NanoSentry: An Edge‐Deployable Unified Neural Architecture for Multi‐Domain Industrial Prognostics and Health Management

ABSTRACT Edge‐deployed prognostics and health management (PHM) systems face a fundamental tension: state‐of‐the‐art deep learning models for remaining useful life (RUL) estimation exceed microcontroller memory limits. We introduce NanoSentry, a sub‐50 KiB unified neural architecture template that performs RUL regression and health‐state classification across heterogeneous industrial domains without quantisation or pruning. NanoSentry combines dilated causal convolutions, a novel CrossSensorGate for adaptive feature weighting, and a gated recurrent unit into a compact pipeline. Domain‐specific instances share this topology, varying only input and output dimensions. Evaluated on NASA C‐MAPSS, NASA Battery and CWRU bearing benchmarks using strict leakage‐controlled splits, NanoSentry matches or outperforms baselines its size. Crucially, it reduces critical‐zone (end‐of‐life) RMSE by up to over comparable CNNs, which pathologically fail near failure, and achieves 97.7% accuracy on leakage‐free bearing fault diagnosis. Hardware profiling on a physically flashed STM32H7 microcontroller confirms sub‐3 ms measured inference latency and up to 26 KiB peak SRAM for RUL tasks. To our knowledge, NanoSentry is the first sub‐50 KiB architecture template simultaneously evaluated across turbofan, battery and bearing PHM tasks with physical on‐device deployment evidence.