classifiers based on recurrent neural networks or convolutional neural networks, weakly supervised triggering models based on multiple instance learning (MIL), unsupervised methods based on timeline anomaly detection and change point detection, and triggering times manually annotated by experts in specialized tools. These models are isomorphic in their output capabilities; they can all provide the initial triggering time of a target phenomenon within the first-shot diagnostic timeline. However, they differ significantly in their principles, retrainability, differentiability, and interpretability interfaces.

[0004] To systematically discover the precursor event chains of target phenomena in magnetically confined plasma discharge, the following significant technical bottlenecks currently exist: 1. The engineering gap from single triggering to event chains: Existing trigger moment identifiers can only output "one moment per shot," failing to answer "what precursor events occurred before this moment, and what their order and mutual exclusion relationships are." To automatically extract nameable precursor stages and their transition relationships from a large number of trigger samples, current methods rely on manual shot-by-shot backtracking, which cannot cover databases of thousands of shots. 2. The obscuring of early precursors by strong evidence in the later stages: Regardless of whether the trigger moment identifier uses threshold rules, traditional classifiers, deep neural networks, or multi-instance learning, its output is almost always attracted by the strongest evidence closest to the occurrence of the target phenomenon. Earlier, weaker precursor events are simply not identified by the model in a single call, resulting in clustering the model output directly only yielding synonymous repetitions of "late-stage events." 3. Lack of a physically interpretable precursor event category system: Existing unsupervised clustering is performed directly on the original feature space or encoder output space. The resulting cluster centers often cannot correspond to states that can be named by physicists, and it is difficult to maintain stability between different batches. When the number of clusters changes, the physical interpretation is completely rewritten, resulting in poor engineering usability. 4. Ignoring the temporal constraints between precursor events: Clustering all precursor events from different batches together can lead to results that violate physical common sense, such as "two events in the same batch within the same batch are far apart" or "events appear in reverse order on the timeline," making the final event chain and evolution path unreliable. 5. Unstable interpretability attribution and loss of directional information: Traditional attribution methods such as SHAP fluctuate greatly in small-sample clustering and usually cluster based on absolute values, losing the physical differences between "positive contributions" and "negative contributions" (for example, a physical quantity may be highlighted for both increases and decreases, but the physical meanings are completely different). 6. Lack of readable output on causal / path relationships between precursor events: Even with event clustering, there is a lack of statistical summaries on transition probabilities, path divergences, and mutual exclusivity between events, making it difficult to transform into a readable evolutionary path map for operators. 7. Difficulty in ensuring consistency of trigger times across multiple iterations: Simply discarding all shots at a uniform depth using a time window will significantly shift the model's input distribution during deep iterations, and triggers discovered in earlier iterations may be incorrectly relocated in deeper iterations, disrupting the consistency of the event chain time series. 8. Strong binding between the method and the trigger time identifier type: Existing precursor research for different trigger time identifiers (e.g., threshold rules vs. deep networks vs. multi-instance learning) often designs dedicated downstream processes for each, lacking a general precursor event chain discovery framework that is decoupled from specific trigger time identifiers and reusable across different types of trigger time identifiers. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an automatic detection method for precursor event chains of magnetic confinement plasma discharge phenomena, thereby solving the technical problem that it is difficult to automatically detect precursor event chains of target phenomena (such as plasma rupture) during the discharge process of magnetic confinement fusion devices in existing technologies.

[0006] To achieve the above objectives, according to one aspect of the present invention, an automatic detection method for precursor event chains of magnetically confined plasma discharge phenomena is provided, comprising: Step S1: Obtain any model that can output the first trigger time of the target phenomenon in the single-shot diagnostic timing sequence as a trigger time identifier, and call the trigger time identifier on the complete single-shot diagnostic timing sequence without time stripping to obtain the first trigger time; Step S2: Starting from the second round, a progressive multi-round time stripping iteration with a stop condition determination is executed. Each round of iteration includes: for each gun carrying a target phenomenon tag, the time window from the trigger time detected in the previous round to the end of the discharge is discarded as the boundary, forming a stripped time sequence; the trigger time identifier is called on the stripped time sequence to obtain the new trigger time of the gun in this round; wherein, the stripping depth of different guns is independently determined by their respective trigger times detected in the previous round. Step S3: Validity filtering is performed on the trigger times obtained in all rounds, and for each valid trigger, interpretability attribution is performed in accordance with the type of trigger time identifier from which it originates, to obtain a signed attribution value. Then, the signed attribution value is split into two independent non-negative coordinates according to the positive and negative components to construct a signed split attribution embedding vector. Step S4: Based on the signed split attribution embedding vector, perform confidence-weighted clustering and integrate time-series hard constraints on all valid triggers to divide the triggers into different precursor stages. The time-series hard constraints include: the stage label sequence of each trigger in the same shot must be monotonically non-decreasing on the time axis, and the time distance between two triggers in the same shot that are classified into the same stage must not exceed a preset upper bound. Step S5: According to the occurrence time sequence of each effective trigger within each shot, the continuously repeating stage labels are concatenated and merged to obtain the event chain of that shot. Based on the event chain statistics of all shots carrying target phenomenon labels, a conditional Markov chain representing the divergence of different evolution paths is constructed.

[0007] Preferably, in step S2, the peeling depth of each cannon carrying the target phenomenon tag is determined by taking the upper bound of the trigger time detected by the cannon in the previous round and the maximum value of the historical peeling depth of the cannon; cannons without the target phenomenon tag are not subjected to time peeling, but are re-cut and balanced according to the visible length distribution of cannons carrying the target phenomenon tag in the corresponding round.

[0008] Preferably, the progressive multi-round time stripping iteration stops when any of the following conditions are met: The classification index or detection rate at a fixed peeling depth is lower than the preset lower limit, the length baseline classification index is higher than the preset upper limit, the false positive rate for several consecutive rounds is higher than the preset warning line, or the number of remaining propulsive target-carrying phenomena tag guns is lower than the preset lower limit.

[0009] Preferably, the validity filtering in step S3 includes: Triggers or protection rounds whose classification indicators exceed the high threshold are directly retained. Apply a time-axis-based random permutation significance test to the triggers in the remaining rounds, and retain only the triggers that show significant results; For false positive triggers occurring on guns without target phenomenon labels in all rounds, perform cross-round nearest neighbor merging and perform unstable trigger screening based on multi-round confirmation results.

[0010] Preferably, the interpretability attribution in step S3 is selected in one of the following ways depending on the type of the trigger time recognizer: When the triggering time identifier is a differentiable depth model, gradient-type SHAP or integral gradient is used. When the trigger moment identifier is a black-box classifier, kernel SHAP, LIME, or permutation importance is used. When the triggering time identifier is a timing anomaly detection or change point detection model, a decomposition based on residual contribution is adopted; When the trigger time identifier is a threshold rule model, criterion contribution decomposition is adopted.

[0011] Preferably, the confidence weighting in step S4 is calculated by combining one or more of the following factors: the earlier the trigger time is relative to the occurrence time of the target phenomenon, the higher the weight; the earlier the round in which the trigger time is located, the higher the weight; and the weight of the detection signal type at the trigger time, which is a forward residual peak or a signal based on the amount of evolutionary mutation, is higher than that of a signal based on the instantaneous absolute quantity.

[0012] Preferably, the clustering in step S4 is performed in a two-stage manner: In the first stage, effective triggers with confidence levels higher than the group median are taken as anchor sets, and several initial prototypes are estimated in the signed split attribution embedding space using a weighted mixture model. In the second stage, using the initial prototype as a fixed anchor, a dynamic programming phase allocation with the temporal hard constraints is performed on all valid triggers, where the allocation cost is a confidence-weighted average of the distance from the trigger to the candidate prototype in the signed split attribution embedding space.

[0013] Preferably, the construction of the conditional Markov chain in step S5 includes: First, the transition frequencies of adjacent stages in the entire event chain are counted and normalized to obtain the first-order Markov transition probability matrix. Then, for the key intermediate stages that exhibit multi-path divergence, they are divided into several conditional sub-states according to their preceding stages, and the conditional transition probabilities are re-estimated to obtain the conditional Markov chain.

[0014] Preferably, the trigger time identifier is a single model or a combination of multiple models; when the trigger time identifier is a combination of multiple models, the trigger times given by each model are merged according to time proximity or the multiple models are confirmed and filtered before proceeding to step S3.

[0015] Preferably, the target phenomenon label is a binary label or a multi-class label; when the target phenomenon label is a multi-class label, steps S1 to S5 are executed separately for each type of target label, or aggregated separately for multiple conditions on the shared precursor trigger set.

[0016] According to another aspect of the present invention, an automatic detection system for precursor event chains of magnetically confined plasma discharge phenomena is provided, comprising: The trigger time identifier module is used to output the first trigger time of the target phenomenon in the single-shot diagnostic timing sequence, and to call the trigger time identifier on the complete single-shot diagnostic timing sequence without time stripping to obtain the first trigger time. The progressive multi-round time stripping module is used to, starting from the second round, discard the time window from the triggering time detected in the previous round to the end of the discharge for each gun carrying the target phenomenon tag, forming a stripped time sequence, and calling the triggering time identifier module on the stripped time sequence to obtain the new triggering time of the gun in this round; wherein, the stripping depth of different guns is independently determined by their respective triggering times detected in the previous round. The attribution and embedding module is used to filter the validity of the trigger times obtained in all rounds, and for each valid trigger, to perform interpretable attribution that is compatible with the type of the trigger time identifier module, to obtain a signed attribution value, and then to split the signed attribution value into two independent non-negative coordinates according to the positive and negative components, and to construct a signed split attribution embedding vector. The constrained clustering module is used to perform confidence-weighted clustering and integrate temporal hard constraints on all valid triggers based on the signed split attribution embedding vector, and to divide the triggers into different precursor stages. The temporal hard constraints include: the stage label sequence of each trigger in the same shot must be monotonically non-decreasing on the time axis, and the time distance between two triggers in the same shot that are classified into the same stage must not exceed a preset upper bound. The event chain and path analysis module is used to chain and merge continuously repeating stage tags according to the occurrence time sequence of each effective trigger within each shot to obtain the event chain of that shot, and to construct a conditional Markov chain representing the divergence of different evolution paths based on the event chain statistics of all shots carrying target phenomenon tags.

[0017] Preferably, the progressive multi-round time stripping module is further configured to determine whether to stop iteration based on one of the following conditions: the classification index at the current stripping depth is lower than a preset lower limit; the index of the baseline classifier with only sequence length as input at the current stripping depth is higher than a preset upper limit; the false positive rate is higher than a preset warning line for several consecutive rounds; and the number of remaining shots carrying target phenomenon tags that can continue to be advanced is lower than a preset lower limit.

[0018] Preferably, in the attribution and embedding module, the interpretability attribution is selected according to one of the following methods based on the type of the trigger moment recognizer module: when it is a differentiable deep model, gradient class SHAP or integral gradient is used; when it is a black box classifier, kernel SHAP, LIME or permutation importance is used; when it is a temporal anomaly detection or change point detection model, residual contribution-based decomposition is used; when it is a threshold rule model, criterion contribution decomposition is used.

[0019] Preferably, the constrained clustering module is specifically used to: take valid triggers with confidence scores higher than the population median as anchor sets, estimate several initial prototypes in the signed split attribution embedding space using a weighted mixture model; and then, using the initial prototypes as fixed anchors, perform dynamic programming phase allocation with the time-series hard constraints on all valid triggers, wherein the allocation cost is a confidence-weighted average of the distance from the trigger to the candidate prototype in the signed split attribution embedding space.

[0020] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena proposed in this invention is completely decoupled from trigger model types, possessing broad applicability. The progressive time stripping mechanism, signed splitting attribution embedding method, constrained clustering algorithm, and conditional Markov path analysis defined in this invention are not bound to specific trigger moment identifier types. As long as the selected trigger moment identifier can meet the minimum interface of "outputting the first trigger moment of the target phenomenon for a single-shot diagnostic timing sequence," regardless of whether it is a threshold rule model based on physical criteria, a traditional machine learning classifier, a deep neural network, a multi-instance learning model, or a timing anomaly / change point detection model, it can be directly integrated into the framework of this invention without modifying any module structure. This significantly reduces the engineering cost of building a precursor knowledge base on different fusion devices, different target phenomena, or existing operating systems, achieving cross-model, cross-device, and cross-phenomenon applicability of the method framework.

[0021] 2. The proposed automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena extends the "one shot, one moment" approach to a "one shot, one chain" approach through adaptive time stripping. This invention discards the trigger moment and subsequent time window identified in the previous round for each shot carrying the target phenomenon tag, forcing the initial trigger moment identifier to make judgments only from earlier and weaker precursor signals in the next round. This mechanism distributes the limited attention of a single model to different time scales in a "relay-style" manner, expanding the model's original output of only a "single trigger moment at the end" to a complete event chain that the framework can systematically mine, covering multi-level precursor events from tens to hundreds of milliseconds ago. In particular, this invention employs a single-shot adaptive stripping depth, where the stripping depth of each shot in the next round is independently determined by its own previous trigger moment, rather than uniformly advancing a fixed step size for all shots. This effectively avoids the drawbacks of global uniform stripping causing premature termination for long shots and degradation of training samples for short shots, while ensuring the physically traceable continuity of the trigger trajectory of the same shot across multiple iterations, thus protecting the temporal consistency of the event chain.

[0022] 3. The proposed automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena uses signed decomposition attribution embedding to preserve physical direction information and improve distinguishability. At the attribution layer, this invention adapts corresponding interpretability methods for different types of trigger moment identifiers (gradient-type SHAP or integral gradient for differentiable models, kernel SHAP, LIME, or permutation importance for black-box models, and criterion contribution decomposition for threshold rule models, etc.), uniformly outputting signed attribution contributions. Based on this, this invention innovatively decomposes the signed attribution value of each physical feature into two independent non-negative coordinates according to positive and negative components, constructing a signed decomposition attribution embedding vector. This embedding method enables subsequent clustering spaces to clearly distinguish between situations where "an increase in a physical quantity drives the phenomenon" and "a decrease in a physical quantity drives the phenomenon," which are opposite in direction but similar in absolute value. Traditional absolute value clustering methods lose this crucial physical information. This feature significantly improves the physical distinguishability of precursor attributions without increasing the parameters of the trigger moment identifier model.

[0023] 4. The proposed automatic discovery method for precursor event chains of magnetic confinement plasma discharge significantly improves the physical nomenclature and consistency of precursor stages through two-stage constrained clustering. First, this invention uses effective triggers with confidence levels higher than the population median as anchor sets. Within a signed split attribution embedding space, a weighted mixture model is used to estimate initial prototypes, avoiding low-confidence noise samples contaminating cluster centers. Then, using these prototypes as fixed anchors, dynamic programming stage allocation with hard constraints is performed on all effective triggers. The hard constraints include: the stage label sequence of each trigger within the same shot must be monotonically non-decreasing on the time axis (ensuring no time reversal), and the time distance between two triggers within the same shot and classified into the same stage must not exceed a preset upper bound (ensuring temporal locality of events). These two types of hard constraints fundamentally prevent unreasonable results such as events being out of order on the time axis and two events within the same shot and classified into the same cluster being too far apart. Simultaneously, the confidence-weighted mechanism gives higher-confidence triggers greater weight in clustering, further improving the robustness of stage division.

[0024] 5. The proposed automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena uses conditional Markov chains to reveal the divergence and mutual exclusion of evolutionary paths within the same target phenomenon. After obtaining the event chains of all shots carrying the target phenomenon tag, this invention first statistically analyzes and normalizes the transition frequencies between adjacent stages to construct a first-order Markov transition probability matrix. Based on this, for key intermediate stages exhibiting multi-path divergence, the methods are divided into several conditional sub-states according to their preceding stages, and the conditional transition probabilities are re-estimated to construct a conditional Markov chain. This conditional Markov chain can quantitatively answer the question, "Under different upstream histories, which path is a certain intermediate stage more likely to turn to?", thereby revealing the divergence positions and mutual exclusion relationships of different evolutionary paths within the same target phenomenon.

[0025] 6. The proposed automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena enhances the robustness of precursor triggering through cross-round permutation significance testing and multi-round confirmation. For triggers generated in all rounds, except for those in protection rounds or rounds with high classification indices which are directly retained, a time-axis-based random permutation significance test is applied to the triggers in the remaining rounds, retaining only those triggers with significant results. Simultaneously, false positive triggers appearing on shots without target phenomenon labels in all rounds undergo cross-round nearest neighbor merging, and unstable triggers are screened based on multi-round confirmation results. This filtering mechanism ensures that the precursor triggers ultimately entering the cluster do not depend on the random fluctuations of a single round or a single trigger moment identifier, significantly reducing the sensitivity of the event chain construction process to the randomness of a single inference and improving the output stability of the entire framework.

[0026] 7. The overall framework of the automatic discovery method for precursor event chains of magnetic confinement plasma discharge proposed in this invention can be automatically completed on a dataset with only gun-level labels. Starting from the original multi-channel time series data, this invention sequentially completes the initial trigger detection, progressive time stripping, cross-cycle trigger sorting, attribution embedding construction, stage clustering and path analysis, without any manual labeling of precursor events.

[0027] 8. The automatic discovery method for precursor event chains of magnetic confinement plasma discharge proposed in this invention can seamlessly interface with existing triggers in operating systems. Any mature existing triggers, such as rupture alarm thresholds, mode-lock alarm thresholds, and MARFE alarm thresholds, can be used as trigger moment identifiers in this invention. Therefore, this invention can upgrade the "single trigger" output of existing trigger systems to operational decision support with precursor event chains and evolution paths without replacing them, thereby improving the intelligence level of existing fusion device operation and control systems at a relatively low modification cost.

[0028] 9. The automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena proposed in this invention is compatible with multi-class target labels and is flexibly expandable. This invention is not only applicable to binary target labels (such as "whether a rupture has occurred"), but can also be extended to multi-class cases. When the target phenomenon label is multi-class, the steps of this invention can be executed separately for each type of target label, or aggregated separately according to multiple conditions on a shared precursor trigger set, thereby simultaneously discovering the precursor chains of multiple related target phenomena and their interrelationships. This scalability enables this invention to adapt to the more complex physical analysis needs of fusion devices.

[0029] Overall, this invention systematically solves several bottleneck problems in the automatic discovery of precursor event chains of magnetic confinement plasma discharge phenomena through steps such as progressive adaptive time stripping, signed splitting attribution embedding, confidence-weighted hard-constraint clustering, and conditional Markov path analysis. It has high practical value and promising prospects for industrial application. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the automatic discovery method for precursor event chains of magnetically confined plasma discharge phenomena according to the present invention.

[0031] Figure 2 This is a flowchart of the automatic discovery method for precursor event chains of magnetic confinement plasma discharge phenomena according to the present invention.

[0032] Figure 3 This is a schematic diagram illustrating the coverage time extension of target phenomenon precursor identification before and after multiple rounds of progressive stripping in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] Example 1 like Figure 1-3 As shown, this invention proposes an automatic detection method for precursor event chains of magnetically confined plasma discharge phenomena, comprising: Step S1: Obtain any model that can output the first trigger time of the target phenomenon in the single-shot diagnostic timing sequence as a trigger time identifier, and call the trigger time identifier on the complete single-shot diagnostic timing sequence without time stripping to obtain the first trigger time; Step S2: Starting from the second round, a progressive multi-round time stripping iteration with a stop condition determination is executed. Each round of iteration includes: for each gun carrying a target phenomenon tag, the time window from the trigger time detected in the previous round to the end of the discharge is discarded as the boundary, forming a stripped time sequence; the trigger time identifier is called on the stripped time sequence to obtain the new trigger time of the gun in this round; wherein, the stripping depth of different guns is independently determined by their respective trigger times detected in the previous round. Step S3: Validity filtering is performed on the trigger times obtained in all rounds, and for each valid trigger, interpretability attribution is performed in accordance with the type of trigger time identifier from which it originates, to obtain a signed attribution value. Then, the signed attribution value is split into two independent non-negative coordinates according to the positive and negative components to construct a signed split attribution embedding vector. Step S4: Based on the signed split attribution embedding vector, perform confidence-weighted clustering and integrate time-series hard constraints on all valid triggers to divide the triggers into different precursor stages. The time-series hard constraints include: the stage label sequence of each trigger in the same shot must be monotonically non-decreasing on the time axis, and the time distance between two triggers in the same shot that are classified into the same stage must not exceed a preset upper bound. Step S5: According to the occurrence time sequence of each effective trigger within each shot, the continuously repeating stage labels are concatenated and merged to obtain the event chain of that shot. Based on the event chain statistics of all shots carrying target phenomenon labels, a conditional Markov chain representing the divergence of different evolution paths is constructed.

[0035] This embodiment uses a discharge database (containing thousands of plasma discharges from each shot, with several physical channels collected at a preset sampling rate and accompanied by shot-level tags) of a typical magnetic confinement fusion device (including but not limited to tokamak, stellarator, and reverse field pinch device) as an example. The shot-level tags can be, but are not limited to, any one or more binary or multi-class tags such as "whether a fracture occurred," "whether mode locking occurred," "whether a MARFE occurred," "whether H-mode was entered," or "whether a given type of ELM appeared." The specific process is as follows: Step A: Trigger Timing Recognizer Access and Data Preparation. Select the trigger timing recognizer for this framework. The trigger timing recognizer must meet the minimum interface requirement of "outputting the first trigger timing of the target phenomenon for a single shot diagnostic timing sequence." Divide the discharge dataset into training, validation, and test sets according to a preset ratio (if the selected trigger timing recognizer is a pre-trained fixed model, the entire set can be used directly as the inference set). Truncate and balance all shot sequences without target phenomenon labels according to the visible length distribution of shots with target phenomenon labels. This step eliminates the possibility of sequence length becoming a shortcut for trigger timing recognizer classification and completes unified input interface adaptation.

[0036] Step B, First-round trigger time identifier detection. On the complete sequence without any time window discarding, the trigger time identifier is invoked, outputting the first-round trigger time for each shot carrying a target phenomenon tag; information such as trigger location, round number, detection signal type, time difference from the target phenomenon occurrence time, and source model handle is written into the trigger time identifier record table. This step obtains the final trigger set under the condition that no time information has been discarded.

[0037] Step C: Adaptive peel depth calculation for each shot and preparation of data for the next round. For each shot carrying a target phenomenon label, based on the trigger token position detected in the previous round, an upper bound is taken towards an earlier time step. Combined with the existing peel depth of that shot, the maximum value of the two is taken as the peel depth for the next round. The inference input for the next round is constructed according to the independent peel depth of each shot, so that each shot is not visible to the triggers and their subsequent windows discovered in the previous round in the next round. Shots without target phenomenon labels are never discarded in time windows, but participate in the re-matching of length. This step can form a non-overlapping time window peel hierarchy across multiple rounds.

[0038] Step D, the trigger moment identifier invocation protocol. Each round uses the following invocation method depending on the type of the selected trigger moment identifier: If the selected trigger moment identifier is a pre-trained model with fixed parameters, forward inference is performed on the stripped input of that round; if the selected trigger moment identifier allows retraining, a two-step training protocol is used. The first step trains on standard augmented target phenomenon labeled shots and all non-target phenomenon labeled shots with a target loss suitable for the trigger moment identifier. The second step uses the high-confidence true positive samples from the training set of the first step as a new "clean" positive sample set. A "second chance" mechanism is enabled for persistent false positive samples near the decision boundary, temporarily retaining them for training without immediate removal, and removing them only when they consistently fall below the absolute threshold. If the selected trigger moment identifier is a threshold rule model, this step degenerates into re-evaluating the stripped input according to the rules. This step can stabilize the output quality of the trigger moment identifier in each round and is adapted to the type of the selected trigger moment identifier.

[0039] Step E: Determining the Stop Condition for Each Round. After each round, the classification performance (such as AUPRC, detection rate, false positive rate, etc.) of the classifier is evaluated at a fixed peel depth based on the metrics available to the classifier at the selected trigger time. The metrics of the baseline classifier, which uses only sequence length as input, are also evaluated at the current peel depth. Stop is triggered when the fixed peel depth metric falls below a preset lower limit, or the length baseline metric exceeds a preset upper limit, or the false positive rate exceeds a preset warning line for several consecutive rounds, or the remaining number of target-carrying phenomenon tags that can be further advanced falls below the lower limit. This step adaptively terminates the iteration, avoiding meaningless over-peeling.

[0040] Step F: Cross-round trigger timing identifier filtering and aggregation. For triggers generated in all rounds, they are filtered according to the classification index of their respective rounds as follows: triggers belonging to protection rounds or whose classification index exceeds a high threshold are directly retained; for triggers from other rounds, a time-axis-based random permutation significance test is applied, retaining only triggers with significant results; false positive triggers appearing on non-target phenomenon tag guns in all rounds undergo cross-round nearest neighbor merging and require multi-round confirmation. This step yields a robust, cross-round consistent set of effective precursor triggers.

[0041] Step G: Interpretable Attribution and Signed Split Embedding Construction. For each valid precursor trigger, the handle of its source round trigger time identifier is read from the trigger time identifier record table and loaded. An appropriate interpretable attribution method is selected based on the type of the trigger time identifier (gradient-based SHAP or integral gradient for differentiable deep models, kernel SHAP, LIME, or permutation importance for black-box classifiers, residual contribution-based decomposition for time-series anomaly / change point detection models, and criterion contribution decomposition for threshold rule models). Attribution is performed on the output of the trigger time identifier in the neighborhood of the trigger token to obtain the signed attribution contribution of the trigger on each physical feature. For the signed attribution value of each physical feature, its positive and negative parts are placed in two independent dimensions, and the global mean over all triggers is removed to obtain the signed split attribution embedding vector of the trigger. This step preserves physical direction information while ensuring the embedded coordinates are numerically non-negative, facilitating subsequent measurement.

[0042] Step H, Confidence Scoring and Prototype Learning. For each precursor trigger, a comprehensive confidence score is calculated based on its detection time, round, and signal type. Precursors with confidence scores higher than the population median are selected as the anchor set. Several initial prototypes are estimated within the signed split attribution embedding space using a weighted mixture model. The number of prototypes is selected based on the partition scores within a preset candidate range (considering signature strength, maximum cluster ratio, profile coefficient, long-range collision rate, and phase reversal rate). This step prevents the prototypes from being contaminated by low-confidence noise.

[0043] Step I, Constrained Dynamic Programming Stage Allocation. Using the prototypes obtained in the previous step as fixed anchors, perform dynamic programming stage allocation with hard constraints on all precursor triggers: the allocation cost is a confidence-weighted average of the distances from the precursor trigger to the candidate prototype within the signed split attribution embedding space; the hard constraints include that the sequence of stages within the same shot must be monotonically non-decreasing, and the token distance between two precursor triggers assigned to the same stage within the same shot must not exceed a preset upper bound; finally, output a stage label for each precursor trigger. This step yields a set of temporally monotonically monotonically and physically nameable precursor stages.

[0044] Step J, Event Chain Construction. For each shot, its stage tags are concatenated in chronological order of the occurrence of its effective precursor triggers, and consecutively repeating stages are merged and deduplicated to obtain the event chain for that shot. The event chains of all shots carrying target phenomenon tags are statistically analyzed to obtain the event chain pattern distribution. The stages are merged according to their parent category (named according to the target phenomenon, for example, rupture can be divided into early / magnetohydrodynamic instability / final) to obtain the parent category event chain distribution. This step yields a readable "chain spectrum" for the target phenomenon database.

[0045] Step K involves first-order and conditional Markov chain analysis. The transition frequencies of adjacent stages in the entire event chain are statistically analyzed and normalized to a first-order Markov transition probability matrix. For key intermediate stages exhibiting multi-path divergence, they are broken down into several conditional sub-states based on their preceding stages. The conditional transition probabilities are re-estimated, and a conditional Markov chain is constructed. A heatmap and directed graph of the transition probabilities are output. This step can quantitatively present the divergence positions and mutual exclusion relationships of different evolutionary paths within the same target phenomenon.

[0046] Step L, Physically Interpretable Summary and Backtracking. For each stage, output its internal confidence mean, signature strength, core signature strength, percentage, typical time position, and the first few significant physical features and their directions, and write them into the physical interpretation file; for each shot, output its ordered stage sequence, stage time, stage confidence, and source round, facilitating backtracking by shot. This step can construct an event chain knowledge base that can be directly reviewed by physicists.

[0047] Example 2 In one embodiment of the present invention, the trigger timing identifier is a threshold rule-based trigger timing identifier: when the device has already deployed a mature rupture alarm threshold (e.g., a hard trigger based on mode-locked amplitude exceeding the threshold, radiation ratio rise exceeding the threshold, etc.), the existing trigger is connected as the trigger timing identifier of the present invention, and step D degenerates into re-evaluating the rule on the stripped data; the interpretability attribution in step G is performed according to the criterion contribution decomposition (i.e., the relative amplitudes of multiple criterion triggers are used as the attribution vector); the remaining steps are exactly the same as in embodiment 1. Therefore, the present invention can upgrade the "single alarm" output to operational decision support with a precursor event chain and evolution path without replacing the existing rupture alarm system.

[0048] Example 3 In one embodiment of the present invention, the trigger moment recognizer is a deep multi-instance learning trigger moment recognizer: when the trigger moment recognizer is a bidirectional Transformer multi-instance learning model with a local classification head and a global attention head, step D adopts a two-step training protocol; the interpretability attribution of step G is performed on the local classification head output of the model using kernel SHAP; the remaining steps are exactly the same as in embodiment 1.

[0049] Example 4 In one embodiment of the present invention, the trigger moment identifier is a change point detection type trigger moment identifier: when the trigger moment identifier is an unsupervised model based on Bayesian online change point detection or sliding window CUSUM, step D degenerates into rerunning the change point detection algorithm on the stripped data; the interpretability attribution in step G is based on the mean / variance difference of each physical feature before and after the change point as a signed attribution vector; the remaining steps are exactly the same as in embodiment 1.

[0050] Example 5 In one embodiment of the present invention, when the target phenomenon is mode locking, the gun-level label is "whether mode locking has occurred," and the remaining steps are exactly the same as in Embodiment 1. The event chain output by the present invention typically manifests as a three-stage chain of "increased tear mode amplitude → decreased mode frequency → phase locking," and the divergence probability of "increased tear mode amplitude" turning into "phase locking" or "self-healing" under different upstream parameter conditions is given through a conditional Markov chain. When the target phenomenon is L–H transition, the gun-level label is "whether entering H mode," and the event chain output by the present invention typically manifests as a three-stage chain of "decreased edge Dα → increased density → confirmed constraint improvement," and the two L–H transition modes of "gradual increase" and "rapid jump" can be distinguished at the conditional Markov level.

[0051] Example 6 This invention also proposes an automatic detection system for precursor event chains of magnetically confined plasma discharge phenomena, comprising: The trigger time identifier module is used to output the first trigger time of the target phenomenon in the single-shot diagnostic timing sequence, and to call the trigger time identifier on the complete single-shot diagnostic timing sequence without time stripping to obtain the first trigger time. The progressive multi-round time stripping module is used to, starting from the second round, discard the time window from the triggering time detected in the previous round to the end of the discharge for each gun carrying the target phenomenon tag, forming a stripped time sequence, and calling the triggering time identifier module on the stripped time sequence to obtain the new triggering time of the gun in this round; wherein, the stripping depth of different guns is independently determined by their respective triggering times detected in the previous round. The attribution and embedding module is used to filter the validity of the trigger times obtained in all rounds, and for each valid trigger, to perform interpretable attribution that is compatible with the type of the trigger time identifier module, to obtain a signed attribution value, and then to split the signed attribution value into two independent non-negative coordinates according to the positive and negative components, and to construct a signed split attribution embedding vector. The constrained clustering module is used to perform confidence-weighted clustering and integrate temporal hard constraints on all valid triggers based on the signed split attribution embedding vector, and to divide the triggers into different precursor stages. The temporal hard constraints include: the stage label sequence of each trigger in the same shot must be monotonically non-decreasing on the time axis, and the time distance between two triggers in the same shot that are classified into the same stage must not exceed a preset upper bound. The event chain and path analysis module is used to chain and merge continuously repeating stage tags according to the occurrence time sequence of each effective trigger within each shot to obtain the event chain of that shot, and to construct a conditional Markov chain representing the divergence of different evolution paths based on the event chain statistics of all shots carrying target phenomenon tags.

[0052] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena, characterized in that, include, Step S1: Obtain any model that can output the first trigger time of the target phenomenon in the single-shot diagnostic timing sequence as a trigger time identifier, and call the trigger time identifier on the complete single-shot diagnostic timing sequence without time stripping to obtain the first trigger time; Step S2: Starting from the second round, a progressive multi-round time stripping iteration with a stop condition determination is executed. Each round of iteration includes: for each gun carrying a target phenomenon tag, the time window from the trigger time detected in the previous round to the end of the discharge is discarded as the boundary, forming a stripped time sequence; the trigger time identifier is called on the stripped time sequence to obtain the new trigger time of the gun in this round; wherein, the stripping depth of different guns is independently determined by their respective trigger times detected in the previous round; the stripping depth of each gun carrying a target phenomenon tag is determined by taking the upper bound of the trigger time detected in the previous round and the maximum value of the historical stripping depth of the gun; guns without target phenomenon tags are not stripped, but are re-trunculated and balanced according to the visible length distribution of guns carrying target phenomenon tags in the corresponding round; Step S3: Validity filtering is performed on the trigger times obtained in all rounds, and for each valid trigger, interpretability attribution is performed in accordance with the type of trigger time identifier from which it originates, to obtain a signed attribution value. Then, the signed attribution value is split into two independent non-negative coordinates according to the positive and negative components to construct a signed split attribution embedding vector. Step S4: Based on the signed split attribution embedding vector, perform confidence-weighted clustering and integrate time-series hard constraints on all valid triggers to divide the triggers into different precursor stages. The time-series hard constraints include: the stage label sequence of each trigger in the same shot must be monotonically non-decreasing on the time axis, and the time distance between two triggers in the same shot that are classified into the same stage must not exceed a preset upper bound. Clustering is performed in two stages: In the first stage, effective triggers with confidence levels higher than the group median are taken as anchor sets, and several initial prototypes are estimated in the signed split attribution embedding space using a weighted mixture model. In the second stage, the initial prototype is used as a fixed anchor, and dynamic programming phase allocation with the temporal hard constraints is performed on all valid triggers, wherein the allocation cost is the confidence weight of the distance from the trigger to the candidate prototype in the signed split attribution embedding space. Step S5: According to the occurrence time sequence of each effective trigger within each shot, the continuously repeating stage labels are concatenated and merged to obtain the event chain of that shot. Based on the event chain statistics of all shots carrying target phenomenon labels, a conditional Markov chain representing the divergence of different evolution paths is constructed.

2. The method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena according to claim 1, characterized in that, The progressive multi-round time stripping iteration stops when any of the following conditions are met: The classification index or detection rate at a fixed peeling depth is lower than the preset lower limit, the length baseline classification index is higher than the preset upper limit, the false positive rate for several consecutive rounds is higher than the preset warning line, or the number of remaining propulsive target-carrying phenomena tag guns is lower than the preset lower limit.

3. The method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena according to claim 1, characterized in that, The validity filtering in step S3 includes: Triggers or protection rounds whose classification indicators exceed the high threshold are directly retained. Apply a time-axis-based random permutation significance test to the triggers in the remaining rounds, and retain only the triggers that show significant results; For false positive triggers occurring on guns without target phenomenon labels in all rounds, perform cross-round nearest neighbor merging and perform unstable trigger screening based on multi-round confirmation results.

4. The method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena according to claim 1, characterized in that, The confidence weighting in step S4 is calculated by combining one or more of the following factors: the earlier the trigger time is relative to the occurrence time of the target phenomenon, the higher the weight; the earlier the round in which the trigger time is located, the higher the weight; and the weight of the detection signal type at the trigger time, which is a forward residual peak or a signal based on the amount of evolutionary mutation, is higher than that of a signal based on the instantaneous absolute quantity.

5. The method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena according to claim 1, characterized in that, The construction of the conditional Markov chain in step S5 includes: First, the transition frequencies of adjacent stages in the entire event chain are counted and normalized to obtain the first-order Markov transition probability matrix. Then, for the key intermediate stages that exhibit multi-path divergence, they are divided into several conditional sub-states according to their preceding stages, and the conditional transition probabilities are re-estimated to obtain the conditional Markov chain.

6. The method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena according to claim 1, characterized in that, The trigger time identifier can be a single model or a combination of multiple models. When the trigger time identifier is a combination of multiple models, the trigger times given by each model are merged according to time proximity or the multiple models are confirmed and filtered before proceeding to step S3.

7. The method for automatically detecting precursor event chains of magnetically confined plasma discharge phenomena according to claim 1, characterized in that, The target phenomenon label is a binary label or a multi-class label; when the target phenomenon label is a multi-class label, steps S1 to S5 are executed separately for each type of target label, or aggregated separately for multiple conditions on the shared precursor trigger set.

8. An automatic detection system for precursor event chains of magnetically confined plasma discharge phenomena, characterized in that, include: The trigger time identifier module is used to output the first trigger time of the target phenomenon in the single-shot diagnostic timing sequence, and to call the trigger time identifier on the complete single-shot diagnostic timing sequence without time stripping to obtain the first trigger time. A progressive multi-round time stripping module is used, starting from the second round, to discard the time window from the trigger time detected in the previous round to the end of the discharge for each gun carrying a target phenomenon tag, forming a stripped time sequence. The trigger time identifier module is then called on the stripped time sequence to obtain the new trigger time of the gun in this round. The stripping depth of different guns is independently determined by their respective trigger times detected in the previous round. The stripping depth of each gun carrying a target phenomenon tag is determined by taking the upper bound of the trigger time detected in the previous round and the maximum value of the historical stripping depth of the gun. Guns without target phenomenon tags are not stripped, but are re-trunculated and balanced according to the visible length distribution of guns carrying target phenomenon tags in the corresponding round. The attribution and embedding module is used to filter the validity of the trigger times obtained in all rounds, and for each valid trigger, to perform interpretable attribution that is compatible with the type of the trigger time identifier module, to obtain a signed attribution value, and then to split the signed attribution value into two independent non-negative coordinates according to the positive and negative components, and to construct a signed split attribution embedding vector. The constrained clustering module is used to perform confidence-weighted clustering and incorporate temporal hard constraints on all valid triggers based on the signed split attribution embedding vector, classifying the triggers into different precursor stages. The temporal hard constraints include: the stage label sequence of each trigger within the same shot must be monotonically non-decreasing on the time axis, and the time distance between two triggers classified into the same stage within the same shot must not exceed a preset upper bound. Clustering is performed in a two-stage manner: In the first stage, valid triggers with confidence scores higher than the population median are taken as the anchor set, and clustering is performed in the signed split attribution embedding space by weighting confidence scores. The weighted hybrid model estimates several initial prototypes; in the second stage, using the initial prototypes as fixed anchors, dynamic programming stage allocation with the aforementioned temporal hard constraints is performed on all valid triggers, where the allocation cost is a confidence-weighted average of the distance from the trigger to the candidate prototype in the signed split attribution embedding space; the event chain and path analysis module is used to concatenate and merge continuously repeating stage labels according to the occurrence time sequence of each valid trigger within each shot to obtain the event chain of that shot, and to construct a conditional Markov chain representing the divergence of different evolutionary paths based on the event chain statistics of all shots carrying target phenomenon labels.

Images