A Robotic Detection Method and System for Building Fire Protection Facilities Based on Multimodal Global Perception and Large Model Intelligence

Abstract

The present invention relates to the technical field of intelligent fire-fighting facility detection, and specifically discloses a method and system for detecting building fire-fighting facilities based on multi-modal global perception and large model intelligence. The present invention obtains building information and automatically matches the detection standard system and detection process, collects multi-modal data through a multi-modal perception module, carries the multi-modal perception module and the fire-fighting facility trigger module through an autonomous mobile platform, and performs existence detection, compliance detection, responsiveness detection and linkage detection through a pre-trained large model to generate a comprehensive judgment result and a detection report. The present invention realizes the standardized generation of detection schemes, the all-dimensional data collection, the hierarchical reasoning and judgment based on the logic of fire-fighting detection engineering, and the automatic output of detection reports, constructs an intelligent detection closed loop, and improves the standardization degree, detection coverage, result reliability and operation efficiency of building fire-fighting facility detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent fire protection facility detection technology, and in particular to a robot detection method and system for building fire protection facilities based on multimodal global perception and large-scale model intelligence. Background Technology

[0002] Building fire protection facility testing refers to the on-site verification of the completeness of configuration, operation function and linkage response capability of various fire protection facilities installed in a building, such as fire water supply and fire hydrant systems, automatic sprinkler systems, automatic fire alarm systems, and smoke control and exhaust systems, in accordance with national fire protection technical standards. This ensures that the building fire protection facilities can be activated normally and respond in coordination as required by design when a fire occurs, and perform the expected fire extinguishing and evacuation functions.

[0003] Currently, the testing industry predominantly employs manual testing methods. Testing personnel execute each item according to legal standards, resulting in low data collection efficiency, difficulty in synchronously recording multi-source data, reliance on manual operation for cross-system linkage verification, and subjective differences in test results due to human experience. This leads to insufficient consistency in the testing process. Existing automated testing equipment only provides partial functional replacement; some equipment can perform single-item data collection or single-device start-stop testing for specific fire protection facilities. However, it lacks deep integration with fire protection regulations and standards, as well as the linkage logic of fire protection systems. Therefore, it cannot comprehensively replace manual testing in terms of testing depth and breadth, and cannot achieve automated verification of the linkage response between multiple fire protection subsystems.

[0004] In summary, existing technologies suffer from problems such as limited data dimensions, low efficiency in cross-system collaborative verification, poor consistency of test results, and insufficient traceability of the testing process. There is an urgent need to propose a fire protection facility testing solution that can achieve full-dimensional data collection, automated cross-system collaborative verification, and complete recording of the testing process. Summary of the Invention

[0005] To address the problems existing in the prior art, the main objective of this invention is to provide a robot detection method and system for building fire protection facilities based on multimodal full-domain perception and large-scale model intelligence. This system can realize full-dimensional data collection, cross-system linkage logic verification, and intelligent detection report generation for building fire protection facilities during dynamic response processes, thereby improving the integrity, reliability, and efficiency of fire protection facility detection.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for detecting building fire protection facilities using robots based on multimodal global perception and large-scale model intelligence, comprising the following steps: The system obtains the usage, fire hazard category, and building type of the building under inspection, and automatically matches the corresponding testing standard system and testing process based on the identification results. The multimodal sensing module collects multimodal data including building environment status, operating parameters of fire protection facilities, and linkage signals of fire protection facilities. The fire protection facility trigger module sends a simulated fire signal to the fire protection system and activates the fire protection system after fire confirmation. The multimodal sensing module and fire-fighting facility triggering module are carried out through an autonomous mobile platform; Using a pre-trained large model, based on the detection process, the collected multimodal data is processed and analyzed, and existence detection, compliance detection, responsiveness detection, and linkage detection are performed to generate a comprehensive judgment result. Based on the testing standard system and the judgment results, a test report is generated.

[0007] Preferably, the existence detection verifies whether the fire protection subsystem or its components are set up as required by the design, generating a result indicating presence or absence; the compliance detection determines whether the installation location, specifications, installation method, and quantity configuration of the fire protection subsystem or its components meet the requirements of fire protection technical standards, generating a result indicating compliance or non-compliance; the responsiveness detection verifies whether the fire protection equipment can make the expected response action when it receives a trigger signal, generating a result indicating compliance or non-compliance; and the linkage detection verifies whether all devices that should be linked are activated after the trigger signal is issued, generating a result indicating compliance or non-compliance. Through the above scheme, the fire protection testing task is decomposed into four levels: existence detection, compliance detection, responsiveness detection, and linkage detection, corresponding to whether components are set up as designed, whether components meet standard requirements, whether a single device can respond to a trigger signal, and whether the system linkage is complete. This achieves complete coverage and logical layering of the fire protection testing task from component presence to system coordination.

[0008] Preferably, the existence detection, compliance detection, responsiveness detection, and linkage detection are executed in a hierarchical order, with existence detection taking precedence. When the existence detection result is missing, compliance detection, responsiveness detection, and linkage detection are no longer performed on the missing detection object, and it is marked as inapplicable. This scheme establishes a hierarchical execution order for existence detection, compliance detection, responsiveness detection, and linkage detection, with existence detection taking precedence and subsequent detections terminated when the existence detection result is missing. This avoids performing invalid detections on missing components and optimizes the execution efficiency of the detection process.

[0009] Preferably, the testing standard system is linked to a pre-built field-based standard knowledge base, which includes: a standard index field, a testing object field, a setting location field, a pass / fail judgment rule field, a testing method description field, and a set of linked devices field. The testing process is generated by the autonomous mobile platform based on the spatial information of the inspected building and the field-based standard knowledge base. The testing process includes: a sequence of testing points, a standard index corresponding to each testing point, the fire protection subsystem and components to be tested corresponding to each testing point, a trigger operation type corresponding to each testing point, and a multimodal data type to be collected corresponding to each testing point. Through the above scheme, the testing standard system is constructed as a field-based standard knowledge base containing a standard index field, a testing object field, a setting location field, a pass / fail judgment rule field, a testing method description field, and a set of linked devices field. Based on the spatial information of the inspected building and the field-based standard knowledge base, a testing process containing a sequence of testing points, a standard index, components to be tested, trigger operation types, and data types to be collected is generated, achieving precise association and automated generation of testing standards and execution instructions.

[0010] Preferably, the multimodal sensing module includes: The visual perception group is used to acquire image data and spatial coordinate data. The acoustic sensing group is used to collect sound pressure level data and ultrasonic data; The physical quantity sensing group is used to collect pressure data, flow data, liquid level data, and distance data; The electrical parameter sensing group is used to collect current data, voltage data, and frequency data. The signal acquisition interface group is used to acquire dry contact signals and analog signals from fire protection facilities via hard-wiring. Through the above scheme, the multimodal sensing module is configured into a visual sensing group, an acoustic sensing group, a physical quantity sensing group, an electrical parameter sensing group, and a signal acquisition interface group. These groups respectively acquire image data and spatial coordinate data, sound pressure level data and ultrasonic data, pressure data, flow data, liquid level data and distance data, current data, voltage data and frequency data, as well as dry contact signals and analog signals via hard-wiring, achieving comprehensive data acquisition of the appearance, operating parameters, and electrical signals of fire protection facilities.

[0011] Preferably, the fire protection facility triggering module includes: Mechanical triggering unit, used to operate valves or buttons of fire protection facilities via a robotic arm and end effector; An electrical signal triggering unit is used to send a simulated contact closure signal to the fire protection facility control cabinet via an electrical interface; The simulated heat source generating unit is used to apply a simulated heat source to the temperature sensor. The simulated smoke generating unit is used to apply simulated smoke to the smoke detector.

[0012] The above scheme configures the fire protection facility triggering module into a mechanical triggering unit, an electrical signal triggering unit, a simulated heat source generating unit, and a simulated smoke generating unit. These are used to operate valves or buttons through a robotic arm and end effector, send simulated contact closure signals through an electrical interface, apply simulated heat sources to heat detectors, and apply simulated smoke to smoke detectors, thereby realizing the active simulation capability of multiple triggering methods for fire protection facilities.

[0013] Preferably, the autonomous mobile platform plans its path based on the spatial information of the building under inspection and moves to each inspection point determined in the inspection process. At each inspection point, for the object to be inspected that needs to be triggered, a simulated trigger signal is sent to the object through the fire protection facility trigger module, and inspection data is collected by the multimodal perception module. The collected multimodal data is then transmitted to the pre-trained large model for analysis and processing. Through this scheme, path planning is performed based on the spatial information of the building under inspection, and the platform moves to each inspection point determined in the inspection process. At each inspection point, for the object to be inspected that needs to be triggered, a simulated trigger signal is sent to the object through the fire protection facility trigger module, and inspection data is collected by the multimodal perception module. The collected multimodal data is then transmitted to the pre-trained large model for analysis and processing, achieving coordinated automation of spatial positioning, triggering, and data collection in the inspection execution process.

[0014] Preferably, before being processed by the pre-trained large model, the multimodal data is first converted into structured data fields. Each structured data field includes: a standard index, the object being detected, the location of the sensor, the sensor type, the numerical value, the unit of measurement, the acquisition timestamp, and the spatial coordinate label. This scheme converts the multimodal data into structured data fields containing a standard index, the object being detected, the location of the sensor, the sensor type, the numerical value, the unit of measurement, the acquisition timestamp, and the spatial coordinate label before processing by the pre-trained large model. This achieves a standardized conversion of unstructured perceptual data into computable fields, providing a unified data input format for the inference and judgment of the large model.

[0015] Preferably, the pre-trained large model constructs a three-level context architecture during processing and analysis, and performs detection and analysis using a hybrid reasoning architecture that combines language reasoning and numerical computation. The three-level context architecture includes: A short-term context is used to carry the structured data fields collected in real time at the current detection point; The context is used to carry a structured summary text generated by the pre-trained large model, which contains a summary of the judgment results of all detection items in a subsystem or a fire compartment; Long-term context is used to retrieve the current detection record, historical detection archives, and the fieldized standard knowledge base from the cloud database as context by enhancing the generation architecture through retrieval; In the hybrid inference architecture, the language inference module identifies the detection task type and calls the computation execution module to perform numerical comparison, time difference calculation and set comparison operations.

[0016] The above scheme constructs a three-level context architecture: short-term context, medium-term context, and long-term context. These are used to carry the structured data fields of the current detection point, the summary of the judgment results of the subsystem or fire compartment, and the current detection record and historical detection archive, respectively. A hybrid reasoning architecture that combines language reasoning and numerical calculation is used to perform detection analysis. The language reasoning module identifies the detection task type and calls the calculation execution module to perform numerical comparison, time difference calculation, and set comparison operations, thereby achieving efficient organization and accurate calculation of detection data.

[0017] In a second aspect, the present invention provides a building fire protection facility robot detection system based on multimodal global perception and large-scale model intelligence, comprising: The building type matching module is used to obtain the usage nature, fire hazard category and building type of the building under inspection, and automatically match the corresponding testing standard system and testing process based on the identification results; The multimodal sensing module is used to collect multimodal data, including building environment status, operating parameters of fire protection facilities, and linkage signals of fire protection facilities; The fire protection facility triggering module is used to send a simulated fire signal to the fire protection system and activate the fire protection system after fire confirmation. An autonomous mobile platform is used to carry the multimodal sensing module and the fire-fighting facility triggering module; The data processing and judgment module integrates a pre-trained large model; the pre-trained large model is configured to process and analyze the collected multimodal data based on the detection process, perform existence detection, compliance detection, responsiveness detection and linkage detection, and generate a comprehensive judgment result. The report generation module is used to generate a test report based on the test standard system and the judgment results.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention establishes a standardized testing scheme generation mechanism by identifying information about the inspected building and automatically matching it with corresponding testing procedures and standards. Through an autonomous mobile platform equipped with a multimodal perception module and a fire protection facility triggering module, it achieves autonomous navigation, fixed-point docking, simultaneous multi-source data acquisition, and active triggering of simulated fire signals within the building space, constructing a comprehensive data acquisition capability from environmental perception to proactive intervention. A pre-trained large model processes and analyzes the acquired multimodal data, performing hierarchical reasoning and judgment based on fire protection testing engineering logic, achieving fully automated testing across the entire chain, from component existence, compliance, and responsiveness to system linkage. By generating testing reports based on the integration of various judgment results according to the testing standards system, it achieves a streamlined process connecting data acquisition, analysis, judgment, and result output. The synergistic effect of these technical features constructs a fully intelligent testing closed loop encompassing testing standard matching, autonomous perception and triggering, hierarchical reasoning and judgment, and report generation, comprehensively improving the standardization, coverage, reliability, and efficiency of building fire protection facility testing. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the overall process of the robot detection method for building fire protection facilities based on multimodal global perception and large model intelligence according to an embodiment of the present invention.

[0020] Figure 2 This is a flowchart illustrating the detection standard matching and detection process generation in the building fire protection facility robot detection method based on multimodal global perception and large model intelligence according to an embodiment of the present invention.

[0021] Figure 3 This is a flowchart illustrating the multimodal data acquisition and structured conversion process in the robot detection method for building fire protection facilities based on multimodal global perception and large model intelligence, according to an embodiment of the present invention.

[0022] Figure 4 This is a flowchart illustrating the simulated fire signal triggering and linkage response in the building fire protection facility robot detection method based on multimodal global perception and large model intelligence according to an embodiment of the present invention.

[0023] Figure 5 This is a flowchart illustrating the multi-dimensional detection and analysis of a large model in the robot detection method for building fire protection facilities based on multimodal global perception and large model intelligence, according to an embodiment of the present invention.

[0024] Figure 6 This is a flowchart illustrating the generation of an inspection report in a robot inspection method for building fire protection facilities based on multimodal global perception and large-scale model intelligence, according to an embodiment of the present invention.

[0025] Figure 7This is a functional block diagram of a building fire protection facility robot detection system based on multimodal global perception and large model intelligence according to an embodiment of the present invention.

[0026] Figure 8 This is a schematic diagram of the structure of the autonomous mobile platform of the building fire protection facility robot detection system based on multimodal global perception and large model intelligence according to an embodiment of the present invention. Detailed Implementation

[0027] To better illustrate the objectives, technical solutions, and advantages of the present invention, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0028] The technical solutions provided in this application can be widely applied to various building fire protection facility testing scenarios, including system integration testing and fire protection acceptance testing of newly built buildings during the completion and acceptance phase, functional testing and related system verification of renovated buildings after the renovation is completed, single-point testing and overall system integration testing of expanded buildings in newly added areas, and annual regular testing, quarterly inspections, and special functional testing of in-use buildings in accordance with national or local fire protection technical standards.

[0029] The fire protection systems applicable to this application include fire water supply and fire hydrant systems, automatic sprinkler systems, automatic tracking and positioning jet fire extinguishing systems, deluge and water curtain and water mist fire extinguishing systems, gas fire extinguishing systems, foam fire extinguishing systems, fine water mist fire extinguishing systems, fixed fire monitor fire extinguishing systems, dry powder fire extinguishing systems, kitchen equipment fire extinguishing devices, automatic fire alarm systems, smoke prevention and smoke exhaust facilities, fire doors and windows and fireproof roller shutters, fire door monitoring systems, fire power supply and its distribution, fire emergency lighting and evacuation guidance systems, building fire extinguishers, electrical fire monitoring systems, fire equipment power supply monitoring systems, and urban fire remote monitoring systems.

[0030] like Figures 1 to 6 As shown in the figure, this application discloses a robot detection method for building fire protection facilities based on multimodal global perception and large model intelligence, including the following steps: S1. Obtain the usage nature, fire hazard category and building type of the building under inspection, and automatically match the corresponding testing standard system and testing process based on the identification results; S2. Collect multimodal data, including building environment status, operating parameters of fire protection facilities, and linkage signals of fire protection facilities, through the multimodal sensing module; S3. Send a simulated fire signal to the fire protection system through the fire protection facility trigger module and start the fire protection system after fire confirmation; S4. The multimodal sensing module and fire-fighting facility triggering module are carried on an autonomous mobile platform; S5. Using a pre-trained large model, based on the detection process, process and analyze the collected multimodal data, perform existence detection, compliance detection, responsiveness detection and linkage detection, and generate a comprehensive judgment result. S6. Based on the testing standard system and the judgment results, generate a test report.

[0031] In this embodiment, S1 specifically includes the following sub-steps.

[0032] S11. Building Information Input and Analysis. The design documents of the inspected building are read through the building type matching module, or the building's usage, fire hazard category, and building type are entered through the human-computer interaction interface. Usage categories include civil buildings, industrial buildings, and warehouse buildings. Fire hazard categories are classified according to national fire protection technical standards into light hazard, medium hazard, severe hazard, and warehouse hazard levels. Building types include high-rise civil buildings, large shopping malls, underground buildings, industrial plants, and warehouse buildings.

[0033] S12. Testing Standard System Retrieval and Matching. Based on the identification results, the corresponding testing standard system is automatically retrieved and matched from the pre-built fieldized standard knowledge base. The testing standard system contains national fire protection technical standard clauses applicable to the inspected building. The fieldized standard knowledge base converts the testing items, pass / fail thresholds, testing method descriptions, and pass / fail rules in the fire protection facility testing standards into searchable text field structures and associates them with specific standard indexes. During the matching process, the pre-trained large model uses semantic understanding capabilities to parse structured fields such as the usage nature, fire hazard category, and building type of the inspected building, maps them to the corresponding entries in the fieldized standard knowledge base, generates query conditions for retrieval, and recalls the corresponding set of standard clauses from the knowledge base.

[0034] S13. Detection Process Generation. Based on the matched detection standard system and combined with the spatial information of the building under inspection, the detection process is automatically generated. The detection process includes the sequence of detection points, the standard index corresponding to each detection point, the fire protection subsystems and components to be tested for each detection point, the trigger operation type corresponding to each detection point, and the multimodal data types to be collected for each detection point. Spatial information includes the floor plan layout of each floor of the building, fire compartment division, and coordinates of the installation locations of fire protection facilities. This information is pre-collected by the autonomous mobile platform using LiDAR synchronous positioning and mapping technology or imported from the Building Information Modeling (BIM) system. When generating the detection process, the pre-trained large model, based on the correlation between spatial information and standard clauses, extracts the standard index of each detection point from the fieldized standard knowledge base through a retrieval-enhanced generation architecture, and generates the corresponding trigger operation type and multimodal data types to be collected based on the detection method description in the standard index.

[0035] Therefore, this embodiment achieves standardized generation and precise adaptation of testing schemes, eliminating the uncertainty brought about by manually reviewing standards and formulating schemes.

[0036] In this embodiment, S2 specifically includes the following sub-steps.

[0037] S21. Autonomous Navigation and Point-to-Point Docking. Utilizing a multimodal perception module mounted on an autonomous mobile platform, the system autonomously navigates within the building and docks at pre-defined detection points. Based on the detection point sequence generated in S13, and combined with a building space map constructed using simultaneous positioning and mapping technologies, the autonomous mobile platform plans the optimal movement path and achieves centimeter-level positioning accuracy through lidar and inertial measurement units, sequentially docking at each detection point.

[0038] S22. Visual Data Acquisition. At the detection points, image data and spatial coordinate data are acquired through a visual perception group. The visual perception group includes a high-definition visible light camera and a 3D laser scanner. At the detection points, the high-definition visible light camera acquires images of the appearance of fire protection facilities to identify their presence, model identification, signal valve handle position, and control cabinet manual / automatic status; the 3D laser scanner acquires the spatial coordinates of the facility's installation location for spatial comparison with the location settings in design drawings or a field-based standard knowledge base. The acquired image data is input into the target detection module of a pre-trained large model, which outputs component categories and bounding box information, which are then associated with the standard index in the detection process and stored in structured data fields.

[0039] S23. Acoustic Data Acquisition. Sound pressure level and ultrasonic data are acquired through an acoustic sensing group. This group includes a sound pressure level meter and an ultrasonic sensor. At the detection points, the sound pressure level meter measures the sound pressure level of equipment such as hydraulic alarm bells and smoke exhaust fans when they are activated, recording the data in dB. The ultrasonic sensor detects leaks in gas extinguishing system pipelines or the tightness of smoke exhaust valve closures, determining the location and extent of leaks by analyzing ultrasonic frequency signals. The acquired acoustic data is converted into digital values ​​after signal conditioning and stored in structured data fields after being associated with a standard index in the detection process.

[0040] S24. Physical Quantity Data Acquisition. Pressure, flow, level, and distance data are acquired through a physical quantity sensing group. This group includes a pressure sensor, flow sensor, level sensor, and laser rangefinder. At the testing points, the pressure sensor connects via a quick-connect interface to the end-of-line test device, fire pump outlet main, and elevated fire water tank outlet pipe, collecting pressure values ​​in MPa. The flow sensor is connected in series or clamped to the pipeline to collect flow values ​​in L / s. The level sensor extends into the elevated fire water tank to collect level values ​​in meters (m). The laser rangefinder measures spatial dimensions such as sprinkler spacing and smoke curtain droop height in millimeters (mm). The acquired physical quantity data is stored in structured data fields after being associated with the standard index in the testing process.

[0041] S25. Electrical Parameter Data Acquisition. Current, voltage, and frequency data are acquired through an electrical parameter sensing group. This group includes a non-contact current clamp, a voltage probe, and a frequency meter. At the detection point, the non-contact current clamp is attached to the power supply cable of equipment such as fire pumps and smoke exhaust fans to collect the operating current (A). The voltage probe is connected in parallel to the control cabinet terminal block to collect the supply voltage (V). The frequency meter measures the inverter's output frequency (Hz). The acquired electrical parameter data is used to determine whether the equipment starts normally, whether the main / standby power supply switching is normal, and whether the automatic inspection is operating at the set frequency. After being associated with the standard index in the detection process, the data is stored in structured data fields.

[0042] S26. Acquisition of Digital and Analog Signals. Dry contact signals and analog signals from fire protection facilities are acquired via a hard-wired signal acquisition interface group. The signal acquisition interface group includes digital and analog acquisition modules. At the detection points, the digital acquisition module is connected in parallel with wires to the normally open contacts of devices such as pressure switches, flow switches, water flow indicators, and signal valves, acquiring the moment when the contacts close or open, with a time resolution better than 1ms. The analog acquisition module is connected to the current loop of devices such as pressure transmitters and level transmitters, acquiring continuously changing quantities. The acquired signals are used to accurately record the timestamps of each device's actions, providing a time reference for subsequent responsiveness and linkage detection.

[0043] Therefore, this embodiment realizes a closed loop of full-dimensional data acquisition from environmental perception to active triggering, solving the problems of incomplete data acquisition and difficulty in synchronizing multi-source data in traditional detection.

[0044] In this embodiment, S3 specifically includes the following sub-steps.

[0045] S31. Trigger Unit Selection. Based on the trigger operation type determined in the detection process generated in S13, select the corresponding trigger unit from the pre-configured trigger unit library. The trigger operation type is derived from the detection method description associated with the standard index in the detection process, including mechanical operation type, electrical signal type, environmental simulation type, etc.

[0046] S32. Mechanical Trigger Operation. When the trigger operation type is mechanical, the operation is executed through a mechanical trigger unit. The mechanical trigger unit includes a multi-degree-of-freedom robotic arm and an end effector. The end effector is configured as a screw gripper, a pressing head, or a lever depending on the object being operated on. At the detection point, the robotic arm, based on the spatial coordinates of the target object provided by the vision perception group, uses visual servo control to precisely position the end effector to the operating position, turning the valve handwheel with a predefined torque. While executing the operation, the mechanical trigger unit uses a force sensor to provide feedback on the contact force, ensuring the operation is in place and does not damage the equipment.

[0047] S33. Electrical Signal Simulation. When the trigger operation type is electrical signal simulation, the operation is executed through the electrical signal trigger unit. The electrical signal trigger unit includes a programmable digital output module and an analog output module, which are isolated and converted before being connected to the terminal block of the control cabinet of the device under test through test clamps or standard interfaces. The signal type, amplitude, and duration output by the electrical signal trigger unit are determined by the pass / fail judgment rules associated with the standard index in the testing process, and the output timestamp is recorded by the system clock synchronization mechanism.

[0048] S34. Fire detector simulation. When the trigger operation type is fire detector simulation, the operation is performed through a simulated heat source generating unit or a simulated smoke generating unit.

[0049] S35. Synchronous Triggering and Feedback Monitoring. Before the trigger operation is executed, the signal acquisition interface group begins to synchronously monitor the action feedback signals of relevant devices and records the precise timestamp of each trigger signal. After the trigger operation is completed, the system waits for the feedback signal according to the preset timeout period, and stores the trigger time and feedback time in a structured data field after associating them.

[0050] Therefore, this embodiment realizes the ability to actively simulate multiple triggering methods of fire protection facilities, and can verify the response status of the fire protection system without an actual fire.

[0051] S4. The multimodal sensing module and fire-fighting facility triggering module are carried on an autonomous mobile platform; In this embodiment, S4 specifically includes the following sub-steps.

[0052] S41. Platform Composition and Autonomous Navigation. The autonomous mobile platform adopts a tracked or wheeled chassis and integrates a lidar synchronous positioning and mapping module, an inertial measurement unit, and a high-precision encoder to achieve autonomous navigation and centimeter-level positioning within the building environment. Based on the detection point sequence generated in S13, the platform plans the optimal movement path using a building space map constructed by synchronous positioning and mapping technology. It then uses lidar to scan environmental features in real time and combines this with data fusion calculations from the inertial measurement unit to achieve pose estimation and path tracking.

[0053] S42. Point Dwelling and Collaborative Execution. The platform moves sequentially to each detection point. During the docking process, the visual perception group continuously identifies environmental features and corrects positioning errors in real time, ensuring that the deviation between the docking position and the preset detection point does not exceed a threshold. At each detection point, the platform provides a stable spatial coordinate reference for the multimodal perception module and a precise operational positioning reference for the fire protection facility triggering module, realizing the collaborative automation of spatial positioning, triggering operation, and data acquisition during the detection execution process.

[0054] In this embodiment, S5 specifically includes the following sub-steps.

[0055] S51. Multimodal Data Field Conversion. The multimodal data collected in S2 is converted into structured data fields. Each structured data field contains a standard index, detection object, installation location, sensor type, numerical value, unit of measurement, acquisition timestamp, and spatial coordinate label. The standard index field uniquely identifies the fire protection technical standard clause corresponding to the data; the detection object field identifies the name of the component being measured (e.g., wet alarm valve pressure switch); the installation location field records the theoretical installation location coordinates of the component in the building; the sensor type field identifies the type of sensor used to collect the data (e.g., digital acquisition module, pressure sensor, high-definition visible light camera); the numerical value field stores the raw or processed values ​​measured by the sensor (e.g., contact closure, 0.85MPa, 0.95); the unit of measurement field identifies the unit of the numerical value (e.g., MPa, dB, mm, s); the acquisition timestamp field records the time of data acquisition; and the spatial coordinate label field records the actual spatial location coordinates corresponding to the data. The transformation of structured data fields is achieved through pre-trained large model prompt words engineering. The prompt words contain field mapping rules and format constraints. The pre-trained large model parses unstructured sensor output text (e.g., the string "75dB" output by the sound pressure level meter) or image recognition results (e.g., the "signal valve" category and confidence level output by the object detection model) into preset JSON format fields. Numerical fields retain their original precision, timestamps adopt ISO 8601 format (e.g., 2025-03-28T10:30:15.123Z), and spatial coordinates are represented in Cartesian coordinate system (e.g., x, y, z coordinate values ​​in meters).

[0056] S52. Construction of a Three-Level Context Architecture. A three-level context architecture is constructed to organize and manage the detection data. This architecture is implemented through the context window of a pre-trained large model in collaboration with an external vector database. The short-term context directly utilizes the context window of the large model to carry the real-time data of the current detection point, while the medium-term and long-term contexts are stored and retrieved through an external vector database to expand the effective context capacity of the large model.

[0057] S521. Establish a short-term context. The short-term context carries the structured data fields collected in real time at the current detection point for immediate analysis and judgment. For example, it performs instantaneous comparison and pump start-up time calculation on the pressure value, flow switch action signal, and pressure switch action signal of the end-point test device at the current detection point. The short-term context directly occupies the context window of the current dialogue round of the pre-trained large model, with a capacity limit of 128k tokens. When the data volume of a single detection point exceeds the limit, a sliding window mechanism is used to retain the key data fields of the most recent moment (such as the latest pressure value, flow switch action time, and pressure switch action time) and discard redundant information (such as duplicate intermediate state records or historical dialogue rounds) to ensure the real-time nature of the immediate judgment.

[0058] S522. Establish a mid-time context. The mid-time context carries a structured summary text generated by a pre-trained large model, which contains a summary of the judgment results of all detection items within a subsystem or a fire compartment. After completing the full inspection of a subsystem (e.g., a wet automatic sprinkler system) or a fire compartment, the pre-trained large model compresses the raw data fields in the short-term context into a structured summary by generating summary prompts. The summary content includes the standard index of all inspection items within the subsystem or fire compartment, existence judgment results (e.g., "existing" or "missing"), compliance judgment results (e.g., "qualified" or "unqualified"), responsiveness judgment results (e.g., "response qualified" or "response unqualified"), linkage judgment results (e.g., "linkage qualified" or "linkage unqualified"), inspection time range (e.g., 2025-03-28T10:30:00Z to 2025-03-28T10:35:00Z), spatial location identifier (e.g., alarm valve room on floor B1), and related linkage event records (e.g., trigger signal issuance time, feedback time of each device action). The summary text is stored in local cache in JSON format for subsequent inspection steps (e.g., cross-system linkage analysis or global report generation) as context input.

[0059] S523. Establish a long-term context. The long-term context is generated by retrieving current detection records, historical detection archives, and a field-based standard knowledge base from the cloud database using a retrieval-enhanced generative architecture. During detection execution, the pre-trained large model constructs a query vector based on the standard index of the current detection point. It then retrieves relevant historical detection records (e.g., the judgment results of previous detections of the same building and subsystem) from the cloud vector database using vector similarity retrieval, related clauses in the field-based standard knowledge base (e.g., standard clauses of other subsystems that are linked to the current detection point), and completed detection records (e.g., the judgment results of completed subsystems). The retrieved content is concatenated with the current context and input into the pre-trained large model to assist in complex judgments (e.g., analyzing the reasons for missing linkage logic) and anomaly analysis (e.g., tracing the reasons for pressure switch action timeouts).

[0060] S53. Hybrid Inference Architecture Execution. A hybrid inference architecture combining language inference and numerical computation is employed for detection and analysis. In this architecture, the language inference module identifies and detects task types based on the natural language understanding capabilities of a pre-trained large model, while the computation execution module runs independently outside the pre-trained large model, performing precise numerical calculations, time series analysis, and set operations via an application programming interface (API).

[0061] S531. Detection Task Identification and Rule Retrieval. The language reasoning module identifies the detection task type based on the sensor type in the standard index fields and structured data fields passed in by the current detection point. Detection task types include existence determination, compliance determination, responsiveness determination, and linkage determination. The language reasoning module parses the standard index fields passed in the detection process, uses them as query conditions to retrieve data from the index table of the fieldized standard knowledge base, and returns a structured description of the compliance determination rules.

[0062] S532. Numerical Comparison and Time Difference Calculation. When a numerical comparison, time difference calculation, or set comparison task is detected, the language inference module calls the computation execution module to perform the corresponding operation. The computation execution module has built-in numerical comparison functions, time difference calculation functions, and set operation functions. After receiving the measured value parameters and threshold parameters from the language inference module, it returns a Boolean result or a numerical result. During numerical comparison, the computation execution module compares the measured value (e.g., the pressure value collected in S24) with the threshold (e.g., the upper and lower pressure limits set in the fieldized standard knowledge base), supporting operators such as greater than, less than, equal to, and interval inclusion. During time difference calculation, the computation execution module converts two timestamps (e.g., the trigger signal issuance time T1 recorded in S35 and the device action feedback time T2 recorded in S26) into millisecond-level Unix timestamps, calculates the difference, and outputs the difference for comparison with the allowed response time. During set comparison, the computation execution module determines whether the measured feedback device set (e.g., the actual feedback device identifier set collected in S26) contains all elements of the set of devices to be linked (e.g., the linked device set associated with the standard index in the fieldized standard knowledge base).

[0063] S54. Hierarchical Detection Execution. The four types of detection are executed in the hierarchical order of existence detection, compliance detection, responsiveness detection, and linkage detection. Each level of detection takes the passing result of the previous level as a prerequisite. If the result of the previous level of detection is unqualified or missing, the subsequent detection is terminated and marked as inapplicable.

[0064] S541, Existence Detection Execution. The pre-trained large model utilizes visual recognition capabilities, based on image data from the structured data fields generated in S51, to identify the existence of the target object using a target detection algorithm. The pre-trained large model inputs image data (e.g., an image of the outlet pipe of a high-level fire water tank) into the pre-trained target detection model, outputting the detection box coordinates and category confidence. When the confidence is greater than a preset threshold (e.g., 0.85), it is considered a successful recognition, and the recognized equipment category (e.g., a flow switch) and its bounding box are output. The spatial coordinates of the recognized equipment (e.g., x, y, z coordinate values ​​obtained from a 3D laser scanner) are spatially matched with the set location field in the fieldized standard knowledge base (e.g., the theoretical coordinate range on the outlet pipe of the high-level fire water tank). The spatial matching distance threshold is set to 0.5 meters. If the equipment coordinates are within the threshold range of the set location coordinates and the category is successfully recognized, an existence judgment result (e.g., "exists") is output; otherwise, a missing judgment result (e.g., "missing, the location should be set to the outlet pipe of the high-level fire water tank") is output.

[0065] S542. Compliance Detection Execution. When the existence detection result is "existence," a compliance detection is executed. The language reasoning module identifies numerical comparison, location comparison, or state comparison tasks, and calls the calculation execution module to compare the measured values ​​in the structured data fields with the compliance judgment rules in the fielded standard knowledge base. The comparison type for compliance detection is automatically determined based on the judgment rules associated with the standard index. For numerical comparison, it checks whether the measured value is within the threshold range; for location comparison, it checks whether the spatial coordinates are within the allowable area (e.g., the installation coordinates of the water flow indicator collected in S22 deviate less than the allowable deviation from the required position coordinates in the design drawings, resulting in a qualified result); for state comparison, it checks whether the state string completely matches the standard value (e.g., the control cabinet's manual / automatic state collected in S22 is "automatic," and the standard state is "automatic," resulting in a qualified result). After the calculation execution module returns a Boolean result, the language reasoning module outputs a qualified or unqualified judgment result, recording a deviation description for unqualified items (e.g., "The control cabinet's manual / automatic state is manual, which does not meet the automatic state requirement").

[0066] S543. Response Test Execution. When the compliance test result is qualified, the response test is executed. The calculation execution module calculates the time difference between the trigger signal issued by the fire protection facility trigger module and the action feedback signal of the tested equipment. The trigger signal timestamp is taken from the trigger output time recorded in S35 (e.g., the time T1 when the end test valve is opened), and the action feedback timestamp is taken from the feedback signal acquisition time recorded in S26 (e.g., the time T2 when the water flow indicator contact closes). The difference between the two is calculated in milliseconds (i.e., the difference between T2 and T1). The calculation execution module compares this time difference with the allowable response time in the fieldized standard knowledge base. The allowable response time is determined according to the judgment rules associated with the standard index. If the time difference is less than or equal to the allowable response time, the response qualification judgment result is output; otherwise, the response failure judgment result is output, and the actual response time is recorded.

[0067] S544, Linkage Detection Execution. When the responsiveness detection result is qualified, the linkage detection is executed. The pre-trained large model compares the set of devices whose action feedback signals are actually collected by the signal acquisition interface group after the fire protection facility triggering module issues a trigger signal with the set of linkage devices corresponding to the trigger signal in the field-based standard knowledge base. The set of linkage devices is determined by the linkage device set field associated with the standard index (for example, the set of devices that should be linked after the end-of-line test device is turned on includes "water flow indicator, alarm valve pressure switch, hydraulic alarm bell, high-level fire water tank flow switch, fire pump"). The actual feedback set is generated by deduplicating the feedback signals of each device collected in S26 according to the device identifier (for example, the actual collected feedback signals include "water flow indicator, alarm valve pressure switch, hydraulic alarm bell, fire pump"). The calculation execution module performs a set comparison operation to determine whether the actual feedback set contains all elements of the linkage device set. If it does (for example, all five devices provide feedback), the linkage qualified judgment result is output; otherwise, the linkage unqualified judgment result is output, and the list of missing devices is recorded (for example, "high-level fire water tank flow switch has no feedback signal").

[0068] In this embodiment, S6 specifically includes the following sub-steps.

[0069] S61. Report Template Determination. The report template is determined based on the testing standards system. The template is a predefined structured document template using Extensible Markup Language (EXPLAIN) format, containing four sections: Basic Information, Sub-item Judgment Results, Key Data Records, and Hazard Analysis and Recommendations. The Basic Information section includes fields such as the name of the inspected building, building type, usage, fire hazard category, testing date, testing standard, and testing organization name. The Sub-item Judgment Results section is divided into multiple sub-tables according to the fire protection subsystem. Each sub-table contains columns such as the tested object, installation location, standard index, existence judgment result, compliance judgment result, responsiveness judgment result, linkage judgment result, and comprehensive judgment conclusion.

[0070] S62. Summarizing Judgment Results. The judgment results of existence detection, compliance detection, responsiveness detection, and linkage detection generated in S54 are summarized according to the detection points and detection objects. The summarization process is executed by a pre-trained large model. The input is all structured data fields and judgment results at each level, and the output is a summary dataset categorized by subsystem. The pre-trained large model extracts the standard index and detection item name corresponding to each detection object from the field-based standard knowledge base through a retrieval-enhanced generation architecture. It associates the judgment results with the standard index and sorts them according to the detection point sequence to generate intermediate data objects used to populate the report template. For missing devices, the summary data records that the detection object is missing and marks its required location; for non-conforming items, the summary data records the deviation value and direction of deviation between the measured value and the standard threshold.

[0071] S63. Report Content Population. The summarized judgment results are filled into the corresponding sections of the report template. The pre-trained large model, based on the field mapping relationship of the report template, fills the values ​​from the intermediate data objects into the corresponding positions in the template. For the sub-table of judgment results, the pre-trained large model generates it row by row, with each row containing the detection object name, setting location, standard index, existence judgment result, compliance judgment result, responsiveness judgment result, linkage judgment result, and comprehensive judgment conclusion. For the key data record section, the pre-trained large model extracts the key measured values ​​recorded in S543 and S544 and fills them into the table in key-value pair format.

[0072] S64. Natural Language Summary and Report Output. This section generates the natural language summary portion of the inspection report. Based on all judgment results, the pre-trained large model uses summary generation prompts to convert anomalies such as missing items, non-compliance, response timeouts, and missing linkages into natural language descriptions conforming to standard terminology, which are then filled into the hazard analysis and recommendation section. The generated report is output in a portable document format and stored in a local database, while simultaneously being uploaded to the city's fire remote monitoring system via a communication interface.

[0073] Therefore, this embodiment realizes full automation of the detection data acquisition, analysis and judgment and report output process.

[0074] In summary, the building fire protection facility robot inspection method based on multimodal full-domain perception and large-scale model intelligence provided in this application establishes a standardized inspection scheme generation mechanism by identifying information of the inspected building and automatically matching it with the corresponding inspection process and inspection standard system. Through an autonomous mobile platform equipped with a multimodal perception module and a fire protection facility triggering module, it achieves autonomous navigation, fixed-point docking, simultaneous acquisition of multi-source data, and active triggering of simulated fire signals within the building space, thus constructing a full-dimensional data acquisition capability from environmental perception to active intervention. By processing and analyzing the collected multimodal data through a pre-trained large-scale model, it performs hierarchical reasoning and judgment based on fire protection inspection engineering logic, achieving fully automated inspection across the entire chain, from component existence, compliance, and responsiveness to system linkage. By generating inspection reports by integrating various judgment results according to the inspection standard system, it achieves a streamlined process connecting inspection data acquisition, analysis and judgment, and result output. The synergistic effect of these technical features constructs a fully intelligent inspection closed loop covering inspection standard matching, autonomous perception and triggering, hierarchical reasoning and judgment, and report generation, comprehensively improving the standardization, inspection coverage, result reliability, and operational efficiency of building fire protection facility inspection.

[0075] like Figure 7 and Figure 8As shown in the figure, this application embodiment also provides a building fire protection facility robot detection system based on multimodal global perception and large model intelligence. This system, as a specific implementation carrier of the aforementioned detection method, includes the following modules.

[0076] The building type matching module 10 is used to obtain the usage nature, fire hazard category, and building type of the building under inspection, and automatically matches the corresponding testing standard system and testing process based on the identification results. This module reads building information model files or receives structured data entered through a human-computer interaction interface, parses fields such as the building's usage nature, fire hazard category, and building type, and uses these fields as search conditions to retrieve the corresponding testing standard system from a pre-built field-based standard knowledge base. This testing standard system is output in structured data form, including an index of national fire protection technical standard clauses applicable to the inspected building, along with their corresponding testing items, pass / fail thresholds, testing method descriptions, and pass / fail rules.

[0077] The multimodal sensing module 20 is used to collect multimodal data, including building environment status, operating parameters of fire protection facilities, and linkage signals of fire protection facilities. This module integrates a visual sensing group, an acoustic sensing group, a physical quantity sensing group, an electrical parameter sensing group, and a signal acquisition interface group. The signal acquisition interface group collects dry contact signals and analog signals from fire protection facilities via hard-wired connections. The data collected by each sensing group is converted from analog to digital and then packaged into a unified multimodal data stream, with an additional acquisition timestamp and spatial coordinate label.

[0078] The fire protection facility triggering module 30 is used to send a simulated fire signal to the fire protection system and activate the fire protection system after fire confirmation. The fire protection facility triggering module 30 contains multiple types of triggering units. While each triggering unit performs its triggering operation, the precise timestamp of the trigger signal is recorded by the system clock synchronization mechanism.

[0079] An autonomous mobile platform 60 is used to carry the multimodal sensing module 20 and the fire-fighting facility triggering module 30. This platform uses a tracked or wheeled chassis and integrates a lidar synchronous positioning and mapping module, an inertial measurement unit, and a high-precision encoder to achieve autonomous navigation and centimeter-level positioning within the building's interior environment. Based on the sequence of detection points generated in the detection process by the building type matching module 10, the platform plans the optimal movement path and sequentially stops at each detection point.

[0080] The data processing and judgment module 40 integrates a pre-trained large model. This pre-trained large model is configured to process and analyze the collected multimodal data based on the detection process, performing existence detection, compliance detection, responsiveness detection, and linkage detection to generate a comprehensive judgment result. During processing and analysis, the large model constructs a three-level context architecture, employing a hybrid reasoning architecture that combines language reasoning and numerical computation. For existence detection, the large model utilizes visual recognition capabilities to identify the presence of the detection object based on image data; for compliance detection, the language reasoning module calls the computation execution module to compare the measured values ​​with the compliance judgment rules; for responsiveness detection, the computation execution module calculates the time difference between the trigger signal and the action feedback signal of the detected device; for linkage detection, the large model compares the set of devices that actually collected action feedback signals with the set of linkage devices in the standard knowledge base.

[0081] The report generation module 50 is used to generate a test report based on the test standard system and the judgment results. This module determines the report template according to the test standard system, summarizes the judgment results output by the data processing and judgment module 40 according to the test points and test objects, and fills in the four sections of the report template: basic information, sub-item judgment results, key data records, and hidden danger analysis and suggestions, and finally outputs the test report.

[0082] The system provided in this application embodiment achieves standardized generation of detection schemes through building type matching module 10, realizes full-dimensional data collection and active triggering through the coordinated cooperation of multimodal perception module 20 and fire protection facility triggering module 30, realizes hierarchical reasoning and judgment through data processing and judgment module 40, and realizes automated output of detection reports through report generation module 50. The modules work together to form a closed loop of intelligent detection across the entire chain.

[0083] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0084] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for detecting building fire protection facilities using robots based on multimodal global perception and large-scale model intelligence, characterized in that, Includes the following steps: The system obtains the usage, fire hazard category, and building type of the building under inspection, and automatically matches the corresponding testing standard system and testing process based on the identification results. The multimodal sensing module collects multimodal data including building environment status, operating parameters of fire protection facilities, and linkage signals of fire protection facilities. The fire protection facility trigger module sends a simulated fire signal to the fire protection system and activates the fire protection system after fire confirmation. The multimodal sensing module and fire-fighting facility triggering module are carried out through an autonomous mobile platform; Using a pre-trained large model, based on the detection process, the collected multimodal data is processed and analyzed, and existence detection, compliance detection, responsiveness detection, and linkage detection are performed to generate a comprehensive judgment result. Based on the testing standard system and the judgment results, a test report is generated.

2. The method for robot detection of building fire protection facilities according to claim 1, characterized in that, The existence detection verifies whether the fire protection subsystem or its components are set up as required by the design, and generates a result indicating whether it is present or missing; the compliance detection determines whether the installation location, specifications, installation method, quantity and configuration of the fire protection subsystem or its components meet the requirements of the fire protection technical standards, and generates a result indicating whether it is qualified or unqualified. The responsiveness test verifies whether the fire-fighting equipment can make the expected response action when it receives a trigger signal, and generates a judgment result of whether the response is qualified or unqualified; the linkage test verifies whether all the equipment that should be linked is activated after the trigger signal is issued, and generates a judgment result of whether the linkage is qualified or unqualified.

3. The method for robot detection of building fire protection facilities according to claim 2, characterized in that, The existence detection, compliance detection, responsiveness detection, and linkage detection are executed in a hierarchical order, with the existence detection being executed first. When the existence detection result is missing, compliance detection, responsiveness detection, and linkage detection are no longer executed for the missing detection object, and it is marked as inapplicable.

4. The method for robot-based detection of building fire protection facilities according to claim 1, characterized in that, The testing standard system is linked to a pre-built field-based standard knowledge base, which includes: standard index field, testing object field, setting location field, pass / fail judgment rule field, testing method description field, and linkage equipment set field. The testing process is generated by the autonomous mobile platform based on the spatial information of the building under inspection and the field-based standard knowledge base. The testing process includes: testing point sequence, standard index corresponding to each testing point, fire protection subsystem and components to be tested corresponding to each testing point, trigger operation type corresponding to each testing point, and multimodal data type to be collected corresponding to each testing point.

5. The method for robot-based detection of building fire protection facilities according to claim 1, characterized in that, The multimodal sensing module includes: The visual perception group is used to acquire image data and spatial coordinate data. The acoustic sensing group is used to collect sound pressure level data and ultrasonic data; The physical quantity sensing group is used to collect pressure data, flow data, liquid level data, and distance data; The electrical parameter sensing group is used to collect current data, voltage data, and frequency data. The signal acquisition interface group is used to acquire dry contact signals and analog signals of fire protection facilities through hard-wiring.

6. The method for robot detection of building fire protection facilities according to claim 1, characterized in that, The fire protection facility triggering module includes: Mechanical triggering unit, used to operate valves or buttons of fire protection facilities via a robotic arm and end effector; An electrical signal triggering unit is used to send a simulated contact closure signal to the fire protection facility control cabinet via an electrical interface; The simulated heat source generating unit is used to apply a simulated heat source to the temperature sensor. The simulated smoke generating unit is used to apply simulated smoke to the smoke detector.

7. The method for robot-based detection of building fire protection facilities according to claim 1, characterized in that, The autonomous mobile platform plans its path based on the spatial information of the inspected building and moves to each inspection point according to the inspection points determined in the inspection process. At each inspection point, for the inspection object that needs to be triggered, a simulated trigger signal is sent to the inspection object through the fire protection facility triggering module, and the inspection data is collected through the multimodal perception module. The collected multimodal data is transmitted to the pre-trained large model for analysis and processing.

8. The method for robot detection of building fire protection facilities according to claim 1, characterized in that, Before being processed by the pre-trained large model, the multimodal data is first converted into structured data fields. Each structured data field includes: standard index, detection object, setting location, sensor type, value, unit of measurement, collection timestamp, and spatial coordinate label.

9. The method for robot-based detection of building fire protection facilities according to claim 1, characterized in that, The pre-trained large model constructs a three-level context architecture during processing and analysis, and performs detection and analysis using a hybrid reasoning architecture that combines language reasoning and numerical computation. The three-level context architecture includes: A short-term context is used to carry the structured data fields collected in real time at the current detection point; The context is used to carry a structured summary text generated by the pre-trained large model, which contains a summary of the judgment results of all detection items in a subsystem or a fire compartment; Long-term context is used to retrieve the current detection record, historical detection archives, and the fieldized standard knowledge base from the cloud database as context by enhancing the generation architecture through retrieval; In the hybrid inference architecture, the language inference module identifies the detection task type and calls the computation execution module to perform numerical comparison, time difference calculation and set comparison operations.

10. A robot detection system for building fire protection facilities based on multimodal global perception and large-scale model intelligence, characterized in that, include: The building type matching module is used to obtain the usage nature, fire hazard category and building type of the building under inspection, and automatically match the corresponding testing standard system and testing process based on the identification results; The multimodal sensing module is used to collect multimodal data, including building environment status, operating parameters of fire protection facilities, and linkage signals of fire protection facilities; The fire protection facility triggering module is used to send a simulated fire signal to the fire protection system and activate the fire protection system after fire confirmation. An autonomous mobile platform is used to carry the multimodal sensing module and the fire-fighting facility triggering module; The data processing and judgment module integrates a pre-trained large model; the pre-trained large model is configured to process and analyze the collected multimodal data based on the detection process, perform existence detection, compliance detection, responsiveness detection and linkage detection, and generate a comprehensive judgment result. The report generation module is used to generate a test report based on the test standard system and the judgment results.

Images