<p>With the rapid development of intelligence technology, inspection robots for high-speed Electric Multiple Unit (EMU) maintenance have emerged as a promising solution. However, a mature defect detection algorithm and framework tailored specifically for high-speed EMU inspection robots remains lacking. To address this gap, this paper proposes a position-prior-based multimodal defect detection framework capable of accurately identifying two primary types of component defects: component missing and bolt looseness. The proposed framework in this paper is designed with a two-stage methodology, which includes a component detection stage and a defect detection stage for identifying missing components and bolt looseness. To address the component missing defect, the paper initially established defect-free standard images and corresponding component positions for the same vehicle model. Following this, a Class-Similarity Iterative Closest Point (CS-ICP) approach is introduced. The proposed method represents the component positions from both the defect-free reference and the inspected images as semantic points enriched with category information, and diagnoses component absence by performing point cloud registration on these semantic point sets. To address the bolt looseness defect, the point cloud in the bolt area is first preprocessed via filtering. Then, a RANSAC algorithm incorporating normal vector constraints, as proposed in this study, is employed to precisely segment the top and bottom planes of the bolt. Based on this, the bolt height is computed to identify potential looseness defects. The proposed method has been experimentally validated on both the high-speed EMU inspection robot platform deployed in a high-speed railway maintenance depot and the Train of EMU failures Detection System(TEDS). Experimental results demonstrate that the method accurately diagnoses the primary types of component defects, exhibiting high accuracy and strong practical applicability.</p>