DEVELOPMENT OF AUTONOMOUS INDUSTRIAL INSPECTION ROBOTS USING COMPUTER VISION AND ARTIFICIAL INTELLIGENCE

Abstract

Autonomous inspection robots deployed in industrial facilities have so far relied predominantly on a single sensing modality, typically RGB computer vision, traversing a fixed, pre-programmed route regardless of the actual likelihood that any given asset harbors a developing defect. This design choice constrains both detection reliability, since single-modality vision struggles to distinguish genuine surface degradation from benign visual clutter such as dust, glare, or paint discoloration, and inspection efficiency, since a fixed route allocates equal attention to low-risk and high-risk assets alike. This study proposes Multi-Modal Fusion Autonomous Inspection (MMFAI), an architecture in which a LiDAR-based simultaneous localization and mapping module provides the navigation backbone, RGB, long-wave thermal, and structure-borne acoustic sensing streams are combined through a Bayesian belief-fusion layer that cross-validates candidate defect signatures across modalities before an alert is raised, and an information-theoretic next-best-view planner sequences inspection stops according to expected information gain rather than a fixed patrol route, prioritizing assets whose condition estimate carries the greatest uncertainty or the greatest consequence of failure. The framework was evaluated on a representative petrochemical processing unit through comparative field trials against manual human inspection and an existing single-modality RGB inspection robot operating a fixed route. The proposed system raised defect-detection recall to 94.2% and precision to 91.6%, extended mean early-detection lead time for corrosion and thermal-hotspot onset from 9.6 to 27.3 days, and reduced human exposure time in hazardous zones by 93% relative to manual inspection, with a seven-year net present value nearly three times that of the single-modality robot.