CN-121988000-A - Four-foot robot fire-fighting inspection, studying, judging and disposing closed-loop method, system and medium

CN121988000ACN 121988000 ACN121988000 ACN 121988000ACN-121988000-A

Abstract

The invention discloses a closed-loop method, a system and a medium for the fire inspection, study, judgment and treatment of a quadruped robot, which relate to the technical field of fire safety; the method comprises the steps of generating a structured search request, obtaining a search result related to a current inspection scene, constructing a scene constraint condition set, taking an event map, the scene constraint condition set and a resource state of a quadruped robot as input, generating a structured research judgment conclusion and a domain action language by using a large language model, carrying out consistency check on the domain action language, driving the quadruped robot to execute an atomic skill instruction sequence when the consistency check is passed, obtaining multi-mode feedback data, and carrying out state write-back updating on the event map according to a generated execution evidence chain. By adopting the method, the structured closed-loop execution of fire inspection, study, judgment and treatment can be realized, and the reliability and traceability of the system under a complex scene are improved.

Inventors

YUAN HONGYONG
Huo Yinuo
CHEN TAO
WANG JINGWU
HUANG LIDA
GUAN JINFU
WANG XIAOFEI
QIANG YUKAI
LI HAIPEI

Assignees

清华大学合肥公共安全研究院
清华大学

Dates

Publication Date: 20260508
Application Date: 20260407

Claims (10)

1. The method for performing fire inspection, studying, judging and handling closed loop on the four-foot robot is characterized by comprising the following steps of: acquiring multi-mode data acquired by a quadruped robot in the inspection process of an area to be inspected, identifying abnormal symptoms in the multi-mode data, extracting and generating a plurality of event nodes based on the abnormal symptoms, and constructing an event map dynamically updated along with the inspection process; Generating a structured search request based on the event map, acquiring a search result related to the current inspection scene from a knowledge base according to the structured search request, and constructing a scene constraint condition set based on the search result; taking the event map, the scene constraint condition set and the resource state of the quadruped robot as inputs, and generating a structured research and judgment conclusion and a corresponding domain action language by using a large language model; Performing consistency verification on the generated domain action language, and driving the quadruped robot to execute an atomic skill instruction sequence corresponding to the domain action language when the consistency verification is passed, and acquiring multi-mode feedback data in the process of executing the atomic skill instruction sequence; And generating an execution evidence chain according to the multi-mode feedback data and the event map, and carrying out state write-back updating on the event map based on the execution evidence chain.
2. The four-legged robot fire inspection, and judgment treatment closed-loop method according to claim 1, wherein before generating the plurality of event nodes based on the abnormality symptom extraction, the method further comprises: And carrying out time synchronization processing and space coordinate alignment processing on the multi-mode data, and uniformly mapping the data acquired by different sensors to the same time axis and a uniform space coordinate system.
3. The four-foot robot fire inspection, study and judgment treatment closed loop method of claim 2, wherein generating a plurality of event nodes based on abnormal symptom extraction comprises: carrying out space-time overlapping degree analysis on the extracted event nodes, and carrying out de-overlapping merging treatment on the repeatedly-occurring event nodes; and extracting attribute data of each event node after de-duplication combination processing, establishing a time sequence relation and a causal relation between different event nodes based on the attribute data, and constructing the event map by taking the event nodes as vertexes and the relation as edges.
4. The four-foot robot fire inspection, study, and decision handling closed loop method of claim 1, wherein the constructing a set of scene constraints comprises: Identifying a target entity and corresponding attribute information related to the event map; Generating a limit field of the structured search request based on the target entity and the attribute information, and matching corresponding security procedures and operation limits in the knowledge base according to the limit field; and converting the security procedure and the operation limit obtained by matching into a machine-readable constraint object.
5. The four-foot robot fire inspection, study and judgment treatment closed loop method according to claim 1, wherein the consistency check at least comprises one or more of the following check modes: Carrying out compliance verification on whether the domain action language violates the hard constraint in the scene constraint condition set; Based on the patrol map or BIM space information in the scene constraint condition set, carrying out reachability verification on whether the target pose corresponding to the domain action language is located in a passable path range and whether the operating space of the mechanical arm meets the safety distance; performing resource verification on whether the resource state of the four-legged robot meets the execution condition; Checking whether the action parameters in the domain action language fall into a preset parameter range or not; And judging whether the domain action language comprises a preset safety step or not according to the risk grade corresponding to the structural research judgment conclusion so as to carry out risk verification.
6. The four-legged robot fire inspection, and judgment treatment closed-loop method according to claim 1 or 5, comprising, after consistency verification of the generated domain action language: And when the consistency check fails, generating structural error reason information, and feeding back the error reason information to the large language model to regenerate or correct the domain action language.
7. The four-legged robot fire inspection, and study treatment closed loop method according to claim 1, wherein during the driving of the robot to execute the atomic skill instruction sequence, the method further comprises: monitoring safety parameters in the multi-mode feedback data in real time; When the security parameters exceed a preset security threshold, the atomic skills currently executed are forcefully interrupted, and a predefined security action is executed.
8. The four-legged robot fire inspection, and study treatment closed-loop method according to claim 1, wherein in driving a robot to execute the atomic skill instruction sequence, the method comprises: Hash binding is carried out on the multi-mode feedback data and corresponding event nodes in the event map so as to form an execution evidence chain which is related to the research conclusion and the on-site real-time perception; recording the treatment whole process data of the execution evidence chain, and generating a playable inspection report according to the state migration track of the execution evidence chain and the event map after the inspection is finished.
9. The utility model provides a four-legged robot fire control inspection is ground and is handled closed loop system, its characterized in that includes high in the clouds and robot end, wherein, the robot end includes: the four-foot robot is provided with a mechanical arm and a fire-fighting treatment tool and is used for executing fire-fighting treatment or operation tasks; the multi-mode sensing unit is deployed on the quadruped robot and is used for collecting multi-mode data and reporting the resource state of the quadruped robot in the process of inspecting the area to be inspected of the quadruped robot; the robot execution control unit is used for receiving the execution instruction from the cloud and driving the four-legged robot to execute an atomic skill instruction sequence corresponding to the domain action language; the multi-modal sensing unit is further used for collecting multi-modal feedback data in the process of executing the atomic skill instruction sequence; The cloud comprises: The sensing and calculating module is used for acquiring multi-mode data of the area to be inspected; The event map construction module is used for identifying abnormal symptoms in the multi-mode data, generating a plurality of event nodes based on the extraction of the abnormal symptoms, and constructing an event map which is dynamically updated in real time along with the inspection process; the knowledge retrieval and scene constraint module is used for generating a structured retrieval request based on the event map, acquiring a retrieval result related to the current inspection scene from a knowledge base according to the structured retrieval request, and constructing a scene constraint condition set based on the retrieval result; The research judgment and task compiling module is used for taking the event map, the scene constraint condition set and the resource state of the quadruped robot as inputs, and generating a structured research judgment conclusion and a corresponding domain action language by using a large language model; the consistency verification module is used for carrying out consistency verification on the domain action language; the execution management module is used for generating an execution instruction and issuing the execution instruction to the robot end when the consistency check is passed; And the evidence binding module is used for generating an execution evidence chain according to the acquired multi-mode feedback data and the event map, and carrying out state write-back updating on the event map based on the execution evidence chain.
10. A computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the four-legged robot fire inspection, study, and treatment closed loop method of any of claims 1 to 8.

Description

Four-foot robot fire-fighting inspection, studying, judging and disposing closed-loop method, system and medium Technical Field The invention relates to the technical field of fire safety, in particular to a closed-loop method, a system and a medium for fire inspection, study and judgment treatment of a quadruped robot. Background With the continuous improvement of fire safety management requirements of industrial devices, warehouse logistics, electric power facilities, rail transit and other places, the inspection work is gradually changed from a mode of purely relying on manual inspection to a mode of cooperation of manual and automatic equipment. In complex environments such as high temperature, high humidity, toxic and harmful gas, smoke shielding or space limitation, the manual inspection has high personal risks, and long-term stability is difficult to maintain in aspects of inspection frequency, coverage range and execution consistency. Automatic inspection is carried out based on a mobile robot, and a risk source is handled in an auxiliary way when necessary, so that the automatic inspection system has become an important development direction in the fields of fire protection and safety engineering. In the related art, a patrol robot platform generally has mature motion control and task arrangement capability, can execute patrol actions according to a preset path, and acquires multi-source data such as images, thermal imaging, gas concentration or laser point cloud. The platform forms a perfect technical system in the aspects of task scheduling, path execution and data feedback, but the task expression mode is mainly based on scripts or task trees, so that the problems of reaching a designated position and executing designated actions are mainly solved, and the structural expression capability of fire risk situations in complex scenes is relatively limited. The integration capability of multi-source evidence required for fire risk research and judgment is relatively limited, even if data such as thermal imaging, images, point cloud, gas concentration and the like can be acquired, related data are usually presented in a discrete record or alarm form, and a coherent situation expression is difficult to form, so that subsequent research and judgment and treatment still highly depend on manual experience, and stable and reproducible automatic processing flows are difficult to realize. In the related system and scheme of the fire-fighting four-foot robot, the related system and scheme are generally designed around modules such as hardware integration, map construction positioning, target recognition, fire extinguishing device control and the like, and can finish detection or fire extinguishing tasks in specific scenes. However, such schemes are more embodied as integrated implementations of functional modules, and the disclosure of how the research and decision process develops conclusions based on multi-source evidence, and how the conclusions affect subsequent action arrangements is relatively weak, so that it is difficult to meet engineering requirements for the interpretation, auditability, and future complex disc analysis of the research and decision process in complex scenarios. In practical engineering application, when facing complex scenes of multi-factor superposition and dynamic change of risk level, the system often needs to rely on manual participation to carry out comprehensive judgment, and the degree of automation is limited to a certain extent. Meanwhile, in recent years, intelligent research attempts have been made to introduce large language models to participate in task planning and skill selection, and candidate actions are screened or ranked through language understanding and reasoning capabilities. The method has the potential of converting from language to action in a general scene, and external knowledge is introduced by means of retrieval enhancement and the like to improve the reasoning capability. However, related research is multi-faceted to general task scene, systematic design is less conducted by combining professional knowledge system, scene management specification and multi-modal evidence features in the fire-fighting field, and a verifiable and playable disposal flow is difficult to directly form by a language reasoning result when facing a high-risk environment, so that challenges in reliability and compliance still exist in practical floor application. In addition, the differences among different industrial sites in terms of space structure, hazard source type and management specification are obvious, related systems are deployed in a scene customization mode, and a structuring situation expression and task compiling framework capable of being multiplexed across scenes is lacking. When the system is migrated to a new application scene, a large amount of reconfiguration and adaptation work is needed, so that the implementation cost is high, the iteration period is long