Rivets are critical mechanical fasteners in steel bridges, and rivet defects may cause catastrophic failure. This study proposes a convolutional neural network (CNN)‐based inspection system for fast rivet identification and diagnosis. Rivet states are classified as normal, rusted, loose, and missing. A CNN‐based training workflow was introduced to develop a reliable rivet diagnosis system. A multiscale moving window searching technique was proposed to solve the challenge of small rivet identification. A continuous dataset enrichment strategy was applied, which improves training efficiency and minimizes training time. The model performance was assessed based on a historical bridge in Gjerstad. The proposed multiscale moving window searching technique significantly enhances the rivet identification rate to 96.3%. The classification accuracy and model robustness were evaluated, and conditions leading to unidentified rivets were discussed and summarized.