Optimized reverse distillation method for nut and bolts anomaly detection on edge devices

Nut and bolts, critical yet defect-prone components in industrial machinery, traditionally rely on costly and error-prone manual inspection. In contrast, vision-based alternative methods offer efficiency but struggle in complex industrial environments, where defects are small and often obscured by background noise. Moreover, real-time processing is hindered by the reliance on remote cloud servers. To address these challenges, this paper proposes an optimized reverse distillation method specifically designed for nut and bolts anomaly detection on resource-constrained edge devices. Our core innovation lies in the Lightweight Attention Student Decoder, which integrates lightweight upsampling with a CBAM-based attention mechanism. This design significantly enhances the ability to discern small, often occluded defects characteristic of nuts and bolts within noisy environments, while maintaining computational efficiency for edge deployment. The entire system, deployed on an edge device (Raspberry Pi 4B), achieves real-time inference and utilizes a LoRa-based wireless network for result transmission. Extensive experiments on both the MVTec AD and MIAD datasets demonstrate that our optimized reverse distillation approach achieves state-of-the-art detection accuracy and location for nut and bolt anomalies. Crucially, the system proves effective when deployed in practical, resource-limited industrial scenarios.