Real-Time Multimodal Defect Detection and MES Feedback on Edge Devices via Bayesian Fusion and Causal Adaptation

Conventional single-modality or loosely coupled inspection systems often degrade under partial occlusion, transient sensor dropout, or illumination drift, and their feedback to Manufacturing Execution Systems (MES) can be delayed or non-deterministic. We present D3Lite-MES, a real-time multimodal defect detection framework that i) aligns heterogeneous sensors at both hardware and software levels, ii) fuses them via a Bayesian–Causal Weighted Fusion (BCWF) module, and iii) adapts model parameters on edge devices through a lightweight causal adaptation routine. Deployed on commodity edge hardware, the system sustains line-takt operation and writes verified defect events back to MES with bounded latency. On three production lines covering [product families anonymized], D3Lite-MES attains a mean F1 of 98.2% and mAP@0.5 of 97.6%, reducing false alarms by 98.2% compared with lightweight baselines, at an end-to-end latency of 34 ms and average power of 8.9 W. Ablations show that BCWF contributes + 1.7% F1 under partial observations, while causal adaptation yields + 1.7% under illumination drift. The MES feedback loop decreases manual triage time by 34% and shortens recovery from minor stops by 34%.