A review of multimodal-based defect detection for industrial parts

Abstract

Defect detection of industrial parts is a crucial link to ensure product quality, production safety, and stable operation of intelligent manufacturing systems. Traditional machine-vision inspection is efficient and low-cost, but single-modal methods are often limited by illumination changes, reflective surfaces, complex textures, internal cracks, and scarce defective samples. Multimodal defect detection addresses these limitations by fusing complementary information from vision, spectroscopy, acoustics, mechanics, thermal imaging, and 3D sensing, thereby improving the robustness and interpretability of defect recognition and localization. This paper reviews the main sensing modalities and their engineering characteristics, summarizes representative industrial anomaly detection datasets and computer-related algorithms, and discusses data-level, feature-level, and decision-level fusion strategies with typical mathematical formulations and a generic anomaly detection procedure. Furthermore, it analyzes application scenarios in aerospace, automobile manufacturing, precision electronics, and metallurgy, and discusses key challenges including multimodal registration, heterogeneous feature fusion, missing or degraded modalities, small-sample learning, real-time deployment, and trustworthy inspection. The review aims to provide a concise and reliable reference for researchers and engineers developing multimodal industrial defect detection systems.