A nuclear security digital twin drilling and decision-making method based on real combat

Abstract

The present application relates to the technical field of digital twin modeling, dynamic simulation and computer data processing, in particular to a nuclear security digital twin exercise and decision-making method based on real combat. The method collects personnel action, command transmission, region state and disposal result data, encodes them into standard event units and constructs a multi-layer time sequence event graph; a predicted event sequence is generated by a digital twin model, aligned with the real combat event sequence, and the occurrence order, duration interval, state transition and upstream and downstream consistency residual are obtained; according to the residual propagation range, the local state transition relationship or cross-chain association relationship is calibrated layer by layer; based on the calibrated model and adjacent event constraint, the unobserved intermediate state is inversed, and a virtual exercise branch is established. The method can inhibit the diffusion of local bias, keep the twin state and the real combat process in continuous correspondence, and provide a traceable state basis for missing state completion and subsequent deduction.

Description

Technical Field

[0001] This invention relates to the fields of digital twin modeling, dynamic simulation and computer data processing technology, specifically to a nuclear security digital twin exercise and decision-making method based on live-fire confrontation. Background Technology

[0002] Digital twin drills for nuclear facility security typically use the nuclear facility area, security personnel, response tasks, and event outcomes as modeling objects. Pre-defined scenario parameters, personnel action rules, command flow rules, and area state transition rules are written into the simulation model. The model then runs according to a pre-programmed threat script, obtaining process data such as personnel arrival, command execution, area containment, and response completion. During live-fire drills, the simulation process is displayed through drill records, action trajectories, command logs, and status reports. Some solutions also import data generated from live-fire drills into the digital twin scenario to reproduce the sequence of live-fire actions or update the state of objects in the scenario. The focus of these solutions is the synchronous representation of the real drill process in the virtual scenario. Data processing typically uses timestamps, personnel identifiers, and task numbers as association criteria, and pre-defined rules merge data from different sources onto the same drill timeline, thus forming a mapping relationship between the live-fire process and the virtual scenario.

[0003] Existing nuclear security digital twin systems often use trajectory points, state values, or task results as the smallest data objects when processing drill events. Personnel actions, command transmissions, area state changes, and response handling are usually stored and processed separately. Different data are associated with each other based on temporal proximity or manually set correspondences. When the recorded times are inconsistent, the duration of events overlaps, or intermediate states are missing, the system usually completes alignment through time interpolation, nearest neighbor matching, or fixed priority rules. The alignment results are then input into the digital twin model for playback. When there are differences between the model output and the actual results, the overall error, average delay, or result consistency rate are often used as comparison indicators. Then, technicians adjust the behavioral parameters, response time, or state transition conditions. This approach can complete basic drill reproduction, but it does not subdivide the dependencies between events, cross-chain constraints, and the propagation range of errors in different state chains, nor does it form an event-level residual structure that can be called for local calibration.

[0004] In terms of model calibration, conventional techniques typically involve offline fitting of the digital twin model based on a batch of historical exercise data, or uniformly correcting model parameters according to preset evaluation indicators after a single exercise. The correction targets generally include personnel speed, task time, response latency, and state transition probability. Some solutions adopt a rolling update approach, recalculating model parameters and overwriting the original parameters when new data arrives. Other solutions compare different parameter combinations by setting up multiple simulation scenarios, and then select the parameter group with the smallest difference from the actual results as the subsequent model version. All of the above solutions take the overall output being close to the actual results as the calibration target, without distinguishing whether the deviation comes from event time misalignment, local state omission, behavioral rule mismatch, or cross-chain relationship error, nor do they limit the correction to the local state transition relationship or the correlation between multiple sub-models. Therefore, local abnormal data may cause irrelevant rules to change synchronously, and subsequent models may find it difficult to maintain the stability within the verified range.

[0005] The lack of a two-way calibration mechanism between the live-fire exercise and the digital twin model, which focuses on event semantics and residual propagation range, means that live-fire data is typically only used to drive scene display or overall parameter updates. The digital twin model cannot determine the specific source of inconsistencies between predicted events and live-fire events based on the relationships between personnel action chains, command transmission chains, regional state chains, and response and handling chains. It also cannot select appropriate local or joint correction methods based on whether the residual is limited to a single event chain or spans multiple event chains. When time recording deviations, missing intermediate states, or local behavioral changes occur during the exercise, the model is prone to propagating local deviations to irrelevant state transition rules, causing the digital twin state to continuously deviate from the real process as the exercise progresses. Consequently, the unobserved states inverted based on the current model and the virtual exercise branches extended from the current state lack a continuous, reliable, and traceable state foundation for calibration. Summary of the Invention

[0006] The purpose of this invention is to provide a nuclear security digital twin exercise and decision-making method based on live-fire confrontation, which can solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] A method for nuclear security digital twin drills and decision-making based on live-fire confrontation includes: collecting data on personnel actions, command transmission, area status and handling results in live-fire confrontation, and encoding them into standard event units containing event subject, target, time, pre-state and execution result;

[0009] Based on the standard event unit, a multi-layered time-series event diagram is constructed, consisting of personnel action chain, instruction transmission chain, area status chain, and response and handling chain;

[0010] The digital twin model generates a predicted event sequence based on the multi-layer time-series event graph. The predicted event sequence is then time-aligned with the actual event sequence to obtain the occurrence order, duration interval, state transition, and upstream-downstream consistency residual.

[0011] Based on the propagation range of various residuals in the multi-layer time-series event graph, local state transition relationships or cross-chain association relationships are selected for hierarchical calibration;

[0012] Based on the calibrated digital twin model and adjacent event constraints, the unobserved intermediate state is inverted, the inversion results and uncertainties are written into the multi-layer time series event graph, and a virtual exercise branch is established with the calibrated current state.

[0013] Preferably, the standard event unit further includes an event source identifier, a time confidence interval, a state confidence level, and an associated evidence index;

[0014] When constructing the multi-layer time-series event graph, directed edges within the layer are established according to the continuous action relationship of the same event subject, the instruction sending and receiving relationship, the successive relationship of the regional state, and the correspondence between the handling action and the result. Inter-layer constraint edges are established based on the shared fields of the same standard event unit in different event chains.

[0015] For multiple standard event units with inconsistent source times, their original timestamps are retained, and relative order constraints are generated using common preceding events and common succeeding events. The relative order constraints are written into the directed edges in the corresponding layer, and parallel event clusters are formed for events whose unique sequential relationship cannot be determined.

[0016] The digital twin model generates predicted event sequences using the parallel event clusters as branch nodes, and merges branches consistent with the actual event sequences when subsequent live-fire events arrive.

[0017] Preferably, aligning the predicted event sequence with the live-fire event sequence in time includes: constructing a candidate matching key using the event subject, the target of action, and the preceding state, and generating candidate event pairs from the predicted event sequence and the live-fire event sequence according to the candidate matching key;

[0018] Based on the sequential relationship of each candidate event pair in the corresponding event chain and the cross-chain relationship between different event chains, candidate event pairs that disrupt the existing event order relationship are eliminated.

[0019] For the remaining candidate event pairs, establish joint matching costs for occurrence time offset, continuous interval overlap, state transition consistency, and upstream and downstream adjacency consistency, and determine the correspondence according to the constraint of maintaining the continuity of the event chain;

[0020] For predicted events and live-fire events that do not form a corresponding relationship, they are marked as virtual redundant events and live-fire newly added events, respectively, and the correlation between the two in the multi-layer time series event graph is used as the component data of state transition residual and upstream and downstream consistency residual.

[0021] Preferably, the digital twin model is subjected to hierarchical calibration, including: dividing the digital twin model into scene state sub-model, personnel behavior sub-model, instruction coordination sub-model and result evolution sub-model, and establishing a mapping table between the correctable state variables, state transition rules and cross-sub-model input-output relationships of each sub-model;

[0022] Based on the event chain type, start node, and end node corresponding to various residuals, the state variables, state transition rules, or cross-sub-model input-output relationships that intersect with the residual propagation path are selected from the mapping table to form a candidate correction set;

[0023] Local calibration is performed according to the minimum coverage of the candidate correction set in the multi-layer time series event graph, and extended to cross-sub-model joint calibration when the same residuals persist across at least two event chains;

[0024] For each calibration, the model version before calibration, the calibration basis event, the correction object, the applicable event chain scope, and the model version after calibration are recorded to form a model version chain with inheritance relationship.

[0025] Preferably, when updating the multi-layer timing event graph, incremental consistency verification is performed on the newly arrived standard event unit, including: locating the local subgraph that shares the preceding or succeeding state with the newly arrived standard event unit based on the event subject and the target of action;

[0026] The reachability order of the newly arrived standard event unit is compared with that of the events in the local subgraph based on the time confidence interval, and the inter-layer constraint edge is called to determine the acceptable order set when the time confidence intervals overlap.

[0027] When the newly arrived standard event unit causes the directed edge in the existing layer to form a state rollback, repeated transfer or broken transfer, the directed edge in the corresponding layer is converted into a verifiable edge, and a local reconnection structure containing alternative preceding events and alternative succeeding events is generated around the verifiable edge.

[0028] Each local reconnection structure is input into the digital twin model and replayed within the event interval corresponding to the local subgraph. The local reconnection structure that simultaneously satisfies the subsequent events of the actual troops and the cross-chain constraints is retained, and the parallel event clusters and their branch nodes are updated.

[0029] Preferably, the object to be corrected is determined based on the propagation range of various residuals in the multi-layer time-series event graph, including: taking the candidate event pair that generates the residual as the starting node of the residual, and tracing the reachable nodes of the residual in the personnel action chain, instruction transmission chain, regional status chain and response and handling chain along the preceding and following associations of the corresponding event chain and the cross-chain associations between different event chains.

[0030] For each reachable node, record the first occurrence position of the residual, the consistency of the direction after the continuous maintenance interval and cross-chain mapping, and classify the residuals that are only located in a single event chain and disappear at the subsequent matching event into local state transition residuals.

[0031] Residuals that span different event chains and maintain the same offset across multiple subsequent matching events are classified as cross-chain associated residuals.

[0032] For the residuals caused by the newly added military event, compare the event chain connectivity before and after the addition of the newly added military event, and write the new association that eliminates the upstream and downstream consistency residuals after the addition into the state transition relationship or cross-chain association relationship to be corrected.

[0033] Preferably, when forming the candidate correction set, the basic correction range is first determined based on the shortest reachable subgraph between the residual start node and the residual end node, and then the correctable state variables and state transition rules that directly correspond to the state transition relationship within the basic correction range are extracted from the mapping table.

[0034] For the basic correction range that includes cross-chain associations, the cross-sub-model input-output relationship of the sub-models to which both ends of the cross-chain association belong is added to the candidate correction set;

[0035] For each correction object in the candidate correction set, a single-object calibration copy is generated while keeping other correction objects unchanged. The direction of residual change obtained by replaying each single-object calibration copy is compared with the original residual propagation direction, and correction objects that expand the original residual propagation range are eliminated.

[0036] The remaining correction objects are merged into joint correction groups according to the residual paths they share, and calibration is performed using the joint correction group that requires the fewest correction objects to cover all residual paths.

[0037] Preferably, when inverting unobserved intermediate states, event intervals located between adjacent observed events and lacking state records are extracted from the multi-layer time-series event graph. A forward candidate state sequence is generated based on the execution result of the previous observed event, and a reverse candidate state sequence is generated based on the preceding state of the next observed event. The intersection of the forward candidate state sequence and the reverse candidate state sequence is obtained while maintaining the reachability of each event chain and the consistency of cross-chain association.

[0038] For each candidate intermediate state in the intersection, the calibrated digital twin model is invoked to perform local replay. Based on the degree of matching between the local replay results and the adjacent observed events in terms of occurrence time, state transition, and upstream and downstream correlation, a state confidence value is generated.

[0039] The first candidate intermediate state is selected from high to low according to the state confidence value and added to the main event chain. The remaining candidate intermediate states are reserved as alternative branches, and the complementary value of the normalized state confidence value is used as the uncertainty.

[0040] Preferably, residual playback verification is performed on the model version after the cross-sub-model joint calibration, including: extracting the stable event interval before the joint calibration, the residual event interval that triggered the joint calibration, and the subsequent event interval after the residual termination from the model version chain;

[0041] The three event intervals were continuously replayed using the pre-calibration model version and the post-calibration model version to form the corresponding residual propagation trajectory;

[0042] Compare the changes of the two sets of residual propagation trajectories at the residual source node, cross-chain diffusion correlation and residual termination node, and when the calibrated model version generates new residuals in the stable event interval, split the joint correction group that causes the new residuals into multiple sub-correction groups.

[0043] The model is replayed sequentially according to the coverage relationship of each sub-correction group to the original residual path. The sub-correction group that does not change the stable event interval and covers the original residual path is selected and replaced with the corresponding calibrated model version in the model version chain.

[0044] Preferably, the virtual training branch is established in the calibrated state, including: encapsulating the observed state in the main event chain, the verified model version, and the alternative branches in the unobserved event interval as a branch start snapshot;

[0045] For each candidate branch, inheriting intermediate states and uncertainties, and using subsequent exercise event inputs, generate a parallel virtual event sequence;

[0046] When the parallel virtual event sequences meet the same preceding state, event order and cross-chain association at the same state node, the parallel virtual event sequences are merged into a shared successor branch, and their state confidence values ​​are normalized and summarized according to the complementary uncertainty values ​​of each candidate intermediate state.

[0047] When subsequent live-fire events arrive, they are sorted according to their consistency with each parallel virtual event sequence in terms of event subject, target, time, preceding state, and execution result. The virtual event sequence with the highest ranking is connected to the main event chain, and the remaining virtual event sequences are retained in the branch start snapshot.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] 1. By encoding personnel actions, command transmission, area status, and handling results in live-fire exercises into standard event units with event subjects, target objects, time, pre-state, and execution results, and constructing a multi-layered temporal event graph consisting of personnel action chains, command transmission chains, area status chains, and response and handling chains, exercise data from different sources can maintain sequential dependencies and cross-chain constraints under a unified event semantics. After temporal alignment between the predicted event sequence generated by the digital twin model and the live-fire event sequence, occurrence order, duration interval, state transition, and upstream-downstream consistency residuals are formed respectively. Then, based on the propagation range of the residuals in the multi-layered temporal event graph, local state transition relationships or cross-chain correlation relationships are selected for hierarchical calibration, ensuring that the correction object corresponds to the source of the deviation, limiting the spread of local anomalies to irrelevant model rules, and enabling the calibrated digital twin state to continuously fit the live-fire exercise process, providing a continuous current state for unobserved state inversion and virtual branch establishment.

[0050] 2. The time reliability interval, state reliability level, and associated evidence index in the standard event unit can be used to preserve the original time information when the time of multi-source records is inconsistent, and form a verifiable event sequence by combining common preceding events, common succeeding events, and inter-layer constraint edges; the incremental consistency verification and local reconnection structure enable newly arriving events to update the multi-layer time series event graph without disrupting the existing event chain. The model version chain records the calibration basis, correction object, and applicable scope, and compares the propagation trajectory of the model before and after calibration in the stable event interval, residual event interval, and succeeding event interval by residual replay, reducing the interference of new corrections on existing stable relationships; for event intervals lacking state records, candidate intermediate states and their uncertainties are determined by using forward candidate state sequences, reverse candidate state sequences, and local replay results, and then the observed states, verified model versions, and candidate branches are encapsulated as branch start snapshots, so that subsequent live events can be used to screen and continue the virtual event sequence consistent with the real process. Attached Figure Description

[0051] Figure 1 A flowchart of a nuclear security digital twin exercise and decision-making method based on live-fire confrontation;

[0052] Figure 2 Flowchart for standard event unit encoding, multi-layer timing event graph construction, and timing alignment of predicted events with live-fire events;

[0053] Figure 3 Flowchart for residual propagation range identification, candidate correction object determination, hierarchical calibration of digital twin model, and model version verification;

[0054] Figure 4 This is a flowchart for the inversion of unobserved intermediate states, the generation of virtual exercise branches, and the subsequent continuation of live-fire events. Detailed Implementation

[0055] In one embodiment, please refer to Figure 1 A method for nuclear security digital twin exercises and decision-making based on live-fire confrontation includes: collecting data on personnel actions, command transmission, area status, and handling results in live-fire confrontation, and encoding them into standard event units containing event subjects, targets, time, pre-states, and execution results; constructing a multi-layered time-series event graph based on the standard event units, consisting of personnel action chains, command transmission chains, area status chains, and response and handling chains; generating predicted event sequences from the digital twin model based on the multi-layered time-series event graph, aligning the predicted event sequences with the live-fire event sequences in time, and obtaining the occurrence order, duration interval, state transition, and upstream and downstream consistency residuals; selecting local state transition relationships or cross-chain correlation relationships for hierarchical calibration based on the propagation range of various residuals in the multi-layered time-series event graph; inverting unobserved intermediate states based on the calibrated digital twin model and adjacent event constraints, writing the inversion results and uncertainties into the multi-layered time-series event graph, and establishing virtual exercise branches with the calibrated current state.

[0056] In this embodiment, the live-fire confrontation data consists of exercise records, command logs, area status records, and handling result records. All types of records are converted into a unified data structure before entering the digital twin model. The event subject is used to identify the exercise object that implements the action or causes a state change. The target object is used to identify the area, task, or another event subject targeted by the action. The time includes the start time and the end time. The pre-event state represents the personnel state, command state, or area state that must be established before the event occurs. The execution result represents the state change formed after the event ends. During the conversion process, conflicting values ​​in the source data are not directly deleted. Instead, the original fields are retained and the source index is added so that subsequent time alignment can distinguish between differences in the actual state and differences in the records.

[0057] For details, please refer to Figure 2 The multi-layered time-series event graph uses standard event units as nodes, the preceding and following relationships within the same layer as directed edges within the layer, and the state dependencies between different event chains as inter-layer constraint edges. The personnel action chain records the continuous actions of personnel from receiving a task to completing the disposal. The instruction transmission chain records the transmission process of instruction generation, sending, receiving, confirmation, and execution. The area state chain records the state changes of an area from normal, restricted, and sealed to recovery. The response and disposal chain records the triggering, assignment, implementation, and result confirmation of disposal tasks. The digital twin model reads the current end node of each event chain and cross-chain constraints, generates predicted events that satisfy the preceding states, and appends the predicted events to the candidate successor positions of the corresponding event chains.

[0058] Furthermore, the temporal alignment between predicted event sequences and live event sequences does not use a single time difference as the matching basis. Instead, it simultaneously utilizes the event subject, the target, the preceding state, the occurrence interval, and the upstream and downstream relationships to establish candidate correspondences. When the same live event can correspond to multiple predicted events, the correspondence relationship that does not disrupt the existing order of the event chain and cross-chain dependencies is preserved. When multiple data sources provide different time records for the same event, a relative order is first formed based on common preceding events and common succeeding events. Then, the time offset is determined within the range of the relative order. This yields the occurrence order residual, the duration interval residual, the state transition residual, and the upstream and downstream consistency residual. Each type of residual is then associated with its starting node, ending node, and the inter-layer constraint edges it passes through.

[0059] ;

[0060] The above formula is used to calculate the joint matching cost between predicted events and actual military events, where, Let $\frac{a}{b}$ represent the matching cost between the predicted event $a$ and the actual military event $b$. This represents the normalized deviation after correction for relative order constraints at the time of occurrence. This represents the proportion of the non-overlapping portion of the continuous intervals to the union of the two intervals. This indicates the proportion of inconsistencies between the preceding state and the execution result out of the total number of state items. This indicates the proportion of inconsistent upstream and downstream adjacent events to the total number of adjacent events. to They are respectively represented as , , and The corresponding non-negative weights sum to 1, and in a calculation example, their values ​​are 0.25, 0.20, 0.35, and 0.20 respectively. When the four class biases are 0.08, 0.10, 0.00, and 0.05 respectively, If the value is 0.05, the candidate event pair is retained and entered into the event chain continuity check.

[0061] In a preferred embodiment, the standard event unit further includes an event source identifier, a time reliability interval, a state reliability level, and an associated evidence index. When constructing a multi-layer time-series event graph, directed edges within the layer are established according to the continuous action relationship of the same event subject, the instruction sending and receiving relationship, the successive relationship of regional states, and the correspondence between handling actions and results. Inter-layer constraint edges are established based on the shared fields of the same standard event unit in different event chains. For multiple standard event units with inconsistent source times, their original timestamps are retained, and relative order constraints are generated using common preceding events and common succeeding events. The relative order constraints are written into the directed edges within the corresponding layer, and parallel event clusters are formed for events whose unique sequential relationship cannot be determined. The digital twin model generates predicted event sequences with parallel event clusters as branch nodes, and merges branches consistent with the actual event sequence when subsequent actual events arrive.

[0062] The event source identifier is composed of data category, record batch, and original record location. The time credibility interval is determined by the original timestamp, record delay, and order of records from the same source. The state credibility level is classified according to whether the state is directly recorded, whether it is mutually verified by multiple sources, and whether it relies on inference. The associated evidence index points to the set of original records used to establish directed edges within the layer or constraint edges between layers. Shared fields include at least the event subject identifier, the target identifier, the task identifier, the preceding state identifier, and the execution result identifier. When establishing constraint edges between layers, it is required that there is a state inheritance relationship between the shared fields. For example, the task identifier contained in the execution confirmation event in the instruction transmission chain is consistent with the task start event in the personnel action chain, and the execution result of the former can satisfy the preceding state of the latter.

[0063] When the time reliability intervals of multiple events overlap and a unique order cannot be determined based on the records within the layer, the system starts from the common preceding node of each event, checks whether the candidate order causes subsequent states to regress or produces repeated state transitions, and checks in reverse from the common succeeding node whether the candidate order can satisfy the preceding state of the subsequent event. If there are still multiple acceptable orders, a parallel event cluster is established. Each order in the parallel event cluster corresponds to a local branch. The digital twin model only copies the local state adjacent to the parallel event cluster and does not copy the entire scene state. After the subsequent real-world events arrive, the local branches are filtered according to the event subject, the target, the preceding state, and the upstream and downstream relationships. The branches that satisfy the constraints of the subsequent events are merged into the main event chain, and the remaining branches are retained as historical candidate relationships.

[0064] In another embodiment, the predicted event sequence and the live event sequence are time-aligned, including: using the event subject, the target, and the preceding state as candidate matching keys, generating candidate event pairs from the predicted event sequence and the live event sequence according to the candidate matching keys; eliminating candidate event pairs that disrupt the existing event order relationship based on the preceding and following relationships of each candidate event pair in the corresponding event chain and the cross-chain relationships between different event chains; establishing joint matching costs for the remaining candidate event pairs based on occurrence time offset, continuous interval overlap, state transition consistency, and upstream and downstream adjacency consistency, and determining the correspondence relationship according to the constraint of maintaining the continuity of the event chain; for predicted events and live events that do not form a correspondence relationship, they are marked as virtual redundant events and newly added live events, respectively, and the preceding and following relationships of the two in the multi-layer time-series event graph are used as the component data of the state transition residual and the upstream and downstream consistency residual.

[0065] Candidate matching keys are generated using a hierarchical comparison method. Events with completely identical event subjects and target objects enter the direct candidate set. Events with identical event subjects but target objects belonging to the same superior region or the same task chain enter the extended candidate set. Preceding states are compared through state inclusion relationships. When the preceding state of the predicted event is a subset of the preceding state of the actual event, it is allowed to enter the candidate set, but the missing state item is recorded as a state transition residual. After candidate event pairs are formed, the relative order of any two candidate event pairs in the personnel action chain, command transmission chain, regional state chain, and response and handling chain is compared. If the matching relationship causes the predicted event to occur first to correspond to the actual event that occurs later and there is no parallel event cluster to support it, then the candidate event pair with the higher joint matching cost is deleted.

[0066] Event chain continuity constraints are achieved through local dynamic matching. The system takes the confirmed end-matching event of each event chain as the starting point, expands the prediction event window and the actual event window to a limited depth, calculates the joint matching cost of all candidate event pairs in the window, and requires that the selected correspondence maintains the reachability order of directed edges within the layer and the direction of the constraint edges between layers. When the window slides, the correspondence that has been supported by two or more subsequent matching events is frozen. For unmatched prediction events, if their subsequent prediction events can correspond to actual events, they are marked as virtual redundant events and their preceding and following associations are retained. For unmatched actual events, if their preceding and following actual events have been matched, they are marked as newly added actual events and inserted into the local position of the corresponding event chain.

[0067] In a preferred embodiment, please refer to Figure 3The digital twin model undergoes hierarchical calibration, including: dividing the digital twin model into scene state sub-models, personnel behavior sub-models, instruction coordination sub-models, and result evolution sub-models, and establishing a mapping table between the correctable state variables, state transition rules, and cross-sub-model input-output relationships of each sub-model; selecting state variables, state transition rules, or cross-sub-model input-output relationships that intersect with the residual propagation path from the mapping table to form a candidate correction set based on the event chain type, start node, and end node corresponding to various residuals; performing local calibration according to the minimum coverage of the candidate correction set in the multi-layer time-series event graph, and expanding to cross-sub-model joint calibration when the same type of residual continuously appears across at least two types of event chains; and recording the model version before calibration, the calibration basis event, the correction object, the applicable event chain range, and the model version after calibration for each calibration, forming a model version chain with inheritance relationships.

[0068] The scene state sub-model stores the region state and the reachability relationships between regions; the personnel behavior sub-model stores the personnel task state and action transition conditions; the instruction coordination sub-model stores the constraints between instruction generation, transmission, confirmation, and execution; and the result evolution sub-model stores the association between the disposal action and subsequent state changes. The mapping table uses the event chain node type as an index to map each node to the state variables and state transition rules that can generate or change that node, and records the input-output relationships across sub-models. For example, the execution confirmation state output by the instruction coordination sub-model serves as the input to the task start rule in the personnel behavior sub-model, and the disposal completion state output by the personnel behavior sub-model serves as the input to the region state update rule in the result evolution sub-model.

[0069] ;

[0070] The above formula is used to quantify the degree of propagation correlation between residuals and candidate correction objects, where, This represents the association value of the v-th candidate correction object. This represents the set of residual paths from the residual start node to the residual end node, passing through the correspondence of the v-th candidate correction object. This represents the weight of the q-th residual path determined by the event source credibility level. This represents the proportion of nodes where the residual changes in the same direction along the q-th path. This represents the ratio of the number of edges covered by the v-th candidate correction object in the q-th path to the total number of edges in that path. In a calculation example, a candidate correction object corresponds to 2 paths. They are 0.6 and 0.4 respectively. They are 1 and 0.5 respectively. If they are 0.5 and 0.25 respectively, then The value is 0.35, and the local calibration range is determined based on this value after sorting with other candidate correction objects.

[0071] Layered calibration uses the residual propagation path coverage as the boundary and does not directly modify state variables and rules that do not intersect with the residual path. During local calibration, the smallest set of objects that can cover all abnormal edges between the residual start node and the termination node is selected from the candidate correction set. A calibration copy is generated for each object and replayed in the corresponding event interval. If the residual is only located in a single event chain and terminates at the subsequent matching event, only the state transition relationship within the corresponding sub-model of that event chain is corrected. If the same-direction residual enters another event chain through the inter-layer constraint edge and continues in multiple subsequent matching events, the relevant cross-sub-model input-output relationship is included in the joint calibration. The model version chain uses the parent version identifier and incremental correction record to save the inheritance relationship, so that any calibration result can be traced back to the calibration basis event and the applicable event chain range.

[0072] In a preferred embodiment, when updating the multi-layer time-series event graph, incremental consistency verification is performed on newly arrived standard event units, including: locating local subgraphs that share the preceding or succeeding states with the newly arrived standard event unit based on the event subject and the target of action; comparing the reachability order of the newly arrived standard event unit with the events in the local subgraph based on the time confidence interval, and calling the inter-layer constraint edges to determine the acceptable order set when the time confidence intervals overlap; when the newly arrived standard event unit causes the existing layer-directed edges to form state rollback, repeated transition, or broken transition, converting the corresponding layer-directed edges into edges to be verified, and generating a local reconnection structure containing candidate preceding events and candidate succeeding events around the edges to be verified; inputting each local reconnection structure into the digital twin model, replaying it in the event interval corresponding to the local subgraph, retaining the local reconnection structure that simultaneously satisfies the subsequent events of the actual force and the cross-chain constraints, and updating the parallel event clusters and their branch nodes.

[0073] Incremental consistency verification only processes local subgraphs that share a state or a task identifier with newly arrived standard event units. The boundaries of the local subgraphs are determined by tracing back to the first confirmed matching event and tracing back to the first stable state event. State rollback means that the execution result of a new event causes the state to return to a historical state that has been excluded by subsequent events. Repeated transition means that the same subject repeatedly enters the same exclusive state without a recovery event. Break transition means that the preceding state of a new event cannot be satisfied by the execution result of any existing preceding event. When reconnecting the edge to be verified, we try to place the new event before, between, and after the original edge, and delete the structure that violates the inter-layer constraint direction. Then, the digital twin model performs local replay of the remaining structure.

[0074] Local replay takes the model state saved at the starting node of the local subgraph as the initial state, executes the state transition rules sequentially according to the event order in the local reconnection structure, and compares the events generated by the replay with the subsequent events of the actual troops after the local subgraph. The corresponding reconnection structure is retained only when the preceding state of the subsequent event can be satisfied, the main state of the event does not conflict, and the cross-chain association direction is consistent. When multiple reconnection structures satisfy the constraints, the differences are converted into parallel event clusters. Each reconnection structure shares the same preceding model state and subsequent stable state. When the subsequent new event can exclude some structures, only the branch relationship within the parallel event cluster is updated, and the event chain outside the local subgraph is not recalculated.

[0075] In another embodiment, the objects to be corrected are determined based on the propagation range of various residuals in the multi-layer time-series event graph. This includes: taking the candidate event pair that generates the residual as the starting node of the residual, and tracing the reachable nodes of the residual in the personnel action chain, instruction transmission chain, regional state chain, and response and handling chain along the preceding and following associations of the corresponding event chain and the cross-chain associations between different event chains; recording the first occurrence position of the residual for each reachable node, maintaining the consistency of the direction after the interval and cross-chain mapping, and classifying the residuals that are only located in a single event chain and disappear at the subsequent matching event as local state transition residuals; classifying the residuals that cross different event chains and maintain the same direction offset in multiple subsequent matching events as cross-chain association residuals; for residuals caused by new events added by actual troops, comparing the event chain connectivity before and after the addition of the new events, and writing the new associations that eliminate the upstream and downstream consistency residuals after the addition into the state transition relationship or cross-chain association relationship to be corrected.

[0076] Residual propagation tracing starts simultaneously from the prediction node and the actual node corresponding to the candidate event pair. The different state items in the execution results on both sides are used as the initial residual labels. The subsequent nodes are checked along the directed edges within the layer to see if they continue to depend on the state items. The state items are also checked along the inter-layer constraint edges to see if they are used as the preceding state by another event chain. If the direction of the state difference of the subsequent matching event is consistent with the initial residual, the continuous holding interval is extended. If the state difference disappears and the next two matching events do not reappear, the node is recorded as the residual termination node. The directional consistency after cross-chain mapping is determined based on whether the state difference is shown as early, delayed, missing, or added in both the source event chain and the target event chain. This distinguishes between local rule deviation and cross-chain input-output relationship deviation.

[0077] The processing of newly added events in actual combat is based on the connectivity of the event chains before and after insertion. When the execution result of the previous matched event cannot satisfy the prerequisite state of the next matched event before insertion, and the prerequisite state of the newly added event is satisfied by the previous event and its execution result can satisfy the next event after insertion, the state transition relationship corresponding to the newly added event is written into the set to be corrected. When the newly added event connects two nodes in different event chains at the same time and eliminates inter-layer constraint conflicts, the newly added association is written into the cross-chain association relationship to be corrected. If the newly added event only changes the time order and does not change the state reachability relationship, it is retained as the basis for time sequence correction and is not written into the state transition rules.

[0078] In a preferred embodiment, when forming a candidate correction set, the basic correction range is first determined based on the shortest reachable subgraph between the residual start node and the residual end node. Then, the correctable state variables and state transition rules directly corresponding to the state transition relationships within the basic correction range are extracted from the mapping table. For the basic correction range containing cross-chain associations, the cross-sub-model input-output relationships of the sub-models to which the two ends of the cross-chain association belong are added to the candidate correction set. For each correction object in the candidate correction set, a single-object calibration copy is generated while keeping other correction objects unchanged. The residual change direction obtained by replaying each single-object calibration copy is compared with the original residual propagation direction, and correction objects that expand the original residual propagation range are eliminated. The remaining correction objects are merged into a joint correction group according to the residual paths they commonly cover, and calibration is performed using the joint correction group that requires the fewest correction objects to cover all residual paths.

[0079] ;

[0080] The above formula is used to select the joint correction group to perform the calibration, where, Represents the set of candidate correction objects. Indicates the selected joint correction group, Indicates the number of objects within the joint correction group. This represents the number of new residuals generated by the joint correction group during replay of a stable event interval, normalized to the event number. This represents the number of cross-submodel input-output relations included in the joint correction group, normalized to the total number of candidate relations. This represents the set of residual paths covered by the v-th correction object. This represents the set of all residual paths to be eliminated. , and When the values ​​are 0.4, 0.4 and 0.2 respectively, if the three values ​​of the two feasible correction groups are 2, 0 and 0.5 and 3, 0.1 and 0.2 respectively, then the target values ​​are 0.9 and 1.28 respectively. The former correction group is selected for calibration.

[0081] Single-object calibration copies are created on the same model parent version. Each copy changes only one state variable, state transition rule, or cross-sub-model input-output relationship. The changes are derived from the corresponding state differences in the live event. For example, if the predicted event is earlier than the live event is later, the arrival time of the state referenced by the triggering rule is adjusted without changing the order of the preceding events. If the original residual is shortened after replay but spreads in the same direction in other event chains, the corresponding correction object is marked as expanding the propagation range and removed. The remaining correction objects are merged according to the intersection relationship of covering the same residual path. The joint correction group must cover all residual paths and not introduce state transitions that conflict with the stable event interval.

[0082] In a preferred embodiment, please refer to Figure 4 When inverting unobserved intermediate states, event intervals located between adjacent observed events and lacking state records are extracted from the multi-layer time-series event graph. A forward candidate state sequence is generated based on the execution result of the previous observed event, and a reverse candidate state sequence is generated based on the preceding state of the next observed event. The intersection of the forward and reverse candidate state sequences is calculated while maintaining the reachability of each event chain and the consistency of cross-chain associations. For each candidate intermediate state in the intersection, a calibrated digital twin model is invoked for local replay. A state confidence value is generated based on the degree of matching between the local replay result and the adjacent observed events in terms of occurrence time, state transition, and upstream and downstream associations. The first candidate intermediate state is selected from high to low according to the state confidence value and incorporated into the main event chain. The remaining candidate intermediate states are retained as alternative branches, and the complementary value of the normalized state confidence value is used as the uncertainty.

[0083] The forward candidate state sequence starts from the execution result of the previous observed event, enumerates the state transition rules in the calibrated model that satisfy the current state, and expands to the start time of the next observed event. The reverse candidate state sequence starts from the previous state of the next observed event, traces back along the input-output mapping of the state transition rules to the end time of the previous observed event. During the expansion process, sequences that violate the exclusive relationship of personnel action chain tasks, the confirmation relationship of instruction transmission chain, the reachability relationship of regional state chain, or the result dependency relationship of response and disposal chain are deleted. The intersection of the forward sequence and the reverse sequence is determined by the state item set, the overlap of occurrence intervals, and the consistency of inter-layer constraints. When the intersection is empty, it is expanded to one inter-layer constraint node directly adjacent to the missing event interval to regenerate the candidate sequence.

[0084] ;

[0085] The above formula is used to calculate the state confidence value and uncertainty of the k-th candidate intermediate state, where, Indicates the state confidence value. This represents the normalized time deviation between a locally replayed event and its adjacent observed events. This indicates the proportion of inconsistency between the local replay result and the preceding state of the subsequent observed event. This represents the proportion of inconsistent edges among the directed edges within a layer and the constraint edges between layers formed by local replay. , and This represents the fixed weight coefficients of the three categories of components, and their sum is 1. Let m represent the uncertainty, and m represent the total number of candidate states. In a calculation example, let m be a value that represents the uncertainty. , , When the three types of biases for a candidate state are 0.1, 0, and 0.2, If the sum of the confidence values ​​of the three candidate states is 2.10, then... It is approximately 0.567.

[0086] When candidate intermediate states are written into a multi-layered time-series event graph, the first candidate state is incorporated into the main event chain as an inferred state node, and the previous observed event, the next observed event, the model version, and the local replay record on which it is based are retained. The remaining candidate states are established as alternative branches. The alternative branches share the observed nodes before the missing interval and the target nodes after the missing interval with the main event chain. Uncertainty is saved as a branch attribute. When subsequent live events arrive, the consistency between each candidate state and the preceding state and upstream and downstream relationships of the new event is recalculated. Alternative branches that can satisfy the constraints of the new event are retained, while branches that cannot are marked as excluded but their associated evidence is not deleted.

[0087] In another preferred embodiment, residual replay verification is performed on the model version after cross-sub-model joint calibration, including: extracting the stable event interval before joint calibration, the residual event interval that triggered joint calibration, and the subsequent event interval after residual termination from the model version chain; continuously replaying the three event intervals using the model version before calibration and the model version after calibration to form corresponding residual propagation trajectories; comparing the changes of the two sets of residual propagation trajectories at the residual source node, cross-chain diffusion association, and residual termination node, and when the model version after calibration generates new residuals in the stable event interval, splitting the joint correction group that causes the new residuals into multiple sub-correction groups; replaying sequentially according to the coverage relationship of each sub-correction group to the original residual path, selecting the sub-correction group that does not change the stable event interval and covers the original residual path, and replacing the corresponding model version after calibration in the model version chain.

[0088] The stable event interval is determined by tracing back from the residual start node to the position where there are no state transition residuals in multiple consecutive matching events. The residual event interval extends from the residual start node to the latest terminating node of all residual paths. The subsequent event interval extends from the latest terminating node to the position where the model state re-enters the confirmed stable state. During continuous replay, the three event intervals share the same initial state and keep the live event input unchanged. The residual propagation trajectory records the residual category, direction, source event chain, and cross-chain diffusion association of each event node. When comparing, it is required that the original residual path in the calibration trajectory is shortened or eliminated and that the stable event interval does not contain residuals that did not exist before calibration.

[0089] When splitting the joint correction group, sub-correction groups are formed according to the direct dependencies of state variables, state transition rules, and cross-sub-model input-output relationships. Objects with input-output dependencies remain in the same sub-correction group, while objects without direct dependencies are placed in different sub-correction groups. Each sub-correction group independently performs replay with the pre-calibration model version as the parent version. Sub-correction groups that can cover the original residual path but change the stable event interval are excluded. When multiple sub-correction groups need to be combined to cover all the original residual paths, they are combined in ascending order of the number of new residuals and the number of correction objects. The selected combination forms an alternative model version, and the replaced version is kept in the model version chain as an inactive branch.

[0090] In a preferred embodiment, establishing virtual exercise branches based on calibrated states includes: encapsulating the observed states in the main event chain, the validated model version, and the candidate branches within the unobserved event interval into a branch start snapshot; inheriting candidate intermediate states and uncertainties for each candidate branch, and generating parallel virtual event sequences using subsequent exercise event inputs; when the parallel virtual event sequences satisfy the same preceding states, event order, and cross-chain association at the same state node, merging the parallel virtual event sequences into a shared successor branch, and summarizing their state confidence values ​​according to the complementary uncertainty values ​​of each candidate intermediate state; when subsequent live-fire events arrive, sorting them according to the consistency between the subsequent live-fire events and each parallel virtual event sequence in terms of event subject, target, time, preceding states, and execution results, connecting the top-ranked virtual event sequence to the main event chain, and retaining the remaining virtual event sequences in the branch start snapshot.

[0091] The branch start snapshot includes the snapshot time, the current end node of the main event chain, the current state of each event chain, the model version identifier, the candidate branch identifier, the candidate intermediate state, the uncertainty, and the associated evidence index. Each parallel virtual event sequence starts running from the same snapshot time, but inherits the candidate intermediate state of the corresponding candidate branch. Subsequent exercise event inputs are delivered to the parallel sequences that meet the conditions according to their previous states. If a sequence cannot meet the previous states of the input event, the original state is retained and the input conflict is recorded. The sequence does not continue to run by forcibly modifying the state. Parallel sequences are determined to have reached the same state node when they have the same event main state, region state, task state, and cross-chain association set.

[0092] ;

[0093] The above formula is used to rank the consistency between the d-th parallel virtual event sequence and subsequent live-fire events, where, Indicates the consistency score. The value is 1 if the subject of the event is the same, otherwise it is 0. The value is 1 when the target objects are the same, 0.5 when they belong to the same parent region, and 0 in other cases. This represents the truncation value after dividing the time deviation from the event occurrence by the current event window length. This indicates that a value of 1 is taken when all prerequisite conditions are met, and a value is taken when only some conditions are met, representing the proportion of the met conditions. This indicates the proportion of consistency between the execution result status items. to This represents the weight coefficients corresponding to the five sub-items, and their sum is 1. In a calculation example, each is taken as 0.2. When the five items of a certain sequence have values ​​of 1, 1, 0.9, 0.8, and 1, The score is 0.94, and the sequence of events to be connected to the main event chain is determined by ranking the scores from highest to lowest.

[0094] When merging parallel virtual event sequences, only the subsequent parts after reaching the same state node are merged. Different paths before merging are retained in the branch start snapshot, and the state credibility value of the shared subsequent branches is determined according to the source branches. Normalization is performed and the mapping from the source branch to the shared successor branch is saved. When the scores of subsequent live events are the same as those of multiple parallel sequences, the sequence with the complete satisfaction of the preceding state and the continuous upstream and downstream event chains is selected first. If they still cannot be distinguished, the parallel main branches are retained until the subsequent events arrive. A unique branch is not forcibly determined based on a single time proximity relationship.

[0095] Specifically, the entire operation uses a multi-layered time-series event graph as an intermediate representation between live-fire data and the digital twin model. After a live-fire event arrives, standardization, local consistency verification, and event chain updates are performed first. The digital twin model generates predicted events based on the updated event graph and completes time-series alignment. Residuals are entered into local calibration or cross-sub-model joint calibration according to their propagation range. The calibration results are saved through the model version chain and verified through residual playback. Missing states are bidirectionally inverted based on the constraints of previous and subsequent events. The verified current state, together with the alternative states, forms a branch start snapshot. Subsequent virtual exercises and live-fire events continue to iterate on the same event semantics and model version, thereby maintaining a traceable correspondence between exercise records, model correction basis, state inversion results, and virtual branches.

Claims

1. A method for nuclear security digital twin drills and decision-making based on live-fire confrontation, characterized in that, include: Collect data on personnel actions, command transmission, area status, and handling results during live-fire exercises, and encode them into standard event units that include the event subject, target, time, prior state, and execution result. Based on the standard event unit, a multi-layered time-series event diagram is constructed, consisting of personnel action chain, instruction transmission chain, area status chain, and response and handling chain; The digital twin model generates a predicted event sequence based on the multi-layer time-series event graph. The predicted event sequence is then time-aligned with the actual event sequence to obtain the occurrence order, duration interval, state transition, and upstream-downstream consistency residual. Based on the propagation range of various residuals in the multi-layer time-series event graph, local state transition relationships or cross-chain association relationships are selected for hierarchical calibration; Based on the calibrated digital twin model and adjacent event constraints, the unobserved intermediate state is inverted, the inversion results and uncertainties are written into the multi-layer time series event graph, and a virtual exercise branch is established with the calibrated current state.

2. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 1, characterized in that, The standard event unit also includes an event source identifier, a time confidence interval, a state confidence level, and an associated evidence index; When constructing the multi-layer time-series event graph, directed edges within the layer are established according to the continuous action relationship of the same event subject, the instruction sending and receiving relationship, the successive relationship of the regional state, and the correspondence between the handling action and the result. Inter-layer constraint edges are established based on the shared fields of the same standard event unit in different event chains. For multiple standard event units with inconsistent source times, their original timestamps are retained, and relative order constraints are generated using common preceding events and common succeeding events. The relative order constraints are written into the directed edges in the corresponding layer, and parallel event clusters are formed for events whose unique sequential relationship cannot be determined. The digital twin model generates predicted event sequences using the parallel event clusters as branch nodes, and merges branches consistent with the actual event sequences when subsequent live-fire events arrive.

3. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 1, characterized in that, The time-series alignment of the predicted event sequence with the actual military event sequence includes: constructing a candidate matching key based on the event subject, the target of action, and the preceding state, and generating candidate event pairs from the predicted event sequence and the actual military event sequence according to the candidate matching key; Based on the sequential relationship of each candidate event pair in the corresponding event chain and the cross-chain relationship between different event chains, candidate event pairs that disrupt the existing event order relationship are eliminated. For the remaining candidate event pairs, establish joint matching costs for occurrence time offset, continuous interval overlap, state transition consistency, and upstream and downstream adjacency consistency, and determine the correspondence according to the constraint of maintaining the continuity of the event chain; For predicted events and live-fire events that do not form a corresponding relationship, they are marked as virtual redundant events and live-fire newly added events, respectively, and the correlation between the two in the multi-layer time series event graph is used as the component data of state transition residual and upstream and downstream consistency residual.

4. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 1, characterized in that, The digital twin model is calibrated hierarchically, including dividing the digital twin model into scene state sub-model, personnel behavior sub-model, instruction coordination sub-model and result evolution sub-model, and establishing a mapping table between the correctable state variables, state transition rules and cross-sub-model input-output relationships of each sub-model. Based on the event chain type, start node, and end node corresponding to various residuals, the state variables, state transition rules, or cross-sub-model input-output relationships that intersect with the residual propagation path are selected from the mapping table to form a candidate correction set; Local calibration is performed according to the minimum coverage of the candidate correction set in the multi-layer time series event graph, and extended to cross-sub-model joint calibration when the same residuals persist across at least two event chains; For each calibration, the model version before calibration, the calibration basis event, the correction object, the applicable event chain scope, and the model version after calibration are recorded to form a model version chain with inheritance relationship.

5. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 1, characterized in that, When updating the multi-layer timing event graph, incremental consistency verification is performed on the newly arrived standard event unit, including: locating the local subgraph that shares the preceding or succeeding state with the newly arrived standard event unit based on the event subject and the target of action; The reachability order of the newly arrived standard event unit is compared with that of the events in the local subgraph based on the time confidence interval, and the inter-layer constraint edge is called to determine the acceptable order set when the time confidence intervals overlap. When the newly arrived standard event unit causes the directed edge in the existing layer to form a state rollback, repeated transfer or broken transfer, the directed edge in the corresponding layer is converted into a verifiable edge, and a local reconnection structure containing alternative preceding events and alternative succeeding events is generated around the verifiable edge. Each local reconnection structure is input into the digital twin model and replayed within the event interval corresponding to the local subgraph. The local reconnection structure that simultaneously satisfies the subsequent events of the actual troops and the cross-chain constraints is retained, and the parallel event clusters and their branch nodes are updated.

6. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 3, characterized in that, The objects to be corrected are determined based on the propagation range of various residuals in the multi-layer time-series event graph, including: taking the candidate event pair that generates the residual as the starting node of the residual, and tracing the reachable nodes of the residual in the personnel action chain, instruction transmission chain, regional status chain and response and handling chain along the preceding and following associations of the corresponding event chain and the cross-chain associations between different event chains. For each reachable node, record the first occurrence position of the residual, the consistency of the direction after the continuous maintenance interval and cross-chain mapping, and classify the residuals that are only located in a single event chain and disappear at the subsequent matching event into local state transition residuals. Residuals that span different event chains and maintain the same offset across multiple subsequent matching events are classified as cross-chain associated residuals. For the residuals caused by the newly added military event, compare the event chain connectivity before and after the addition of the newly added military event, and write the new association that eliminates the upstream and downstream consistency residuals after the addition into the state transition relationship or cross-chain association relationship to be corrected.

7. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 4, characterized in that, When forming the candidate correction set, the basic correction range is first determined based on the shortest reachable subgraph between the residual start node and the residual end node, and then the correctable state variables and state transition rules that directly correspond to the state transition relationship within the basic correction range are extracted from the mapping table. For the basic correction range that includes cross-chain associations, the cross-sub-model input-output relationship of the sub-models to which both ends of the cross-chain association belong is added to the candidate correction set; For each correction object in the candidate correction set, a single-object calibration copy is generated while keeping other correction objects unchanged. The direction of residual change obtained by replaying each single-object calibration copy is compared with the original residual propagation direction, and correction objects that expand the original residual propagation range are eliminated. The remaining correction objects are merged into joint correction groups according to the residual paths they share, and calibration is performed using the joint correction group that requires the fewest correction objects to cover all residual paths.

8. The method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 1, characterized in that, When inverting unobserved intermediate states, the event intervals located between adjacent observed events and lacking state records are extracted from the multi-layer time-series event graph. A positive candidate state sequence is generated based on the execution result of the previous observed event, and a reverse candidate state sequence is generated based on the preceding state of the next observed event. The intersection of the positive candidate state sequence and the reverse candidate state sequence is obtained while maintaining the reachability relationship of each event chain and the consistency of cross-chain association. For each candidate intermediate state in the intersection, the calibrated digital twin model is invoked to perform local replay. Based on the degree of matching between the local replay results and the adjacent observed events in terms of occurrence time, state transition, and upstream and downstream correlation, a state confidence value is generated. The first candidate intermediate state is selected from high to low according to the state confidence value and added to the main event chain. The remaining candidate intermediate states are reserved as alternative branches, and the complementary value of the normalized state confidence value is used as the uncertainty.

9. A method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 4, characterized in that, Perform residual playback verification on the model version after the cross-sub-model joint calibration, including: extracting the stable event interval before the joint calibration, the residual event interval after the joint calibration, and the subsequent event interval after the residual termination from the model version chain; The three event intervals were continuously replayed using the pre-calibration model version and the post-calibration model version to form the corresponding residual propagation trajectory; Compare the changes of the two sets of residual propagation trajectories at the residual source node, cross-chain diffusion correlation and residual termination node, and when the calibrated model version generates new residuals in the stable event interval, split the joint correction group that causes the new residuals into multiple sub-correction groups. The model is replayed sequentially according to the coverage relationship of each sub-correction group to the original residual path. The sub-correction group that does not change the stable event interval and covers the original residual path is selected and replaced with the corresponding calibrated model version in the model version chain.

10. A method for nuclear security digital twin drills and decision-making based on live-fire confrontation as described in claim 8, characterized in that, Establishing the virtual exercise branch in the calibrated state includes: encapsulating the observed state in the main event chain, the verified model version, and the alternative branches in the unobserved event interval into a branch start snapshot; For each candidate branch, inheriting intermediate states and uncertainties, and using subsequent exercise event inputs, generate a parallel virtual event sequence; When the parallel virtual event sequences meet the same preceding state, event order and cross-chain association at the same state node, the parallel virtual event sequences are merged into a shared successor branch, and their state confidence values ​​are normalized and summarized according to the complementary uncertainty values ​​of each candidate intermediate state. When subsequent live-fire events arrive, they are sorted according to their consistency with each parallel virtual event sequence in terms of event subject, target, time, preceding state, and execution result. The virtual event sequence with the highest ranking is connected to the main event chain, and the remaining virtual event sequences are retained in the branch start snapshot.

Images