Combining In-Sensor Computing with Reasoning at the Edge for Low-Power Bearing RUL Prediction

The increasing demand for predictive maintenance in industrial environments requires edge-intelligent solutions in which latency, energy consumption, memory footprint, and data movement are strictly constrained. This paper presents a two-stage Edge AI architecture for rolling-bearing prognostics, designed for heterogeneous embedded deployment across two distinct processing levels: a smart programmable sensing unit for local low-complexity signal preprocessing and a low-power embedded multiprocessor for higher-level temporal prognostic reasoning. In the first stage, a tiny neural preprocessor processes high-frequency vibration measurements directly at the sensing level and produces a compact low-dimensional degradation representation, complemented by lightweight health-related physical features. In the second stage, an embedded temporal reasoning model analyzes sequences of these compressed representations to estimate bearing degradation and remaining useful life. In addition to numerical remaining useful life regression, the second stage includes diagnostic reasoning heads that produce maintenance-oriented categories from a restricted vocabulary, enabling interpretable diagnostic summaries without relying on cloud-based language models. The complete pipeline is designed for fully edge-resident operation and exported in deployment-compatible formats, with the objective of supporting practical integration into heterogeneous industrial edge platforms. The proposed framework is trained and evaluated on the PRONOSTIA bearing degradation dataset and positioned against representative recurrent and hybrid prognostic baselines from the literature. From the deployment viewpoint, the sensor-side stage requires 68.62 ms inference time with 2.07 KiB RAM and 1.35 KiB flash/weights, whereas the edge temporal stage runs in 49.2 ms with 90.68 MiB RAM and 67.71 MiB flash/weights. In terms of prognostic performance, the proposed model achieves an average normalized RMSE of 0.1616 and an average normalized MAE of 0.1311 on three held-out bearings, while the weakly supervised diagnostic heads reach accuracies of 0.9066 for degradation trend and 0.8872 for vibration evidence. Experimental results show that the proposed architecture provides an effective trade-off between compact sensor-side processing, temporal prognostic accuracy, monotonic degradation consistency, hardware deployability, and interpretable maintenance-oriented outputs, supporting the feasibility of fully edge-based predictive maintenance systems for rolling-bearing health monitoring.