EDGE AI FOR INDUSTRIAL PREDICTIVE MAINTENANCE: A DUAL-ATTENTION APPROACH WITH EXPLAINABILITY AND UNCERTAINTY QUANTIFICATION

Industrial Internet of Things (IIoT) systems generate continuous sensor data streams from rotating machinery. Cloud based artificial intelligence achieves high accuracy for Remaining Useful Life (RUL) prediction but introduces 100-500 ms round-trip latency, which may be unacceptable for real-time failure prevention. Edge AI offers an alternative but faces computational, explainability, and security constraints that existing approaches rarely address jointly. This paper presents a unified framework integrating four dimensions: (1) a dual-attention neural architecture for RUL prediction that explicitly models channel-wise sensor importance and temporal relevance; (2) SHAP-based explainability with quantified overhead on edge hardware; (3) Monte Carlo Dropout for uncertainty calibration; and (4) a Zero Trust security layer for Operational Technology (OT) deployment. The framework targets the NVIDIA Jetson Orin Nano class of devices with an inference-latency goal of under 50 ms. This is a methodology paper; a six-phase experimental protocol benchmarks the approach against verified state-of-the-art (HMDAM: RMSE 10.82; DAST: RMSE 11.43 on NASA C-MAPSS FD001). Architecture and protocol are complete; experimental results will follow in an extended version. All quantitative claims are explicitly marked as literature-verified or target-under-validation.