Abstract
Heat exchanger tubes in power plants and refineries degrade through fouling, scaling, corrosion and wall thinning, and their internal condition governs both thermal efficiency and plant safety. Inspection of these assets is still performed largely by manual or semi-automated review of remote visual inspection footage, which is time-consuming, subjective, and prone to inconsistent defect interpretation between operators. The difficulty is compounded by the imaging environment itself: tube bores are narrow, illumination is supplied coaxially with the camera and falls off with depth, metallic surfaces produce strong specular reflections, and probe motion introduces blur. This work proposes an AI-driven multi-modal vision framework for automated detection, localization and quantification of tube blockages and surface degradation under these conditions. The framework integrates image preprocessing, deep feature learning, a hybrid convolutional neural network and vision transformer (CNN-ViT) detector, and a reinforcement learning agent that adapts probe traversal speed to the observed defect risk. A quantitative Tube Health Index (THI) is introduced to score severity from corrosion coverage, blockage ratio, defect depth and structural scaling, so that each tube is assigned an objective risk category rather than a binary defect flag. The framework was validated on an industrial dataset comprising 353 heat exchanger tubes, approximately 180 hours of footage and 12,400 labelled defect instances acquired under field conditions. The proposed hybrid model achieved 94.6% accuracy, 92.3% precision and 91.1% recall, outperforming conventional thresholding at 72.4% accuracy and a standalone CNN at 88.1% accuracy. Inference ran at 48 ms per frame on embedded hardware, and inspection time per tube fell by approximately 60% relative to manual practice. The results indicate that combining hybrid feature learning with adaptive scanning and quantitative health scoring offers a deployable pathway toward autonomous industrial non-destructive testing.