基于多模态感知的实验室巡检消防机器人设计OA

[Objective]Laboratory environments contain dense hazardous chemicals,high-precision instruments,and complex electrical systems,resulting in significant fire risks.Conventional manual patrols are inefficient and prone to human error,and single-sensor monitoring systems lack robustness under dynamic conditions such as smoke interference,occlusion,and variable illumination.To address these challenges,this study designs and implements a multimodal perception-based laboratory inspection robot capable of autonomous navigation,intelligent fire detection,and precise fire suppression.The proposed system integrates multisource sensing,real-time mapping,decision-making,and execution,providing an intelligent,all-weather solution for laboratory safety management.[Methods]The proposed system adopted a hierarchical"Perception-Decision-Execution"architecture,implemented on an ROS-based control platform.The perception layer fused data from a 16-line LiDAR,a binocular vision camera,an infrared thermal imager,and an IMU to build a unified environmental model.Temporal and spatial synchronization among sensors was achieved through ROS-based soft time alignment and spatial calibration between vision-LiDAR and infrared-visible cameras,ensuring accurate data correspondence.For navigation and mapping,a SLAM-based grid mapping algorithm was implemented using a two-stage process:a front-end scan matching stage and a back-end pose graph optimization stage.The front-end generated submaps via probabilistic scan matching,while the back-end performed nonlinear global optimization with loop closure detection and inertial constraints,achieving centimeter-level positioning accuracy.Fire detection was realized through an improved YOLOv8 model trained on high-resolution flame datasets.The model performed real-time flame localization using three detection scales(80×80,40×40,and 20×20)with non-maximum suppression to ensure robust classification under complex lighting and occlusion.After fire localization,stereo depth estimation using the SGBM algorithm converted the detected 2D fire coordinates into 3D space,enabling precise spatial positioning of the fire source relative to the robot.The decision layer integrated fire position data,performed path planning,and calculated extinguishing parameters such as nozzle angle and flow rate.The execution layer included a Mecanum-wheeled mobile base and a PWM-controlled water pump and servo assembly for precise aiming and extinguishing.Feedback from the IMU and temperature sensors was continuously transmitted to the decision layer,forming a closed-loop control system for adaptive correction during operation.The entire hardware system was driven by an Orange Pi 5 Plus embedded controller,featuring an RK3588 processor with 6 TOPS NPU computing power,ensuring efficient real-time data processing and modular integration.[Results]Extensive experiments were conducted in a simulated laboratory environment.The autonomous navigation test demonstrated an average positioning error of 0.074 m and an orientation error of<1.5°,confirming the high precision of the SLAM-based localization.The YOLOv8 flame detection model achieved a detection rate of 95.8%,precision of 92.3%,and recall of 94.1%,with an average inference time of 28 ms per frame,ensuring real-time performance.In 50 standardized fire-extinguishing trials using alcohol lamps,paper,and oil-based flames,the robot achieved an overall extinguishing success rate of 96%,with an average response time of 9.2 s.The system effectively handled both stationary and path-planned fire scenarios,maintaining stability even under partial occlusion or low illumination.Failure cases were primarily attributed to minor nozzle misalignment or infrared interference,which could be further optimized through dynamic compensation algorithms.[Conclusions]The proposed multimodal perception-based laboratory-inspection firefighting robot achieved full-process automation from fire detection to extinguishing,demonstrating excellent perception accuracy,navigation precision,and execution reliability.The system effectively bridged the gap between static monitoring and dynamic response,enabling continuous 24/7 fire inspection and rapid suppression within 10 s of ignition.Its modular design ensured scalability for future integration with multiagent collaboration and intelligent scheduling.However,limitations remain in detecting smoldering or early-stage nonflame fires,and current extinguishing media are limited to water-based systems.Future work will focus on enhancing multispectral sensing for early smoke and heat detection,incorporating multiagent coordination,and developing adaptive decision models using deep reinforcement learning to further improve efficiency and safety in laboratory environments.