A safety emergency drill and actual combat collaborative review method and system based on double-layer scoring

Abstract

The present application belongs to the technical field of emergency drill, and particularly relates to a safety emergency drill and actual combat collaborative review method and system based on double-layer scoring. The present application converts an emergency plan into a structured plan with checkpoints as basic units, adopts a Stage arrangement engine with a deterministic topology structure of Start→Host→{Slots}→End, drives the drill and actual combat scene in the same framework by switching the event source and Slot role participation mode, records the full-process structured track in the execution process, calculates the double-layer automatic scoring of the Slot layer behavior score and the Host layer scheduling quality score after the execution, generates the replayable review data with the highlight of the lost points by associating the scoring result with the space-time in the execution process, and establishes the two-way knowledge closed loop of the drill and actual combat. The present application realizes the seamless collaboration of peacetime and wartime and the accurate evaluation of the whole link of emergency disposal, and greatly improves the scientificity and effectiveness of safety emergency management.

Description

Technical Field

[0001] This invention belongs to the field of emergency drill technology, specifically relating to a method and system for collaborative review of safety emergency drills and actual combat based on a two-level scoring system. Background Technology

[0002] Emergency management is a crucial component of safety production systems in high-risk industries such as fire protection, chemicals, mining, and power. Traditional emergency management methods typically rely on pre-prepared written emergency plans (hereinafter referred to as "plans") developed by safety experts. These plans specify accident types, response procedures, role assignments, and resource allocation. During drills or real-world exercises, relevant personnel manually execute various operations according to the plan, and afterwards, managers evaluate or compile statistics based on the results.

[0003] Existing technologies typically construct separate training and operational systems. However, training systems focus on low-cost simulation training, often employing simplified event sources and manual scoring methods. Operational systems, on the other hand, require high reliability and low latency, necessitating integration with real IoT alarms and device control. In existing technologies, the pre-set plans, scoring criteria, and role assignments for the two systems are often maintained separately. This prevents the automatic transfer of validated processes and parameters from training to operational scenarios. Furthermore, issues such as scheduling delays and role omissions exposed during operational scenarios are difficult to systematically feed back into the training system for iterative correction. Consequently, the transition between training and operational scenarios requires manual reconfiguration or even complete redevelopment, resulting in slow response times and high costs.

[0004] Furthermore, existing evaluation mechanisms primarily focus on the execution phase of emergency drills or real-world operations. This includes assessing the compliance of frontline personnel (such as patrol officers and firefighters) – specifically whether frontline personnel or automated equipment completing tasks according to the prescribed actions, timelines, and evidence requirements; and secondly, dispatch quality – indicators such as the timeliness of task allocation by the dispatch center, the accuracy of role matching, and the compliance of escalation chain execution. In existing operational systems, the dispatch process is often decoupled from the specific execution process of the plan, failing to differentiate dispatch quality from execution behavior within the same framework. This misalignment of evaluation objects means that executor scores cannot reflect the performance deficiencies of the dispatch center, while the dispatch center's scores lack linkage with the plan execution process. This results in an incomplete overall emergency effectiveness assessment, hindering targeted optimization of dispatch strategies or plan configurations. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a method and system for collaborative review of safety emergency drills and real-world operations based on a two-layer scoring system. This addresses the issues in existing safety emergency management technologies, such as the disconnect between drill and real-world systems, the inability to transfer knowledge bidirectionally, and the fact that the evaluation system focuses only on frontline execution and fails to differentiate between assessing dispatch quality and execution behavior, resulting in incomplete emergency effectiveness assessments and difficulties in targeted optimization.

[0006] The objective of this invention can be achieved through the following technical solutions: A method for collaborative review of safety emergency drills and real-world scenarios based on a two-level scoring system includes the following steps: S1: Transform the emergency plan into a structured plan with checkpoints as the basic unit. The checkpoints include action description fields, role division fields, completion time limit fields, escalation chain condition fields, and evidence requirement fields. S2: In the Stage orchestration engine, set up a Host role and at least one Slot role; the Host role is used to receive emergency events or drill events and generate task allocation decisions according to structured contingency plans; the Slot role is used to receive and execute the tasks assigned by the Host role and return the execution results. S3: During execution, record the response time, action completion rate, and evidence upload status of each Slot role to each checkpoint, as well as the task allocation time, upgrade trigger time, and role matching record of the Host role; S4: After execution, a two-layer automatic scoring is performed based on the recorded trajectory, including the following process: For each slot role, a weighted summation model is used to calculate the behavior score based on the expected behavior, time limit, and evidence requirements of the corresponding checkpoint in the structured plan. For the Host role, a multi-dimensional weighted model is used to calculate the scheduling quality score based on the timeliness of task allocation, the accuracy of role matching, and the compliance of upgrade chain execution. S5: Spatiotemporally correlate the results of the two-layer automatic scoring with the execution process to generate replayable debriefing data, which is used to show the collaboration process between the Host role and each Slot role and highlight the points lost.

[0007] Preferably, the Stage orchestration engine is a directed graph topology; The directed graph topology includes an event injection node Start, a unique scheduling node Host, at least one execution node Slot, and a result aggregation node End. Directed edges are only allowed to point from the start node to the scheduling node, from the scheduling node to each execution node, and from each execution node to the end node, forming a deterministic execution path Start→Host→{Slots}→End. Tasks are processed in an orderly manner through an event-driven message queue.

[0008] Preferably, the event-driven message queue is EventInbox, which performs idempotent deduplication and deterministic batch processing on each received event, so that the same event is processed only once and the order of events within the same batch is determined. The idempotent deduplication method generates a unique hash key based on the event source identifier and the event sequence number, and the deterministic batch processing uses a FIFO queue combined with a batch window timer.

[0009] Preferably, the structured contingency plan is used in both drill scenarios and actual combat scenarios; The Stage orchestration engine drives training or real-world scenarios within the same framework by switching event sources and Slot role participation modes. The event sources are script events in training scenarios and real IoT alarms in real-world scenarios. The Slot role participation modes include: NPC mode, human-managed mode, and duty robot-managed mode.

[0010] Preferably, S1 also includes narrative interpolation of the general contingency plan template based on the owner's actual personnel configuration, equipment parameters, and service level agreement (SLA) to generate a contingency plan instance specific to the owner.

[0011] Preferably, in S4, a weighted summation model is used to calculate the behavior score, and the formula for calculating the single checkpoint score is: ; In the formula, The score for the i-th checkpoint. The actual response time at checkpoint i. The time limit requirement specified for checkpoint i in the structured contingency plan, For indicator functions, Let i be the number of sub-actions actually completed at checkpoint i. Let i be the total number of sub-actions expected to be completed at checkpoint i. The number of valid evidence types actually uploaded at checkpoint i. The number of evidence types required for checkpoint i, This is an indicator function.

[0012] Preferably, in step S4, the calculation of the behavior score using a weighted summation model further includes the following steps: If the Slot role fails to upload compliant evidence, the behavior score for the corresponding checkpoint is multiplied by a preset decay factor; if the checkpoint is critical and compliant evidence is not uploaded, the checkpoint score is directly determined to be zero.

[0013] Preferably, in S4, a multi-dimensional weighted model is used to calculate the scheduling quality score, and the calculation formula is as follows: ; In the formula, These are the weight coefficients for the four rating dimensions, and they satisfy... , The first distribution of delayed points, Score the character matching score. To upgrade the blockchain compliance score, The score is for closed-loop integrity.

[0014] Preferably, in S5, automatically marking and highlighting the points that have lost marks includes the following process: Set a first threshold and a second threshold. The first threshold is the score threshold for the Slot layer, and the second threshold is the score threshold for the Host layer. Automatically mark checkpoints or scheduling actions where the Slot layer score is below the first threshold or the Host layer score is below the second threshold; During the replay, when the timeline reaches the corresponding point of deduction, the detailed information of that point of deduction is automatically highlighted, including the reason for the deduction, the violation of the plan requirements, the expected behavior, and the actual implementation.

[0015] A safety emergency drill and practical collaborative debriefing system based on dual-level scoring includes: The contingency plan configuration module is used to transform safety emergency plans into a structured set of checkpoints and supports narrative interpolation to generate custom contingency plan instances. The Stage orchestration engine includes a Host scheduling unit and a Slot execution unit for executing collaborative processes with a fixed graph structure; the Stage orchestration engine also includes an EventInbox unit for idempotent deduplication and deterministic batch processing of events. The event access module is used to receive drill script events or real IoT alarms and inject them into the EventInbox unit; The scoring calculation module is used to automatically calculate the behavior score of the Slot layer and the scheduling quality score of the Host layer after execution. The scoring calculation module has a built-in weighted summation model and an adaptive weight adjustment unit. The debriefing and replay module is used to generate debriefing data with timelines and two-level scoring, and supports jumping from the highlighted points of deduction to the corresponding contingency plan checkpoint definition; The scenario package management module is used to store and load scenario packages from different security domains. The scenario packages include checkpoint definitions, role types, device protocol mappings, and scoring weight coefficients.

[0016] The beneficial effects of this invention are as follows: This invention proposes a stage orchestration engine architecture that integrates peacetime and wartime needs. First, it achieves idempotent deduplication and deterministic batch processing through an event-driven message queue (EventInbox), simultaneously meeting the low-cost simulation requirements of drills and the high reliability and low latency requirements of real-world operations. Second, it employs a two-layer automatic scoring model that separates the scheduling quality of the Host layer from the execution behavior of the Slot layer. For the first time, it incorporates scheduling indicators such as task allocation timeliness, role matching accuracy, and escalation chain compliance into a standardized evaluation system, and achieves quantifiable and traceable scoring through mathematical formulas. Third, it innovates in contingency plan adaptation by proposing a method for personalized transformation of general contingency plans based on narrative interpolation, solving the industry problem that general contingency plans cannot be directly adapted to the actual scenarios of owners. Fourth, it innovates in knowledge transfer by establishing a two-way knowledge closed-loop mechanism between drills and real-world operations, enabling the automatic accumulation, transfer, and iteration of emergency response experience. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0018] Figure 1 A flowchart illustrating the steps of the safety emergency drill and practical collaborative review method based on dual-level scoring provided by this invention; Figure 2 This is a schematic diagram of the architecture of the safety emergency drill and practical collaborative review system based on dual-level scoring provided by the present invention. Detailed Implementation

[0019] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided.

[0020] Please see Figure 1 This embodiment provides a method for safety emergency drills and collaborative debriefing based on a two-level scoring system. The method includes the following steps: S1: Transform the emergency plan into a structured plan with checkpoints as the basic unit. The structured plan is used in both drill scenarios and actual combat scenarios. Specifically, in this step, we first obtain a general written emergency plan compiled by industry security experts and break it down into several smallest execution units, namely checkpoints. Each checkpoint is the smallest action unit that can be executed and verified independently.

[0021] Each checkpoint contains fixed structured fields, including action description, role assignment, completion deadline, escalation chain conditions, and evidence requirements. Each field specifically includes: Action description field: Used to store the specific action to be performed for this checkpoint, including the execution standards, operation steps and compliance requirements of the action; Role assignment field: Used to store the execution role corresponding to this checkpoint, including role name, job responsibilities, scope of authority, etc., and supports mapping to specific personnel identifiers as well as AI agent identifiers or robot subsystem identifiers; Completion time limit field: Used to store time constraint parameters such as the maximum allowed completion time, response time limit, and upgrade trigger time limit for this checkpoint; Upgrade Chain Condition Field: This field stores the upgrade handling rules when the checkpoint is not completed as required, including upgrade trigger conditions, target role for upgrade, and post-upgrade handling procedures. Evidence Requirements Field: This field stores the supporting documentation required for the completion of the checkpoint, including at least one of the following: photographic evidence, video stream retrieval, equipment operation records, personnel sign-in records, and task confirmation receipts.

[0022] Furthermore, this step also includes narrative interpolation of the general contingency plan template based on the owner's actual personnel configuration, actual equipment parameters, and Service Level Agreement (SLA) to generate a contingency plan instance specific to the owner. The specific process of narrative interpolation includes: Map the general role field in the general contingency plan template to the actual personnel identifiers and job identifiers of the owner unit; Map the general equipment type field in the general contingency plan template to the actual equipment number, equipment address, and communication protocol type deployed on-site by the owner; Replace the general time limit field in the general contingency plan template with the time limit value corresponding to the Service Level Agreement (SLA) agreed upon by the owner; Based on the site environment, building structure, and emergency resource allocation of the owner, the execution sequence, escalation chain conditions, and evidence requirements of the checkpoints will be adjusted.

[0023] Once the structured contingency plan is configured, it will serve as the core knowledge of the entire system, applicable to both drill and real-world scenarios. It acts as a behavioral constraint for AI agent reasoning and execution, a scoring standard for drill scenarios, and an audit baseline for real-world scenarios.

[0024] S2: In the Stage orchestration engine, set up a Host role and at least one Slot role; the Host role is used to receive emergency events or drill events and generate task allocation decisions according to the structured contingency plan; the Slot role receives and executes the tasks assigned by the Host role and returns the execution results.

[0025] Specifically, the Stage orchestration engine adopts a fixed directed graph topology, which includes an event injection node Start, a unique scheduling node Host, at least one execution node Slot, and a result aggregation node End. Directed edges are only allowed from the start node to the scheduling node, from the scheduling node to each execution node, and from each execution node to the end node, forming a deterministic execution path Start→Host→{Slots}→End. Tasks are processed in an orderly manner through an event-driven message queue. Directed edges are only allowed from Start to Host, from Host to each Slot, and from each Slot to End, prohibiting edges between Slots or from Slot to Host. This provides a deterministic basis at the architectural level for quantifying execution latency into Host scheduling latency and Slot execution latency, and ensures that the execution path of the entire emergency response process is clear and predictable, avoiding process chaos caused by free flow.

[0026] Among them, the Host role is the sole scheduling hub of the entire collaborative process, corresponding to the unique scheduling node Host in the directed graph. In this embodiment, the Host role is implemented by the intelligent AI safety officer, which has built-in large language model LLM reasoning capability. With structured plans as the constraint framework, after receiving an event, it completes the entire process scheduling of event analysis, task splitting, role matching, task assignment, process advancement, escalation triggering, and closed-loop supervision according to the structured plans.

[0027] The Slot role is the execution unit of the collaborative process, corresponding to the execution node Slot in the directed graph. Each Slot role corresponds to an execution subject, which can be any one of the following: front-line personnel, AI agents, or fire control duty robots.

[0028] Furthermore, the Stage orchestration engine drives drill or real-world scenarios within the same framework by switching event sources and Slot role participation modes. In drill scenarios, the event source is a script event, while in real-world scenarios, it is a real IoT alarm. Slot role participation modes include the following three, which can be mixed and used in the same scenario: The NPC mode, which is automatically inferred by a large language model, is automatically inferred and executed by the robot_core large language model. It has built-in structured pre-plan constraints that are the same as those of the Host role, and automatically completes the entire process of task reception, inference execution, and result feedback. It is suitable for fully automated pre-plan quality verification, serving as a training partner in drills, and simulating multi-role collaborative scenarios. The human-hosted mode, operated by real personnel through a client, involves real personnel participating in the execution through the emergency collaboration client, receiving tasks assigned by the Host role, and uploading the execution results and evidence materials through the client after completion. It is suitable for frontline personnel training, multi-role joint drills, and on-site personnel handling in real combat scenarios.

[0029] The on-duty robot managed mode is executed by an independently deployed fire control on-duty robot AI subsystem, which has the ability to continuously monitor 24 / 7. In emergency collaboration, it temporarily accesses the Stage engine's session to convert the host's dispatch instructions into advisory suggestions and present them to the fire control room operator. After the operator confirms, it executes equipment operations and processes. At the same time, it supports the automatic execution of dispatch instructions according to preset authorization policies when there is no operator, and completes operations such as remote equipment control, video image retrieval, and alarm information reporting.

[0030] Furthermore, the Stage orchestration engine itself does not contain any safety and emergency domain knowledge; all domain-related content is injected through scenario packages, achieving decoupling between the engine and domain knowledge. These scenarios include checkpoint definitions, role types, device control protocol mappings, weight coefficients in scoring rules, escalation chain rules, and other content specific to the corresponding safety domain. By changing different scenario packages, the same Stage engine can be used for emergency drills and practical responses in different safety domains such as fire fighting, chemical engineering, mining, or power, significantly improving the system's reusability.

[0031] Furthermore, the Stage orchestration engine processes all events in an ordered manner through the event-driven message queue EventInbox. EventInbox performs idempotent deduplication and deterministic batch processing on each received event, ensuring that the same event is processed only once and that the order of events within the same batch is fixed, effectively shielding the impact of IoT alarm jitter and duplicate reporting on system stability.

[0032] The specific implementation of the idempotent deduplication is as follows: a unique hash key is generated based on the event source identifier and the event sequence number. After receiving the event, EventInbox first checks whether the hash key already exists in the processed event library. If it already exists, the event is discarded directly; otherwise, it is included in the processing queue.

[0033] The specific implementation of deterministic batch processing is as follows: a FIFO first-in-first-out queue is used in combination with a batch window timer. A batch window of fixed duration is preset. Within each batch window, events are sorted according to the order in which they enter the queue. After the batch window ends, the events in the batch are submitted to the Host role for processing in a fixed order, ensuring the determinism of the event processing order.

[0034] S3: During execution, record the response time, action completion rate, and evidence upload status of each Slot role at each checkpoint, as well as the task allocation time, upgrade trigger time, and role matching record of the Host role. Specifically, in this step, the Stage engine performs structured, full-trace recording of all operations, decisions, and execution behaviors throughout the entire session, generating complete execution trajectory data, including: For each Slot role, the following data is recorded according to the checkpoint dimension: task reception time, response time, execution time and result of each sub-action, completion confirmation time, uploaded evidence materials and upload time, and abnormal reporting records. Based on the above data, the response time, action completion rate and evidence upload status of the Slot role to the corresponding checkpoint are calculated. For the Host role, the following data is recorded: event reception time, task assignment time and assignment record, role matching record for each task, upgrade trigger time and upgrade reason, process supervision record, and closed-loop confirmation time, forming a complete scheduling decision trajectory; All recorded data carries a unique timestamp, a corresponding checkpoint identifier, and a corresponding role identifier, forming a structured execution trajectory dataset for subsequent two-level scoring and review playback.

[0035] S4: After execution, a two-layer automatic scoring is performed based on the recorded trajectory, including the following process. Specifically, in this step, after the session of the emergency drill or actual handling is closed, the system automatically calls the scoring calculation module to perform a two-layer automatic scoring based on the recorded complete execution trajectory data. The specific process is as follows: Calculate the behavior score for the Slot layer: For each slot role, a weighted summation model is used to calculate the behavior score based on the expected behavior, time limit, and evidence requirements for the corresponding checkpoint in the structured contingency plan. The weight of each checkpoint is pre-defined according to its importance level in the contingency plan, which is divided into three levels: general, important, and critical, with corresponding weight coefficients of 1, 3, and 5, respectively. The formula for calculating the score for a single checkpoint is as follows: ; In the formula, The score for the i-th checkpoint. The actual response time at checkpoint i. The time limit requirement for this checkpoint is specified in the structured contingency plan. This is an indicator function that takes the value 1 when the condition within the parentheses is true and 0 when it is false. Let i be the number of sub-actions actually completed at checkpoint i. Let i be the total number of sub-actions expected to be completed at checkpoint i. The number of valid evidence types actually uploaded at checkpoint i. The number of evidence types required for checkpoint i, This is an indicator function that takes the value 1 when the condition within the parentheses is true and 0 when it is false. , , The weighting coefficients for the three dimensions of response time, action completion, and evidence compliance, and satisfying the following conditions: The preferred values ​​are 0.4, 0.4, and 0.2 respectively. Users can adjust the weight coefficients through the scenario package according to their actual needs.

[0036] Furthermore, during the Slot layer scoring process, if the Slot role fails to upload records that meet the evidence requirements, the behavior score of the corresponding checkpoint is multiplied by a preset attenuation factor. The attenuation factor ranges from 0 to 1, and in this embodiment, the attenuation factor is 0.3. If the checkpoint is a critical checkpoint and no evidence that meets the requirements is uploaded, the score of the checkpoint is directly determined to be zero.

[0037] The total behavior score of a single slot role is obtained by multiplying the individual checkpoint scores of all checkpoints corresponding to that role by the weight coefficient of the corresponding checkpoint and then summing them up.

[0038] Calculate the host layer scheduling instruction score: For the Host role, a multi-dimensional weighted model is used to calculate the scheduling quality score based on the timeliness of task allocation, the accuracy of role matching, and the compliance of escalation chain execution. This model includes four scoring dimensions: ① Initial dispatch delay score, used to evaluate the timeliness of the Host role's response to the event, is calculated using the following formula: Where d is the actual delay from the event trigger to the Host role's first task assignment, and D is the maximum allowable delay threshold specified in the structured contingency plan; ② Role matching score, used to evaluate the accuracy of Host role task allocation, is calculated using the following formula: ,in, To correctly allocate the number of tasks for each role, The total number of tasks is determined by the fact that the assigned roles are consistent with the execution roles specified in the structured plan. ③ Upgrade Chain Compliance Score, used to assess the Host role's compliance with upgrade chain rules, is calculated using the following formula: ,in, This represents the actual trigger time of the j-th upgrade. J represents the maximum allowable delay time for this upgrade as specified in the structured contingency plan, and J represents the total number of upgrades during this process. ④ Closed-loop integrity score, used to evaluate the Host role's closed-loop control capability over the entire handling process, is calculated using the following formula: ,in, This represents the total number of checkpoint chains that have completed the closed loop. This represents the total number of checkpoint chains that should be closed during this process. Based on the scores from the above four dimensions, a multi-dimensional weighted model is used to calculate the overall scheduling quality score for the Host role. The calculation formula is as follows: ; In the formula, These are the weight coefficients for the four rating dimensions, and they satisfy... In this embodiment, the preferred values ​​are 0.2, 0.3, 0.3, and 0.2, respectively. Users can adjust the weighting coefficients through scene packages according to their actual needs. The first distribution of delayed points, Score the character matching score. To upgrade the blockchain compliance score, The score is for closed-loop integrity.

[0039] Furthermore, this step also includes an adaptive weight adjustment step for Host layer scoring. Specifically, based on the statistical distribution of historical exercise or real-world data, the weight values ​​in the Host role's total score calculation formula are dynamically adjusted to ensure the comparability of scheduling quality scores across different scenarios. This adaptive weight adjustment employs the entropy weight method, automatically adjusting the weight coefficients based on the dispersion of scores for each dimension in historical data; dimensions with greater dispersion receive higher weight coefficients. Alternatively, an online learning algorithm based on gradient descent can be used, with the effectiveness of the handling results as the optimization objective, to iteratively update the weight coefficients online.

[0040] S5: Spatiotemporally correlate the two-layer scoring results with the execution process to generate replayable debriefing data, which is used to demonstrate the collaboration process between the Host role and each Slot role and highlight the points lost.

[0041] Specifically, in this step, the calculated dual-layer scoring results are spatiotemporally correlated with the recorded full-process execution trajectory data. Using the timeline as the axis, the execution record, corresponding role, scoring result, and deduction reason for each checkpoint are bound one by one to generate replayable review data with a complete timeline.

[0042] The review data supports full-process playback, and can synchronously display the scheduling decision process of the Host role, the task execution process of each Slot role, equipment status changes, video footage, evidence materials, etc., according to the timeline, to achieve a 1:1 restoration of the entire handling process.

[0043] Furthermore, this step also includes an automatic highlighting step for points lost, specifically: based on the dual-layer scoring results, automatically marking checkpoints or scheduling actions where the Slot layer score is below the first threshold or the Host layer score is below the second threshold; during playback, when the timeline reaches the corresponding point lost, automatically highlighting and displaying detailed information about that point lost, including the reason for the deduction, the violated contingency plan requirements, the expected behavior, and the actual execution status; at the same time, it supports one-click jump from the highlighted point lost to the corresponding contingency plan checkpoint definition page, making it convenient for users to quickly view the contingency plan requirements and make targeted optimizations.

[0044] The first threshold and the second threshold can be a preset standard deviation multiple of the historical average score. In this embodiment, the historical average score is taken minus 1 standard deviation. Alternatively, they can be percentiles preset by the user. In this embodiment, the 20th percentile is taken. That is, if the score is lower than the 20th percentile of all historical data, it is determined as a point loss.

[0045] S6: Automatically transfer the dual-level scoring results, contingency plan optimization parameters, and character capability profiles accumulated in the training scenarios to the actual combat scenarios; Specifically, in this step, the system will automatically accumulate all the data for each exercise scenario, including the two-layer scoring results, the pre-plan parameters optimized based on the scoring results, the weight coefficients of the scoring rules, the historical score data of personnel / agents corresponding to each slot role, generate the corresponding pre-plan optimization parameter package and role capability profile, and automatically synchronize them to the configuration library of the actual scenario.

[0046] During real-world scenarios, the Host role can directly access the optimized pre-plan parameters and scoring weights from the training scenario without reconfiguration. Simultaneously, based on the role's ability profile, more precise task allocation can be performed, assigning tasks at corresponding checkpoints to roles with higher historical scores and better matching abilities, achieving zero-loss transfer of experience accumulated in training to real-world scenarios. Furthermore, the dual-layer scoring results and pre-plan optimization suggestions generated in the real-world scenario are automatically fed back to the training scenario to update training level configurations and adjust training priorities, achieving a two-way knowledge loop iteration between training and real-world scenarios.

[0047] It also includes equipment linkage execution steps, specifically including the following processes: When the Host or Slot role needs to operate field equipment, it sends control commands to field equipment such as fans, valves, water pumps, fire doors, and elevators through the remote control module. The execution result of the control command is recorded as part of the evidence requirement and used for the action completion determination in the Slot layer scoring. The operation of the remote control module is fully traced. The traced data includes the operation initiator, operation time, target equipment, command parameters, and equipment return status. The traced data is written into a structured operation result object for post-event auditing.

[0048] When an IoT alarm is triggered, the Host role automatically invokes the security and fire protection linkage module to retrieve real-time video streams from nearby related cameras based on the alarm location. The footage is then stored and viewed in real time, and the stored video footage is used as evidence of the alarm in the execution trajectory data. The security and fire protection linkage module also supports proactive risk prediction based on video algorithms, identifying risks such as smoke and fire, passage blockage, and fire lane occupancy. The prediction results are injected as additional events into the Stage engine's EventInbox event queue. The prediction results carry a confidence score. When the confidence score is higher than a third threshold, an automatic escalation process is triggered. In this embodiment, the third threshold is set to 90%.

[0049] Please see Figure 2 This embodiment provides a safety emergency drill and practical collaborative review system based on a two-layer scoring system. The system includes a contingency plan configuration module, a Stage orchestration engine, an event access module, a scoring calculation module, a review and playback module, and a scenario package management module. The specific functions and implementation methods of each module are as follows: The emergency plan configuration module is used to transform safety emergency plans into structured plans based on checkpoints as basic units. It supports users to visually configure each field of the checkpoints. It also supports narrative interpolation based on the owner's actual situation to generate owner-specific plan instances. The configured structured plan is synchronously distributed to the Stage orchestration engine and scoring calculation module. The Stage orchestration engine, including the Host scheduling unit, Slot execution unit, and EventInbox unit, is used to execute collaborative processes with a fixed graph structure and to perform idempotent deduplication and deterministic batch processing of events. The Host scheduling unit, corresponding to the Host role, is used to receive events and complete the entire process scheduling, including event analysis, task allocation, process scheduling, escalation triggering, and closed-loop supervision, according to the structured plan. The Slot execution unit, corresponding to the Slot role, supports three access modes: NPC intelligent agent mode, human client mode, and duty robot subsystem mode. It is used to receive and execute tasks dispatched by the Host scheduling unit and return the execution results. EventInbox unit: Used to receive events injected by the event access module, perform idempotent deduplication and deterministic batch processing on the events, and submit them to the Host scheduling unit to ensure the uniqueness and deterministic order of event processing.

[0050] Event Access Module: Used to connect to external event sources, receive drill script events or real IoT alarm events, and inject the normalized events into the EventInbox unit of the Stage orchestration engine; supports connecting to the fire protection IoT platform via MQTT / HTTP protocol, and connecting to the third-party monitoring platform via REST / Webhook, to receive various alarm events.

[0051] The scoring calculation module automatically calculates the Slot layer behavior score and Host layer scheduling quality score based on the full-process execution trajectory data recorded by the Stage orchestration engine after the session loop ends. This module incorporates a weighted summation model and a multi-dimensional weighted model, along with an adaptive weight adjustment unit, supporting dynamic adjustment of scoring weights based on historical data. Furthermore, the two-layer scoring result generated by the scoring calculation module includes traceable fields, which at least include checkpoint identifiers, expected behavior descriptions, actual execution records, deduction reasons, and scheduling decision timestamps. The deduction reason record records specific violated constraints, including at least one of timeout duration, missing evidence type, and role mismatch details.

[0052] The debriefing and replay module is used to generate debriefing data with timelines and two-level scoring, and supports jumping from the highlighted points of deduction to the corresponding contingency plan checkpoint definition; The scenario package management module is used to store and load scenario packages from different security domains. Scenario packages include checkpoint definitions, role types, device protocol mappings, and scoring weight coefficients. The remote control module interfaces with the Stage orchestration engine to receive control commands from the Host scheduling unit or Slot execution unit. It then sends remote control commands to field equipment such as fans, valves, or pumps, and feeds the execution results back to the Stage orchestration engine. Simultaneously, it records complete operation log data for scoring and auditing. When a Host or Slot role needs to operate field equipment, it sends control commands to at least one of the fans, valves, or pumps. The execution result of the control command is recorded as part of the evidence requirement and used for action completion determination in Slot-level scoring. The remote control module's operations are fully logged, and the log data includes the operation initiator, operation time, target equipment, command parameters, and equipment return status. This log data is written to a structured operation result object for post-event auditing.

[0053] The security and fire protection linkage module interfaces with the Stage orchestration engine to receive invocation commands from the Host scheduling unit. It retrieves video streams from associated cameras based on the alarm location for image retention and real-time viewing. Simultaneously, it incorporates a video algorithm unit to support proactive risk identification and prediction, injecting the prediction results as events into the EventInbox unit. When an IoT alarm is triggered, the Host role automatically invokes the security and fire protection linkage module to retrieve video streams from associated cameras based on the alarm location for image retention and real-time viewing. The security and fire protection linkage module also supports proactive risk prediction based on video algorithms, injecting the prediction results as additional events into the Stage engine's event queue. The prediction results carry a confidence score; automatic escalation is triggered when the confidence score exceeds a third threshold.

[0054] The method and system of this invention first use a unified structured pre-plan as the knowledge core. Through the Stage orchestration engine, it switches between event sources (script events / real IoT alarms) and Slot role participation modes (NPC / human / duty robot), simultaneously driving drills and real-world scenarios within the same framework. This completely eliminates the configuration redundancy and conversion costs associated with maintaining two separate systems. Furthermore, it achieves idempotent deduplication and deterministic batch processing through the event-driven message queue EventInbox, simultaneously meeting the low-cost simulation requirements of drills and the high reliability and low latency requirements of real-world scenarios. It utilizes Host-layer scheduling quality and Slot-layer execution behavior. The separate, two-layer automatic scoring model divides the entire emergency response chain into a scheduling layer and an execution layer for separate quantitative evaluation. Through a fixed Start→Host→{Slots}→End directed graph topology, it achieves precise decomposition of scheduling and execution delays at the architectural level, solving the evaluation misalignment problem caused by traditional evaluations that only focus on frontline execution and neglect the scheduling center. Finally, by establishing a two-way knowledge loop between drills and real-world scenarios, the model automatically transfers the plan parameters and role capability profiles verified in drills to real-world scenarios, while simultaneously feeding back problems exposed in real-world scenarios to the drill system for iterative optimization. This enables the automatic accumulation, transfer, and iteration of emergency response experience. It achieves seamless collaboration and precise end-to-end evaluation between emergency drills and real-world scenarios, significantly reducing system development, deployment, and maintenance costs, and improving the reliability and consistency of emergency response.

[0055] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for collaborative review of safety emergency drills and real-world scenarios based on a two-level scoring system, characterized in that: Includes the following steps: S1: Transform the emergency plan into a structured plan with checkpoints as the basic unit. The checkpoints include action description fields, role division fields, completion time limit fields, escalation chain condition fields, and evidence requirement fields. S2: In the Stage orchestration engine, set up a Host role and at least one Slot role; the Host role is used to receive emergency events or drill events and generate task allocation decisions according to structured contingency plans; the Slot role is used to receive and execute the tasks assigned by the Host role and return the execution results. S3: During execution, record the response time, action completion rate, and evidence upload status of each Slot role to each checkpoint, as well as the task allocation time, upgrade trigger time, and role matching record of the Host role; S4: After execution, a two-layer automatic scoring is performed based on the recorded trajectory, including the following process: For each slot role, a weighted summation model is used to calculate the behavior score based on the expected behavior, time limit, and evidence requirements of the corresponding checkpoint in the structured plan. For the Host role, a multi-dimensional weighted model is used to calculate the scheduling quality score based on the timeliness of task allocation, the accuracy of role matching, and the compliance of upgrade chain execution. S5: Spatiotemporally correlate the results of the two-layer automatic scoring with the execution process to generate replayable debriefing data, which is used to show the collaboration process between the Host role and each Slot role and highlight the points lost.

2. The method for collaborative review of safety emergency drills and practical operations based on dual-level scoring as described in claim 1, characterized in that: The Stage orchestration engine is a directed graph topology. The directed graph topology includes an event injection node Start, a unique scheduling node Host, at least one execution node Slot, and a result aggregation node End. Directed edges are only allowed to point from the start node to the scheduling node, from the scheduling node to each execution node, and from each execution node to the end node, forming a deterministic execution path Start→Host→{Slots}→End. Tasks are processed in an orderly manner through an event-driven message queue.

3. The method for collaborative review of safety emergency drills and practical operations based on dual-level scoring as described in claim 2, characterized in that: The event-driven message queue is EventInbox. EventInbox performs idempotent deduplication and deterministic batch processing on each received event, so that the same event is processed only once and the order of events within the same batch is determined. The idempotent deduplication method generates a unique hash key based on the event source identifier and the event sequence number, and the deterministic batch processing uses a FIFO queue combined with a batch window timer.

4. The method for safety emergency drills and collaborative debriefing based on dual-level scoring as described in claim 1, characterized in that: The structured contingency plan is used in both drill scenarios and real-world scenarios; The Stage orchestration engine drives drill scenarios or real-world scenarios within the same framework by switching the participation modes of event sources and Slot roles. The event sources are script events in drill scenarios and real IoT alarms in real-world scenarios. The participation modes for the Slot role include: NPC mode, human-managed mode, and duty robot-managed mode.

5. The method for safety emergency drills and collaborative debriefing based on dual-level scoring as described in claim 1, characterized in that: S1 also includes narrative interpolation of the general contingency plan template based on the owner's actual personnel configuration, equipment parameters, and service level agreement (SLA) to generate owner-specific contingency plan instances.

6. The method for collaborative review of safety emergency drills and practical operations based on dual-level scoring as described in claim 1, characterized in that: In S4, a weighted summation model is used to calculate the behavior score. The formula for calculating the score of a single checkpoint is as follows: ； In the formula, The score for the i-th checkpoint. The actual response time at checkpoint i. The time limit requirement specified for checkpoint i in the structured contingency plan, For indicator functions, Let i be the number of sub-actions actually completed at checkpoint i. Let i be the total number of sub-actions expected to be completed at checkpoint i. The number of valid evidence types actually uploaded at checkpoint i. The number of evidence types required for checkpoint i, This is an indicator function.

7. The method for safety emergency drills and collaborative debriefing based on dual-level scoring as described in claim 1, characterized in that: In S4, the weighted summation model is used to calculate the behavior score, which also includes the following steps: If the Slot role fails to upload compliant evidence, the behavior score for the corresponding checkpoint is multiplied by a preset decay factor; if the checkpoint is critical and compliant evidence is not uploaded, the checkpoint score is directly determined to be zero.

8. The method for safety emergency drills and collaborative debriefing based on dual-level scoring as described in claim 1, characterized in that: In S4, a multi-dimensional weighted model is used to calculate the scheduling quality score. The calculation formula is as follows: ； In the formula, These are the weight coefficients for the four rating dimensions, and they satisfy... , The first distribution of delayed points, Score the character matching score. To upgrade the blockchain compliance score, The score is for closed-loop integrity.

9. The method for safety emergency drills and collaborative debriefing based on dual-level scoring as described in claim 1, characterized in that: In S5, automatically marking and highlighting points that have lost marks includes the following process: Set a first threshold and a second threshold. The first threshold is the score threshold for the Slot layer, and the second threshold is the score threshold for the Host layer. Automatically mark checkpoints or scheduling actions where the Slot layer score is below the first threshold or the Host layer score is below the second threshold; During the replay, when the timeline reaches the corresponding point of deduction, the detailed information of that point of deduction is automatically highlighted, including the reason for the deduction, the violation of the plan requirements, the expected behavior, and the actual implementation.

10. A safety emergency drill and practical collaborative debriefing system based on dual-level scoring, used to implement the safety emergency drill and practical collaborative debriefing method based on dual-level scoring as described in any one of claims 1-9, characterized in that: include: The contingency plan configuration module is used to transform safety emergency plans into a structured set of checkpoints and supports narrative interpolation to generate custom contingency plan instances. The Stage orchestration engine includes a Host scheduling unit and a Slot execution unit for executing collaborative processes with a fixed graph structure; the Stage orchestration engine also includes an EventInbox unit for idempotent deduplication and deterministic batch processing of events. The event access module is used to receive drill script events or real IoT alarms and inject them into the EventInbox unit; The scoring calculation module is used to automatically calculate the behavior score of the Slot layer and the scheduling quality score of the Host layer after execution. The scoring calculation module has a built-in weighted summation model and an adaptive weight adjustment unit. The debriefing and replay module is used to generate debriefing data with timelines and two-level scoring, and supports jumping from the highlighted points of deduction to the corresponding contingency plan checkpoint definition; The scenario package management module is used to store and load scenario packages from different security domains. The scenario packages include checkpoint definitions, role types, device protocol mappings, and scoring weight coefficients.

Images