Fast fitting method, system, and media of a fire fighter body model with safety score

Abstract

The application discloses a kind of quick adaptation method, system and medium of fire-fighting body model with safety score, wherein, based on the quick adaptation method of fire-fighting body model with safety score of the embodiment of the application, with the unified fire-fighting body data specification as base, real data acquisition and simulation synthesis data are mapped to the same field system, the risk form that real fire scene / hazard scene is difficult to cover is filled in with fire domain randomization and sampling distribution adaptive strategy, with parameter efficient fine tuning (PEFT), new scene adaptation converges to a small number of trainable parameters and forms manageable scene adaptation, and the safety requirement of fire-fighting is quantified as online criterion with automatic evaluation and safety score threshold. Thus, support the continuous improvement of fire-fighting body model in multiple scene long-term operation and reduce iteration cost.

Description

Technical Field

[0001] This invention relates to the field of fire safety and emergency robot technology, and in particular to a rapid adaptation method for a fire-fighting body model with safety scoring, a computer-readable storage medium, and a rapid adaptation system for a fire-fighting body model with safety scoring. Background Technology

[0002] In high-risk environments such as chemical plant areas, warehousing parks, power stations, subways, and tunnels, embodied intelligent systems for inspection and early response need to operate long-term and maintain stable performance under different scenarios and risk profiles. Due to the hazardous and costly nature of data collection from real fire scenes and high-risk hazard scenarios, the industry typically employs a multi-source parallel approach to construct training sets, including routine inspection data, drill data, accident debriefing data, and simulated synthetic data. As model size increases, the training and adaptation process has gradually evolved from offline training in a single scenario to a cross-scenario, continuously iterative engineering process, with data standardization, synthetic data strategies, fine-tuning methods, automatic evaluation, and release gating becoming core components.

[0003] However, the problems with these technologies are: 1) The lack of unified data standards for multiple risk factors in fire protection. For example, the lack of unified standards makes it difficult to reuse and align data across scenarios, data versions are untraceable, regression sets are difficult to maintain stably, and thus it is difficult to form a sustainable data flywheel. 2) The risk distribution of simulated synthetic data is inconsistent with that of real fire protection scenarios. For example, without modeling and sampling strategies oriented towards risk distribution, even large-scale synthetic data may not cover critical failure modes, resulting in good performance in simulation evaluations but significant performance degradation in real environments. 3) Fine-tuning and deployment lack safety scoring thresholds and automatic regression loops. For example, the lack of auditable thresholds leads to model releases relying on human experience, making version rollbacks and accountability difficult, and also hindering large-scale iteration. 4) Cross-scenario adaptation cycles are long and prone to catastrophic amnesia. For example, fine-tuning for new scenarios may damage the capabilities of old scenarios, and the lack of systematic version management, regression set construction, and automated evaluation gating mechanisms will further amplify regression risks. Summary of the Invention

[0004] This invention aims to at least partially address one of the technical problems in related technologies. Therefore, the first objective of this invention is to propose a rapid adaptation method for a fire-fighting vehicle model with a safety score, enabling rapid adaptation of the fire-fighting vehicle model in multiple scenarios, supporting the continuous improvement capability of the fire-fighting vehicle model during long-term operation in multiple scenarios, and reducing iteration costs.

[0005] A second objective of this invention is to provide a computer-readable storage medium.

[0006] The third objective of this invention is to provide a rapid adaptation system for fire-fighting body models with safety scoring.

[0007] To achieve the above objectives, the first aspect of this invention proposes a rapid adaptation method for a fire safety performance model with safety scoring, comprising: constructing a fire safety performance data specification, wherein the fire safety performance data specification is organized in rounds, and each round consists of time-synchronized multimodal observation sequences, action / skill sequences, risk labels, safety-related labels, evidence indexes, and scene metadata, wherein the evidence index is used to bind conclusions with verifiable evidence, each evidence entry is bound to a content hash, and a versioning binding mechanism is established to link training / evaluation / release with dataset versions, regression set versions, and other relevant data. The randomization configuration version, model base version, scene adapter version, and evaluation protocol version are bound together; simulation synthetic data that meets the fire protection data specifications is generated, and fire protection domain randomization configuration is embedded throughout the initialization, sampling, label generation, and evidence generation processes of the simulation synthetic data. The fire protection domain randomization configuration includes smoke / visibility randomization, thermal background drift and thermal reflection pseudo-hotspot randomization, gas diffusion and sensor hysteresis drift randomization, structural accessibility change randomization, ground slip / adhesion change randomization, and communication degradation and latency randomization. The simulation-synthesized data is synthesized and aligned with the real-world collected data. Based on the risk pattern distribution obtained from online collected data or real-machine collected data, the randomized sampling distribution of the next round of synthesis is weighted and updated to ensure that the synthesized data prioritizes coverage of real high-risk boundaries and historical failure patterns. Training / validation / regression sets for the fire-fighting avatar model are constructed. The regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples. Multiple scenario adapters for the fire-fighting avatar model are formed based on the PEFT adaptation path. A scenario router selects the corresponding scenario adapter based on scenario metadata and risk pattern characteristics, enabling rapid switching between multiple scenarios, gray-scale deployment, and one-click rollback. The fire-fighting avatar model is given a safety score. If the fire-fighting avatar model fails the safety score or triggers the red line condition, a structured data supplementation requirement record is automatically generated. Based on the structured data supplementation requirement record, a training flywheel closed loop is performed on the fire-fighting avatar model. The fire-fighting avatar model is released online after passing the safety score and not triggering the red line condition.

[0008] According to embodiments of the present invention, a rapid adaptation method for fire-fighting avatar models with safety scoring is based on a unified fire-fighting avatar data standard. It maps real-world collected data and simulated synthetic data to the same field system. It uses fire domain randomization and adaptive sampling distribution strategies to supplement risk patterns that are difficult to cover in real fire / hazard scenarios. It uses parameter efficient fine-tuning (PEFT) to converge new scenario adaptations to a small number of trainable parameters, forming manageable scenario adaptable models. It uses automatic assessment and safety scoring thresholds to quantify fire safety requirements into deployable criteria. Thus, by establishing a reusable, alignable, and traceable unified standard for cross-scenario multimodal data and safety elements, and constructing a simulated synthetic data generation mechanism with fire domain randomization, the risk patterns, observation distributions, and action distributions are made closer to real-world scenarios. Rapid adaptation is achieved without significantly altering the base model, transforming cross-scenario migration from a long-cycle manual process into a streamlined, lightweight iteration. Simultaneously, the introduction of automatic assessment and safety scoring thresholds enables automated regression, auditable, and secure release processes. Therefore, it supports the continuous improvement of the fire-fighting avatar model's capabilities and reduces iteration costs during long-term operation in multiple scenarios.

[0009] In addition, the rapid adaptation method for the fire-fighting suit model with safety scoring according to the above embodiments of the present invention may also have the following additional technical features: According to one embodiment of the present invention, each evidence entry is bound to a content hash, including: generating a content hash for each evidence entry and referencing it in the evaluation report.

[0010] According to one embodiment of the present invention, the step of synthesizing and aligning the simulated synthetic data with the real collected data further includes: for missing modalities or incomplete rounds, explicitly marking the missing data in the fire protection data specification, and incorporating deduction or blocking rules into the subsequent evidence integrity scoring and threshold determination.

[0011] According to one embodiment of the present invention, the safety score is composed of a heat exposure sub-score, a toxic gas or combustible gas exposure sub-score, a near miss event sub-score, a risk boundary violation sub-score, a false alarm / missed alarm cost sub-score, an evidence integrity sub-score, and a mission success rate sub-score. The red line conditions include cumulative heat exposure exceeding the upper limit or the duration of exceeding the limit being too long, cumulative toxic gas or combustible gas exposure exceeding the upper limit or the peak exceeding the upper limit, the occurrence of a high-level fall / instability hazard or a high-level collision hazard, the occurrence of a high-risk missed alarm reaching a preset number of times, and the lack of necessary evidence for key high-risk alarms, resulting in unverifiable results.

[0012] According to one embodiment of the present invention, the structured data supplementation requirement record is associated with at least the failed rounds, failure risk factors, failure modes, target coverage intervals, and required sample size and priority, and serves as input for the next round of data collection plan and randomized distribution update.

[0013] According to one embodiment of the present invention, each adapter entry is bound to an adapter version and its corresponding dataset version, regression set version, randomization configuration version, evaluation protocol version, training recipe summary and evaluation result summary, and is associated with a security score report and evidence index hash.

[0014] According to one embodiment of the present invention, the scene router selects the corresponding scene adapter based on smoke level, light level, thermal background shift, gas change rate, slipperiness level and communication quality level.

[0015] According to one embodiment of the present invention, during the gray-scale deployment phase, the scenario router routes a small number of rounds to the new adapter according to a preset ratio and continuously monitors the security score and red-line events. When a red-line event is triggered, the score drops significantly, or the number of key risk misses increases, the router quickly switches to the previous stable adapter and retains the evidence index for review and supplementary data generation.

[0016] To achieve the above objectives, a computer-readable storage medium is provided in the second aspect of the present invention, which stores a rapid adaptation program for a fire-fighting suit model with safety scores. When the rapid adaptation program for the fire-fighting suit model with safety scores is executed by a processor, it implements the rapid adaptation method for the fire-fighting suit model with safety scores described in the embodiments of the present invention.

[0017] According to embodiments of the present invention, a computer-readable storage medium can achieve rapid adaptation of a fire-fighting body model with safety scores stored thereon by executing a rapid adaptation program, thereby enabling the fire-fighting body model to be rapidly adapted in multiple scenarios, supporting the continuous improvement capability of the fire-fighting body model in long-term operation in multiple scenarios and reducing iteration costs.

[0018] To achieve the above objectives, the third aspect of this invention proposes a rapid adaptation system for a fire safety performance model with safety scoring, comprising: a specification construction module for constructing fire safety performance data specifications, wherein the fire safety performance data specifications are organized in rounds, and each round consists of time-synchronized multimodal observation sequences, action / skill sequences, risk labels, safety-related labels, evidence indexes, and scene metadata, wherein the evidence index is used to bind conclusions with verifiable evidence, each evidence entry is bound to a content hash, and a versioning binding mechanism is established to ensure that training / evaluation / release is linked to dataset versions, regression set versions, and synthesis versions. The system binds the configuration version, model base version, scene adapter version, and evaluation protocol version; the data generation module is used to generate simulation synthetic data that meets the fire protection data specifications, and embeds fire protection domain randomization configuration throughout the initialization, sampling, label generation, and evidence generation processes of the simulation synthetic data generation. The fire protection domain randomization configuration includes smoke / visibility randomization, thermal background drift and thermal reflection pseudo-hotspot randomization, gas diffusion and sensor hysteresis drift randomization, structural accessibility change randomization, ground slip / adhesion change randomization, and communication degradation and latency randomization; the data processing module… This module is used to synthesize and align the simulated synthetic data with the real-world collected data. Based on the risk pattern distribution statistically obtained from online or real-world collected data, it weights and updates the randomized sampling distribution for the next round of synthesis, ensuring that the synthesized data prioritizes coverage of real high-risk boundaries and historical failure patterns. It also constructs a training / validation / regression set for the fire-fighting stunt model. The regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples. A rapid adaptation module is used to form the fire-fighting stunt model based on PEFT-adapted paths. Multiple scene adapters are available, and the scene router selects the corresponding scene adapter based on scene metadata and risk characteristics, enabling rapid switching between multiple scenes, gray-scale deployment, and one-click rollback. The safety scoring module is used to score the safety of the fire-fighting vehicle model. If the fire-fighting vehicle model fails the safety score or triggers the red line condition, a structured data supplementation requirement record is automatically generated. Based on the structured data supplementation requirement record, a training flywheel closed loop is performed on the fire-fighting vehicle model. After the fire-fighting vehicle model passes the safety score and does not trigger the red line condition, the fire-fighting vehicle model is released online.

[0019] The rapid adaptation system for fire-fighting avatar models with safety scoring, as described in this embodiment of the invention, is based on a unified fire-fighting avatar data standard. It maps real-world collected data and simulated synthetic data to the same field system. It uses fire domain randomization and adaptive sampling distribution strategies to supplement risk patterns that are difficult to cover in real fire / hazard scenarios. It uses parameter efficient fine-tuning (PEFT) to converge new scenario adaptations to a small number of trainable parameters, forming manageable scenario adaptable models. It quantifies fire safety requirements into deployable criteria using automatic assessment and safety scoring thresholds. Thus, by establishing a reusable, alignable, and traceable unified standard for cross-scenario multimodal data and safety elements, and constructing a simulated synthetic data generation mechanism with fire domain randomization, the system makes risk patterns, observation distributions, and action distributions closer to real-world scenarios. It achieves rapid adaptation without significantly altering the base model, transforming cross-scenario migration from a long-cycle manual process into a streamlined, lightweight iteration. Simultaneously, the introduction of automatic assessment and safety scoring thresholds enables automated regression, auditability, and secure controllable release processes. Therefore, it supports the continuous improvement of the fire-fighting avatar model's capabilities and reduces iteration costs during long-term operation across multiple scenarios.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a rapid adaptation method for a fire-fighting body model with safety scoring according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the field set of the fire-fighting equipment data specification according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the process of generating synthetic data and randomizing fire zones according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating the automatic assessment and security scoring process according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the training flywheel closed loop according to an embodiment of the present invention; Figure 6 This is a block diagram of a rapid adaptation system for a fire-fighting suit model with safety scoring according to an embodiment of the present invention. Detailed Implementation

[0022] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0023] The following describes, with reference to the accompanying drawings, a rapid adaptation method for a fire-fighting suit model with safety scoring, a computer-readable storage medium, and a rapid adaptation system for a fire-fighting suit model with safety scoring, according to embodiments of the present invention.

[0024] Figure 1 This is a flowchart illustrating a rapid adaptation method for a fire-fighting suit model with safety scoring according to an embodiment of the present invention.

[0025] Specifically, such as Figure 1 As shown, in some embodiments of the present invention, a rapid adaptation method for a fire-fighting suit model with safety scoring includes: S101 establishes a fire protection embodied data specification. The fire protection embodied data specification is organized in rounds. Each round consists of time-synchronized multimodal observation sequences, action / skill sequences, risk labels, safety-related labels, evidence indexes, and scene metadata. The evidence index is used to bind conclusions with verifiable evidence. Each evidence entry is bound to a content hash, and a versioning binding mechanism is established to bind training / evaluation / release to dataset versions, regression set versions, synthetic randomization configuration versions, model base versions, scene adapter versions, and evaluation protocol versions.

[0026] Specifically, in this embodiment of the invention, such as Figure 2 As shown, the data structure for each round is written as a serializable object, containing the following set of fields: Scene metadata: scene number, scene type, map, task type, risk boundary configuration, environmental conditions, etc.; Robot and sensor configuration: robot model, kinematic parameter version, control frequency, sensor list, intrinsic / extrinsic parameters of each sensor, time synchronization method and alignment error statistics.

[0027] Multimodal observation data: visible light images, thermal imaging images or temperature matrices, point clouds, gas sensor readings, IMU, organism status, with recording units and calibrations for each mode.

[0028] Actions / Skills: Low-level controls (speed, gait parameters, joint targets, etc.), skill-level actions (approaching valves, reading meters, opening cabinet doors, injection suppression, retreating, etc.) and their parameters (target point, target object number, injection volume, time budget, etc.). When using a large embodied model, both input commands and model output action selection results are recorded to support offline playback and consistency regression.

[0029] Risk labels describe environmental risk factors and boundary conditions, using discrete levels and providing level definitions, including smoke obscuration level, visibility level, thermal background drift level, reflectivity / strong light level, gas concentration level, structural accessibility level, ground adhesion / slippery level, communication quality level, etc.

[0030] Safety Labels: Describe the quantitative results of safety and dangerous events, including thermal exposure dose, toxic / combustible gas exposure dose, risk boundary violation, near miss events such as collision / fall / slip, false alarm / missed alarm events and their cost levels, and whether the mission has triggered safety action indicators such as "safe stop / retreat".

[0031] Evidence Index: This index links conclusions with verifiable evidence and includes evidence entry number, corresponding round number, corresponding time window, keyframe / key point cloud fragment reference, thermal image ROI, gas curve window, and associated risk / safety label number. In some embodiments of this invention, a content hash is generated for each evidence entry and referenced in the evaluation report, thereby preventing post-audit tampering and ensuring audit consistency.

[0032] Specifically, the Firefighting Embodied Data Specification is organized in rounds. Each round consists of a time-synchronized multimodal observation sequence, action / skill sequence, risk / safety tags, safety-related tags, evidence index, and scene metadata. To ensure reproducibility, each round must simultaneously record: sensor timestamp sources, coordinate system domain calibration parameters, sampling frequency and frame loss, as well as a configuration summary and random seed for the generation / acquisition process. Furthermore, the Firefighting Embodied Data Specification is compatible with the round organization method of the Reinforcement Learning Data Module (RLDS) to facilitate integration with the general robot learning ecosystem. Additionally, a fire safety element field is introduced to address the issues of quantifiable and auditable fire scene risks.

[0033] Furthermore, in this embodiment of the invention, a versioning binding mechanism is established to bind training / evaluation / release with dataset version, regression set version, synthetic randomization configuration version, model base version, scenario adapter version, and evaluation protocol version. A manifest file is output for each training iteration, listing the sample number, evidence index hash, random seed, data generation configuration summary, and data file hash. This enables third parties to reproduce experimental results and threshold determinations with the same input files, thereby ensuring reproducibility and traceability.

[0034] S102 generates simulated synthetic data that meets the fire protection data specifications, and embeds fire protection domain randomization configuration throughout the entire process of initialization, sampling, label generation, and evidence generation of the simulated synthetic data. The fire protection domain randomization configuration includes smoke / visibility randomization, thermal background drift and thermal reflection pseudo-hotspot randomization, gas diffusion and sensor hysteresis drift randomization, structural accessibility change randomization, ground slip / adhesion change randomization, and communication degradation and latency randomization.

[0035] The following is combined Figure 3 The process of generating simulation synthetic data that meets the fire protection equipment data specifications according to specific embodiments of the present invention will be described below: 1) Construction of simulation scenarios and asset library Before generating the simulation composite data, a fire scene asset library is first established and the semantic and physical modeling of the scene is completed: The asset library contains at least two categories: structural assets, which describe spatial structures such as pipe corridors, warehouses, power distribution rooms, platforms / tunnels, including walls, floors, passageways, steps, doorways, and railings; and object assets, which describe objects that need to be identified or interacted with during inspection and handling, such as valves, meters, cabinet doors, cable trays, fire extinguishers, stacks of combustibles, warning signs, and debris. To ensure usability for ground truth annotation, each asset is bound to a semantic category identifier, interactive attributes (such as openable / rotatable / grabable / non-interactive), and physical attributes (such as mass, coefficient of friction, elasticity, and collider shape) upon import.

[0036] The simulation scene can be obtained from BIM / CAD, point cloud reconstruction or procedural generation. After being imported into the simulation, it is converted into a mesh and collision body that the simulation engine can recognize, and structured annotation information is added to the key areas, including the representation of passable areas, restricted areas / hazardous source areas and key checkpoints. The above structure and annotation together form a versionable scene version for subsequent data writing and audit reproduction.

[0037] 2) Fire risk site modeling First, a smoke / visibility field is constructed, and the smoke density or extinction coefficient is expressed in the scene space using voxel meshes, layered volumes, or equivalent methods. This field is then updated on the time axis based on the source term and wind field. During the rendering stage, the visible light imaging is simulated for attenuation and scattering based on the smoke field, thereby obtaining a low visibility image consistent with the smoke. At the same time, the visibility level or occlusion ratio is output as a risk label to facilitate alignment with real data statistics.

[0038] Secondly, a thermal field and heat source model is constructed, defining the temperature change curve and spatial diffusion range of the heat source over time, and defining the reference temperature and emissivity for ordinary objects. Specifically, during thermal image generation, two types of results are output simultaneously: one is the true temperature field, used for subsequent exposure dose calculation and safety label generation; the other is the thermal image sensor output, used as training and inference input. This process introduces phenomena such as range truncation, noise, background drift, and thermal reflection pseudo-hot spots. The range and truncation rules, noise parameters, and drift parameters are written into the configuration file as part of the randomization configuration to ensure that the generation of thermal image data is reproducible and auditable.

[0039] Furthermore, a gas concentration field and sensor response model is constructed. Gas concentration is expressed in the scene space using voxel meshes or parameterization, and updated over time through wind convection, diffusion, and source term combinations. Considering the hysteresis and drift inherent in real gas sensors under rapid changes, a dynamic response model is introduced when generating sensor readings from the true concentration value. The true concentration is processed by a first-order system with a time constant to generate the reading, and drift and noise are superimposed. The time constant, drift rate, and noise level are all configurable parameters written into the configuration file. Therefore, based on this modeling approach, the phenomenon of rapid concentration increases but lagging readings can be stably reproduced in gas simulations, making training and threshold evaluation closer to real-world conditions.

[0040] 3) Task script and expert action sequence generation Establish a task template library to describe typical fire inspection and response processes. Each task template defines key stages and termination conditions using a state machine or equivalent structure. When generating expert actions, global path planning is performed at the navigation level based on the representation of passable areas, combined with local obstacle avoidance output speed command sequences.

[0041] Specifically, for a quadruped robot, actions can be represented as a sequence of downward velocities changing over time. For example, at the interaction level, scripted action sequences or inverse kinematics and predefined pose libraries can be used to generate joint targets. At the same time, the action generation configuration, random seed, and version are all written into the action list, thereby ensuring the traceability of the training sample source and generation process.

[0042] 4) Run the simulation step by step and collect multimodal data. First, during the initialization phase, parameters for this round are sampled according to the configuration file, including scene layout and object position, smoke / heat / gas and wind field parameters, lighting and material parameters, ground friction and slipperiness, and communication degradation event scripts, etc. The sampling results and random seeds are then fixed into the metadata. Subsequently, the robot's initial pose is set according to the task template, or randomly sampled according to rules within the passable area to ensure that the starting conditions can be reproduced.

[0043] Subsequently, the simulation advances the physics solution with a control step size Δt. At each time step, it receives expert actions and drives the robot's movement, updates the smoke field, thermal field, gas field, and wind field, performs collision detection, contact force calculation, and dynamic integration, and outputs data such as RGB, thermal images, depth / point cloud, gas readings, IMU, and robot status at each sensor sampling time. Simultaneously, safety-related events are recorded in real time, with all events timestamped and graded, providing initial evidence for subsequent safety scoring and evidence indexing.

[0044] 5) Automatically generate tags and evidence indexes Specifically, risk and safety labels are automatically generated during the generation process, including smoke level / visibility level, heat exposure dose, gas exposure dose, and near miss event level. Furthermore, when it is necessary to quantify false alarms / missed alarms, the false alarm / missed alarm events and their risk levels can be obtained by comparing the object's true value with the model output, thus supporting subsequent safety scoring thresholds. Moreover, to ensure that the threshold determination is verifiable and auditable, the generation rules for the evidence index are further defined: for each alarm, anomaly, or red-line event, an evidence entry is generated at the time of its occurrence, and an evidence time window is captured. The evidence entry references keyframes, key point cloud fragments, thermal image ROIs, gas curve windows, and associated labels and event levels to support subsequent tamper-proof verification and consistency auditing.

[0045] Furthermore, in the above embodiments of the present invention, as Figure 3 As shown, the randomization configuration of the fire protection domain is no longer a matter of adding noise to the image after generation, but rather embedded in the entire process of initialization, sampling, and label / evidence generation. Specifically, the randomization configuration of the fire protection domain includes randomization of smoke / visibility, randomization of thermal background drift and thermal reflection pseudo-hotspots, randomization of gas diffusion and sensor hysteresis drift, randomization of changes in structural accessibility, randomization of changes in ground slipperiness / adhesion, and randomization of communication degradation and latency, thereby ensuring full risk coverage. Each type of configuration is described by distribution, range, and segment weight, providing an input basis for subsequent adaptive updates of the sampling distribution based on online statistics.

[0046] It should be noted that randomization parameters can be sampled and solidified according to the configuration file during the round initialization stage. Then, during the generation process, the sensor output and event triggering are affected by the evolution of risk fields such as smoke / heat / gas / communication. In the tag and evidence indexing stage, corresponding tags and evidence are generated from truth values ​​and event criteria, thereby ensuring the consistency between randomization parameters and data, tags, and evidence.

[0047] S103 synthesizes and aligns the simulated synthetic data with the real collected data, and based on the risk pattern distribution obtained from the online collected data or the real machine collected data, it performs a weighted update on the randomized sampling distribution of the next round of synthesis, so that the synthetic data prioritizes the coverage of real high-risk boundaries and historical failure modes, and constructs a training / validation / regression set for the fire protection body model. The regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples.

[0048] The synthesis and alignment of simulated synthetic data and real-world collected data in this invention will be described below with reference to specific embodiments of the invention: Time synchronization and coordinate calibration alignment: 1) Real data records on the edge side use a unified clock timestamp and retain the trigger time and synchronization method of each sensor; 2) When aligning on the cloud side, alignment error statistics are written into the metadata as a quality indicator. The coordinate system includes the robot base, camera, point cloud, and world coordinate system, and the extrinsic parameter matrix is ​​versioned and archived to ensure consistent semantics for subsequent alignment, target localization, and risk boundary determination; 3) The simulation side output uses the same coordinate system convention and similar calibration fields to avoid implicit deviations such as fields with the same name but inconsistent meanings.

[0049] Modal field consistency and unit consistency: Key modalities such as thermal imaging and gas are unified into engineering units. In some embodiments of the present invention, missing modalities or incomplete rounds are explicitly marked as missing in the fire protection data specification, and deduction or blocking rules are included in the subsequent evidence integrity scoring and threshold determination, thereby incorporating the risk of missing data into an auditable technical link rather than relying on manual experience for processing.

[0050] Key sensor response consistency: To reduce the gap between simulation and reality in sensor response, a configurable response mapping is introduced on the generated simulation synthetic data side, making the synthetic data closer to reality in terms of dynamic changes, noise, drift, and truncation. For example, range truncation, noise, background drift, and thermal reflection pseudo-hotspot injection are introduced into thermal imaging data; hysteresis, drift, and noise are introduced into gas data; and low-light noise, compression artifacts, and motion blur are introduced into visual data. The parameters of the above response models are versioned with the configuration file and recorded in the evaluation report, thus ensuring reproducibility with the same configuration and comparability between different configurations.

[0051] Consistency of risk / safety labels: By discretizing key risk forms into level labels, such as smoke level, visibility level, slipperiness level, etc., and providing operational definitions for each level, synthetic labels and real labels can be statistically aligned under unified rules.

[0052] Distribution Alignment: Based on the risk pattern distribution obtained from online data collection or real-world data collection (e.g., smoke level histogram, thermal background offset distribution, gas peak and rate of change distribution, communication degradation ratio, and historical failure mode list), the next round of synthetic randomized sampling distribution is weighted and updated to ensure that the synthetic data prioritizes coverage of real high-risk boundaries and historical failure modes.

[0053] Optionally, reproducibility can be ensured by incorporating the configuration, version, and quality metrics of the alignment and mapping process described above into the data inventory.

[0054] Furthermore, in this embodiment of the invention, the data is divided into a training set, a validation set, and a regression set. The regression set is used to lock in key scenarios and key failure modes over the long term. For example, the regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples. It is maintained through a sample list and content hash index for versioning, thereby ensuring that new versions pass security thresholds without compromising the capabilities of older scenarios. Each training round outputs a data list file and records the sample source and evidence index references, thus providing a clear basis and traceability for the expansion and changes of the regression set.

[0055] S104 is a multi-scenario adapter for fire protection body model based on PEFT adaptation path. It selects the corresponding scenario adapter according to scenario metadata and risk characteristics through scenario router, so as to realize fast switching between multiple scenarios, gray-scale deployment and one-click rollback.

[0056] The following describes the PEFT adaptation path and scene adapter selection process of the present invention with reference to specific embodiments: Constraints on the Location and Scale of Adaptive Parameter Injection: Trainable parameters are injected into predetermined modules of the base model using LoRA and Adapter. Injection locations include the interaction layer for visual feature and language / command fusion, as well as the action output head or skill selection head. Optionally, adaptive parameters can be injected into auxiliary heads for safety-related judgments to enhance risk boundary identification and uncertainty estimation capabilities. To ensure rapid adaptation, the scale of trainable parameters is constrained to not exceed a preset proportion of the base model's parameter count. The injection layer list, LoRA rank, scaling factor, and other recipe information are written into the adaptation recipe and versioned for archiving, ensuring the comparability and reproducibility of different adaptors.

[0057] Training data allocation and safety-oriented objectives: Training employs a mixture of real and aligned synthetic data, with increased sampling weights or corresponding loss weights for high-risk boundary samples. This allows the model to prioritize stability under conditions such as smoke obscuration, thermal reflection spurious hotspots, gas hysteresis, slippery and narrow passages, and communication degradation. Optionally, training objectives can simultaneously cover consistency in instruction understanding and skill selection, imitation of motion trajectories or speed commands, and auxiliary objectives related to safety constraints. The weights and allocation parameters are recorded in the data list as training logs and version summaries, ensuring interpretability and reproducibility of the training process.

[0058] Furthermore, in this embodiment of the invention, to support multi-scenario operation, a scene router selects the corresponding adapter based on scene metadata and risk pattern characteristics, or performs fusion inference among multiple adapters. In some embodiments of the invention, the scene router selects the corresponding scene adapter based on smoke level, light level, thermal background shift, gas change rate, slipperiness level, and communication quality level. It should be noted that the routing strategy can employ rule-based routing to ensure interpretability and auditability, or a lightweight classifier or retrieval unit to adapt to more scene variations. Uncertainty gating can also be incorporated; when a new scene with high uncertainty or a risk pattern exceeding the coverage area is detected, a conservative adapter is preferentially selected or a security policy is triggered.

[0059] It should be noted that, in the above embodiments of the present invention, multiple scenario adapter versions can also be managed by establishing an adapter library. In some embodiments of the present invention, each adapter entry is bound to an adapter version and its corresponding dataset version, regression set version, randomization configuration version, evaluation protocol version, training recipe summary, and evaluation result summary, and is associated with a security score report and evidence index hash, thereby enabling any deployed version to be traced back to its data and evaluation basis. Furthermore, in some embodiments of the present invention, the adapter library supports rollback capability; that is, during the canary deployment phase, the scenario router routes a small number of rounds to the new adapter according to a preset ratio and continuously monitors the security score and red-line events. When a red-line event is triggered, the score drops significantly, or the number of key risk misses increases, it quickly switches to the previous stable adapter and retains the evidence index for review and supplementary data generation.

[0060] S105 performs a safety score on the fire-fighting body model. If the fire-fighting body model fails the safety score or triggers the red line conditions, a structured data supplementation requirement record is automatically generated. Based on the structured data supplementation requirement record, a training flywheel closed loop is performed on the fire-fighting body model. After the fire-fighting body model passes the safety score and does not trigger the red line conditions, the fire-fighting body model is released online.

[0061] Specifically, in some embodiments of the present invention, the safety score consists of a heat exposure sub-score, a toxic gas or combustible gas exposure sub-score, a near miss event sub-score, a risk boundary violation sub-score, a false alarm / missed alarm cost sub-score, an evidence integrity sub-score, and a mission success rate sub-score.

[0062] Specifically, in terms of safety scoring, fire safety requirements are quantified into auditable upper limits. For example, the safety score is defined in the range of 0 to 100 and is composed of multiple sub-scores (e.g., heat exposure sub-score, toxic or combustible gas exposure sub-score, near miss event sub-score, risk boundary violation sub-score, false alarm / missed alarm cost sub-score, evidence integrity sub-score, and mission success rate sub-score).

[0063] Among them, the thermal exposure sub-score can calculate the cumulative exposure exceeding the threshold based on thermal imaging or true temperature sequences and perform normalization processing; the toxic gas / combustible gas exposure sub-score can calculate the cumulative exposure exceeding the threshold and peak risk based on gas readings and consider the hysteresis caused by delays; the near miss event sub-score deducts points based on the level and frequency of events such as collisions, slipping, falling tendencies, and boundary crossings; the risk boundary violation sub-score is based on the number and duration of entering restricted areas or approaching hazards without a safe distance; the false alarm / missed alarm cost sub-score weights false alarms and missed alarms according to risk levels and imposes stronger penalties on high-risk missed alarms; the evidence integrity sub-score is based on whether key alarms have the prescribed modal evidence and whether the time alignment error is within the allowable range; and the task success rate sub-score reflects the completion of closed-loop tasks and the timeout / interruption situation.

[0064] Optionally, the total security score can be calculated using a weighted aggregation method, and the weights and normalization rules can be archived with the versioning of the evaluation protocol to ensure reproducibility.

[0065] The safety scoring process and the release process of the fire-fighting suit model are described below with reference to specific embodiments of the present invention: Specifically, open-loop and closed-loop assessments are conducted on the fire-fighting stunt model, and structured assessment reports and evidence indexes are output. The open-loop assessment includes the evaluation of the accuracy indicators of hazard or anomaly detection, the evaluation of false alarm and false negative rates, the evaluation of the correctness of instruction understanding, the consistency of skill selection, and the evaluation of the accuracy of risk boundary identification. The closed-loop assessment includes the evaluation of task completion rate under dynamic environment and interactive conditions, the evaluation of robustness and resilience, and the evaluation of the safety behavior performance under conditions of path deviation, collision count, slip / fall near loss, retreat strategy trigger count, and communication degradation.

[0066] Specifically, such as Figure 4As shown, the open-loop assessment focuses on offline indicators for perception and decision-making, including accuracy indicators for hazard or anomaly detection, false alarm rate and false negative rate, instruction comprehension accuracy, skill selection consistency, and risk boundary identification accuracy. False negatives are weighted by risk level, and the report provides both the overall false negative rate and the high-risk false negative rate to match the high-risk cost characteristics of fire protection applications. The closed-loop assessment includes simulated closed-loop evaluation, used to measure task completion rate, robustness, and recovery capability under dynamic environmental and interactive conditions. It also records path deviation, collision count, near misses due to slip / fall, retreat strategy trigger count, and safety behavior under communication degradation conditions on the mobile platform.

[0067] It should be noted that after conducting open-loop and closed-loop assessments of the fire-fighting body model, the threshold determination is made verifiable and traceable by outputting a structured assessment report and evidence index. In addition, to ensure reproducibility, the assessment protocol solidifies the list of assessment scenarios, random seed set, termination conditions, success criteria and hazard criteria, and incorporates the assessment configuration and results into the versioned record.

[0068] Optionally, the safety score of the fire-fighting body model can be calculated by using a safety score calculator based on the structured assessment report and evidence index to obtain temperature deviation, gas deviation, body collision, evidence binding, recognition success rate, etc.

[0069] It should be noted that the release control of fire-fighting mannequins has the following settings: The fire safety model passed the safety score and did not trigger the red line conditions: the system was launched according to the gray-scale strategy, and red line events, score drift, key risk omissions and near misses were continuously monitored during operation.

[0070] Triggering rollback conditions: Immediately switch back to the previous stable version and solidify the evidence index for review, thereby implementing security controls from management rules into an executable technical chain.

[0071] If the fire safety model fails the safety score or triggers red-line conditions: A structured data supplementation requirement record is automatically generated, and a training flywheel closed loop is performed on the fire safety model based on this record. In some embodiments of this invention, the structured data supplementation requirement record is associated with at least the failure rounds, failure risk factors, failure modes, target coverage intervals, and required sample size and priority. This record serves as input for the next round of data collection planning and randomized distribution updates, and is also bound to the evidence index reference. This ensures that the iteration direction is driven by threshold failure items and red-line events rather than blindly expanding data collection, thereby forming an auditable and reproducible training flywheel closed loop.

[0072] It should be noted that, to avoid situations where a high score is achieved but a serious accident occurs, a red-line veto rule (i.e., red-line conditions) is set as a strong constraint to set the threshold. In some embodiments of the present invention, red-line conditions include cumulative heat exposure exceeding the upper limit or the duration of exceeding the limit being too long, cumulative exposure to toxic gas or combustible gas exceeding the upper limit or the peak value exceeding the upper limit, occurrence of high-level fall / instability hazards or high-level collision hazards, occurrence of high-risk missed reports reaching a preset number, and lack of necessary evidence for key high-risk alarms, making them unverifiable.

[0073] The overall process of the rapid adaptation method for the fire-fighting suit model with safety scoring according to an embodiment of the present invention will be described below with reference to the accompanying drawings: The specific process of synthetic data generation and fire zone randomization (e.g.) Figure 3 (as shown) Input and configuration: asset library (structural assets, object assets), scene import and annotation (accessible areas, restricted areas and hazardous areas, checkpoints, evacuation points), hardware configuration (internal and external parameters, sampling frequency, noise, delay model).

[0074] The main process of synthesis generation includes: initialization and sampling (sampling scene layout, risk parameters, communication events, setting seed), fire risk field modeling (smoke, visibility field, thermal field / heat source, gas concentration field), task script (task template, path planning, script actions), time step simulation (progressing physical simulation, outputting RGB, thermal images, point clouds, gas, IMU, status), event recording and annotation (recording events such as collision instability and boundary crossing, generating risk labels, generating evidence indexes), and data writing and archiving (simulation data, configuration files).

[0075] Randomized configuration of fire protection domain (initialization and sampling parameters, driving risks and sensor responses in fire risk field modeling, influence dynamics conditions of time step simulation, and randomization consistency of event recording and annotation): visual (lighting / material / noise / blur), smoke obscuration (density / distribution / dynamic obscuration / visibility level), thermal imaging (drift / pseudo-hot spots / range cutoff / noise), gas (source strength / wind field / diffusion barrier / hysteresis drift), and accessibility (obstacles / narrowness / step slope / wet friction).

[0076] The specific process of automatic assessment and safety scoring (e.g.) Figure 4 (as shown) The evaluation model (i.e., the fire-fighting embodied model) undergoes both open-loop evaluation (detection F1, false negative rate, instruction understanding, boundary recognition) and closed-loop evaluation (success rate, collision / slippage and fall, timeout, communication degradation performance) based on the evaluation results of the open-loop and closed-loop evaluations. Then, the safety score calculator obtains temperature deviation, gas deviation, body collision, evidence binding, and recognition success rate based on the evaluation results of the open-loop and closed-loop evaluations and calculates the safety score. Finally, if the red line rule check is passed (i.e., the aforementioned fire-fighting embodied model passes the safety score and does not trigger the red line condition), the model is launched online; otherwise, if the red line rule check is not passed (i.e., the aforementioned fire-fighting embodied model fails the safety score or triggers the red line condition), supplementary data requirements are generated.

[0077] Rapid adaptation process of firefighter mannequins (e.g.) Figure 5 (As shown), the specific steps are as follows: Step 1: Multi-scenario data collection and archiving, then proceed to Step 2; Step 2: Data normalization and quality verification, then proceed to Step 3; Step 3: Synthetic data generation and fire zone randomization, then proceed to Step 4; Step 4: Data alignment, then proceed to Step 5; Step 5: Construct training / validation / regression sets, then proceed to Step 6; Step 6: PEFT fine-tuning to generate adapters, then proceed to Step 7; Step 7: Automatic evaluation and safety scoring, where if successful, proceed to Step 8, otherwise proceed to Step 9; Step 8: Model release; Step 9: Generate supplementary data requirement records, then proceed to Step 10; Step 10: Update sampling weights and randomization distribution, then re-execute Step 7.

[0078] In summary, the rapid adaptation method for fire-fighting vehicle models with safety scores according to embodiments of the present invention uses a unified fire-fighting vehicle data standard as a foundation, maps real-collected data and simulated synthetic data to the same field system, supplements risk forms that are difficult to cover in real fire / hazard scenarios with fire domain randomization and sampling distribution adaptive strategies, uses parameter efficient fine-tuning (PEFT) to converge new scenario adaptation to a small number of trainable parameters and form manageable scenario adaptations, and quantifies fire safety requirements into online criteria with automatic evaluation and safety scoring thresholds. Therefore, by establishing a unified standard for reusable, alignable, and traceable cross-scenario multimodal data and safety elements, and constructing a simulation synthesis data generation mechanism with randomization of the fire protection domain, the risk patterns, observation distributions, and action distributions are made closer to real scenarios. This enables rapid adaptation without altering the base model as much as possible, transforming cross-scenario migration from a long-cycle manual process into a streamlined, lightweight iteration. At the same time, automatic assessment and safety scoring thresholds are introduced to achieve an automated, auditable, and secure release process. This supports the continuous improvement of the capabilities of the fire protection embodied model in long-term operation across multiple scenarios while reducing iteration costs.

[0079] Based on the aforementioned rapid adaptation method for fire-fighting suit models with safety scores in the embodiments of the present invention, the present invention also proposes a computer-readable storage medium storing a rapid adaptation program for fire-fighting suit models with safety scores. When the rapid adaptation program for fire-fighting suit models with safety scores is executed by a processor, it implements the aforementioned rapid adaptation method for fire-fighting suit models with safety scores in the embodiments of the present invention.

[0080] It should be understood that the specific implementation of the computer-readable storage medium in the embodiments of the present invention can be found in the specific implementation of the rapid adaptation method of the fire-fighting body model with safety score in the foregoing embodiments of the present invention, and will not be repeated here to reduce redundancy.

[0081] In summary, the computer-readable storage medium according to embodiments of the present invention, by executing a rapid adaptation program of the fire-fighting body model with safety scores stored thereon, can achieve rapid adaptation of the fire-fighting body model in multiple scenarios, support the continuous improvement capability of the fire-fighting body model in long-term operation in multiple scenarios, and reduce iteration costs.

[0082] Figure 6 This is a block diagram of a rapid adaptation system for a fire-fighting suit model with safety scoring according to an embodiment of the present invention.

[0083] Specifically, in some embodiments of the present invention, such as Figure 6 As shown, the rapid adaptation system 100 for fire-fighting body models with safety scoring includes: a specification construction module 10, a data generation module 20, a data processing module 30, a rapid adaptation module 40, and a safety scoring module 50.

[0084] The specification construction module 10 is used to construct the fire protection data specification. The fire protection data specification is organized in rounds. Each round consists of time-synchronized multimodal observation sequences, action / skill sequences, risk labels, safety-related labels, evidence indexes, and scene metadata. The evidence index is used to bind conclusions with verifiable evidence. Each evidence entry is bound to a content hash, and a version binding mechanism is established to bind training / evaluation / release to dataset versions, regression set versions, synthetic randomization configuration versions, model base versions, scene adapter versions, and evaluation protocol versions. The data generation module 20 is used to generate simulated synthetic data that meets the fire protection data specification. It embeds fire domain randomization configurations throughout the initialization, sampling, label generation, and evidence generation processes of the simulated synthetic data. The fire domain randomization configurations include smoke / visibility randomization, thermal background drift and thermal reflection pseudo-hotspot randomization, gas diffusion and sensor hysteresis drift randomization, structural accessibility change randomization, ground slip / adhesion change randomization, and communication degradation and latency randomization. The data processing module 30 is used to combine the simulated synthetic data with real-world data. The system synthesizes and aligns data, and based on the risk pattern distribution obtained from online or real-world data collection, it weights and updates the randomized sampling distribution for the next round of synthesis. This ensures that the synthesized data prioritizes coverage of real high-risk boundaries and historical failure patterns. It also constructs training / validation / regression sets for the fire-fighting avatar model. The regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples. The rapid adaptation module 40 is used to generate multiple scenario adapters for the fire-fighting avatar model based on the PEFT adaptation path. A scenario router selects the corresponding scenario adapter based on scenario metadata and risk pattern characteristics, enabling rapid switching between multiple scenarios, gray-scale deployment, and one-click rollback. The safety scoring module 50 performs a safety score on the fire-fighting avatar model. If the fire-fighting avatar model fails the safety score or triggers red-line conditions, it automatically generates a structured data supplementation requirement record. Based on this record, it performs a training flywheel closed loop on the fire-fighting avatar model. Finally, if the fire-fighting avatar model passes the safety score and does not trigger red-line conditions, it is released online.

[0085] Furthermore, in some embodiments of the present invention, the specification construction module 10 is also used to generate a content hash for each evidence entry and cite it in the evaluation report.

[0086] Furthermore, in some embodiments of the present invention, the data processing module 30 is also used to explicitly mark missing modalities or incomplete rounds in the fire protection data specification and incorporate deduction or blocking rules into subsequent evidence integrity scoring and threshold determination.

[0087] Furthermore, in some embodiments of the present invention, the safety score is composed of a heat exposure sub-score, a toxic or combustible gas exposure sub-score, a near miss event sub-score, a risk boundary violation sub-score, a false alarm / missed alarm cost sub-score, an evidence integrity sub-score, and a mission success rate sub-score. The red line conditions include cumulative heat exposure exceeding the upper limit or the duration of exceeding the limit being too long, cumulative toxic or combustible gas exposure exceeding the upper limit or the peak exceeding the upper limit, the occurrence of a high-level fall / instability hazard or a high-level collision hazard, the occurrence of a high-risk missed alarm reaching a preset number of times, and the lack of necessary evidence for key high-risk alarms, resulting in unverifiable results.

[0088] Furthermore, in some embodiments of the present invention, the structured data supplementation requirement record is associated with at least the failed rounds, failure risk factors, failure modes, target coverage intervals, and required sample size and priority, and serves as input for the next round of data collection plan and randomized distribution update.

[0089] Furthermore, in some embodiments of the present invention, each adapter entry is bound to an adapter version and its corresponding dataset version, regression set version, randomization configuration version, evaluation protocol version, training recipe summary and evaluation result summary, and associated with a security score report and evidence index hash.

[0090] Furthermore, in some embodiments of the present invention, the quick adaptation module 40 is also used to select the corresponding scene adapter through the scene router based on the smoke level, light level, thermal background offset, gas change rate, slipperiness level and communication quality level.

[0091] Furthermore, in some embodiments of the present invention, the rapid adaptation module 40 is also used to, during the gray-scale deployment phase, route a small number of rounds to the new adapter through the scenario router at a preset ratio and continuously monitor the security score and red line events, and when a red line event is detected, the score drops significantly or the key risk is underreported, quickly switch to the previous stable adapter and retain the evidence index for review and supplementary data generation.

[0092] It should be understood that the specific implementation of the rapid adaptation system 100 for fire-fighting body models with safety scores in the embodiments of the present invention corresponds one-to-one with the specific implementation of the rapid adaptation method for fire-fighting body models with safety scores in the aforementioned embodiments of the present invention. To reduce redundancy, it will not be described again here.

[0093] In summary, the rapid adaptation system for fire-fighting vehicle models with safety scores according to embodiments of the present invention uses a unified fire-fighting vehicle data standard as its foundation, maps real-world collected data and simulated synthetic data to the same field system, supplements risk forms that are difficult to cover in real fire / hazard scenarios with fire domain randomization and sampling distribution adaptive strategies, uses parameter efficient fine-tuning (PEFT) to converge new scenario adaptations to a small number of trainable parameters and form manageable scenario adaptations, and quantifies fire safety requirements into online criteria with automatic evaluation and safety scoring thresholds. Therefore, by establishing a unified standard for reusable, alignable, and traceable cross-scenario multimodal data and safety elements, and constructing a simulation synthesis data generation mechanism with randomization of the fire protection domain, the risk patterns, observation distributions, and action distributions are made closer to real scenarios. This enables rapid adaptation without altering the base model as much as possible, transforming cross-scenario migration from a long-cycle manual process into a streamlined, lightweight iteration. At the same time, automatic assessment and safety scoring thresholds are introduced to achieve an automated, auditable, and secure release process. This supports the continuous improvement of the capabilities of the fire protection embodied model in long-term operation across multiple scenarios while reducing iteration costs.

[0094] It should be noted that the logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0095] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0096] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0097] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0098] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0099] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0100] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0101] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A rapid adaptation method for a fire-fighting suit model with safety scoring, characterized in that, The method includes: A fire safety data specification is constructed, which is organized in rounds. Each round consists of time-synchronized multimodal observation sequences, action / skill sequences, risk labels, safety-related labels, evidence indexes, and scene metadata. The evidence index is used to bind conclusions with verifiable evidence. Each evidence entry is bound to a content hash, and a versioning binding mechanism is established to bind training / evaluation / release to dataset versions, regression set versions, synthetic randomization configuration versions, model base versions, scene adapter versions, and evaluation protocol versions. Simulation synthetic data that meets the fire protection data specifications is generated, and fire protection domain randomization configuration is embedded in the entire process of initialization, sampling, label generation and evidence generation of the simulation synthetic data. The fire protection domain randomization configuration includes smoke / visibility randomization, thermal background drift and thermal reflection pseudo-hotspot randomization, gas diffusion and sensor hysteresis drift randomization, structural accessibility change randomization, ground slip / adhesion change randomization, and communication degradation and latency randomization. The simulated synthetic data is synthesized and aligned with the real collected data. Based on the risk pattern distribution obtained from online collected data or real machine collected data, the randomized sampling distribution of the next round of synthesis is weighted and updated so that the synthetic data can preferentially cover the real high-risk boundaries and historical failure modes. Training / validation / regression sets for fire protection body models are constructed. The regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples. Based on the PEFT adaptation path, multiple scene adapters are formed for the fire protection embodiment model. The scene router selects the corresponding scene adapter according to the scene metadata and risk morphology characteristics to achieve rapid switching between multiple scenes, gray-scale deployment and one-click rollback. The fire-fighting suit model is given a safety score. If the fire-fighting suit model fails the safety score or triggers the red line condition, a structured data supplementation requirement record is automatically generated. Based on the structured data supplementation requirement record, a training flywheel closed loop is performed on the fire-fighting suit model. After the fire-fighting suit model passes the safety score and does not trigger the red line condition, the fire-fighting suit model is released online.

2. The rapid adaptation method for fire-fighting suit models with safety scoring as described in claim 1, characterized in that, Each evidence entry is bound to a content hash, including: Generate a content hash for each of the evidence entries and cite it in the evaluation report.

3. The rapid adaptation method for fire-fighting suit models with safety scoring according to claim 1, characterized in that, The process of synthesizing and aligning the simulated synthetic data with the real-world collected data further includes: For missing modalities or incomplete rounds, the missing information is explicitly marked in the fire protection data specification, and deduction or blocking rules are incorporated into the subsequent evidence integrity scoring and threshold determination.

4. The rapid adaptation method for a fire-fighting suit model with safety scoring according to claim 1, characterized in that, The safety score consists of sub-scores for thermal exposure, toxic or combustible gas exposure, near miss event, risk boundary violation, false alarm / missed alarm cost, evidence completeness, and mission success rate. The red line conditions include cumulative thermal exposure exceeding the upper limit or excessive duration of exceeding the limit, cumulative toxic or combustible gas exposure exceeding the upper limit or peak exceeding the upper limit, occurrence of high-level fall / instability hazards or high-level collision hazards, occurrence of high-risk missed alarms reaching a preset number, and key high-risk alarms lacking necessary evidence, making them unverifiable.

5. The rapid adaptation method for a fire-fighting suit model with safety scoring according to claim 1, characterized in that, The structured data supplementation requirement record is associated with at least the failed rounds, failure risk factors, failure modes, target coverage intervals, and required sample size and priority, and serves as input for the next round of data collection plan and randomized distribution update.

6. The rapid adaptation method for a fire-fighting suit model with safety scoring according to claim 1, characterized in that, Each adapter entry is bound to the adapter version and its corresponding dataset version, regression set version, randomization configuration version, evaluation protocol version, training recipe summary, and evaluation result summary, and is associated with the security score report and evidence index hash.

7. The rapid adaptation method for a fire-fighting suit model with safety scoring according to claim 1, characterized in that, The scenario router selects the appropriate scenario-specific accessories based on smoke level, light level, thermal background shift, gas change rate, slipperiness level, and communication quality level.

8. The rapid adaptation method for a fire-fighting suit model with safety scoring according to claim 1, characterized in that, During the gray-scale deployment phase, the scenario router routes a small number of rounds to the new adapter according to a preset ratio and continuously monitors the security score and red-line events. When a red-line event is triggered, the score drops significantly, or the number of key risk misses increases, it quickly switches to the previous stable adapter and retains the evidence index for review and data supplementation needs.

9. A computer-readable storage medium, characterized in that, It stores a rapid adaptation program for a fire-fighting suit model with a safety score, which, when executed by a processor, implements the rapid adaptation method for a fire-fighting suit model with a safety score as described in any one of claims 1-8.

10. A rapid adaptation system for fire-fighting suit models with safety scoring, characterized in that, The system includes: The specification construction module is used to construct the fire protection data specification. The fire protection data specification is organized in rounds. Each round consists of time-synchronized multimodal observation sequences, action / skill sequences, risk labels, safety-related labels, evidence indexes, and scene metadata. The evidence index is used to bind conclusions with verifiable evidence. Each evidence entry is bound to a content hash, and a version binding mechanism is established to bind training / evaluation / release to dataset versions, regression set versions, synthetic randomization configuration versions, model base versions, scene adapter versions, and evaluation protocol versions. The data generation module is used to generate simulated synthetic data that meets the fire protection data specifications. It embeds fire protection domain randomization configuration throughout the initialization, sampling, label generation, and evidence generation processes of generating the simulated synthetic data. The fire protection domain randomization configuration includes smoke / visibility randomization, thermal background drift and thermal reflection pseudo-hotspot randomization, gas diffusion and sensor hysteresis drift randomization, structural accessibility change randomization, ground slip / adhesion change randomization, and communication degradation and latency randomization. The data processing module is used to synthesize and align the simulated synthetic data with the real collected data, and to perform weighted updates on the randomized sampling distribution of the next round of synthesis based on the risk pattern distribution obtained from online collected data or real machine collected data, so that the synthetic data can preferentially cover the real high-risk boundaries and historical failure modes, and to construct a training / validation / regression set for the fire protection body model. The regression set includes representative inspection route rounds for each scenario, typical rounds under various high-risk boundary conditions, and historical online failure samples and near-failure event samples. The rapid adaptation module is used to form multiple scene adapters for the fire-fighting avatar model based on the PEFT adaptation path, and selects the corresponding scene adapters according to scene metadata and risk morphology characteristics through the scene router, so as to realize rapid switching of multiple scenes, gray-scale deployment and one-click rollback. The safety scoring module is used to score the safety of the fire-fighting suit model. If the fire-fighting suit model fails the safety score or triggers the red line condition, it automatically generates a structured data supplementation requirement record, and performs a training flywheel closed loop on the fire-fighting suit model based on the structured data supplementation requirement record. After the fire-fighting suit model passes the safety score and does not trigger the red line condition, it publishes the fire-fighting suit model online.

Images