BearingNAS brings 99.5% fault diagnosis to microcontrollers using just a laptop

BearingNAS is a Hardware-Aware Neural Architecture Search (HW-NAS) framework that shifts fault intelligence directly onto the sensor die via in-sensor processing. To eliminate dependence on discrete GPUs, the researchers propose a lightweight, derivative-free search strategy paired with a single data-flow search space. A decaying kernel growth formulation prevents parameter explosion, enabling the framework to target extreme micro-budgets: 4 to 8 kiB of RAM and 16 to 32 kiB of Flash. Running entirely on a laptop CPU, the search converges in less than an hour, democratizing neural architecture design for resource-constrained edge devices.

The framework was evaluated on the Case Western Reserve University (CWRU) bearing benchmark, optimizing architectures for three STMicroelectronics targets: two commodity microcontrollers and the LSM6DSO16IS Intelligent Sensor Processing Unit (ISPU). The best in-sensor architecture achieved a highly competitive diagnostic accuracy of 99.50% on the ISPU, demonstrating that machine learning workloads can be effectively moved inside the sensor package. Accepted to IEEE COINS 2026, this work paves the way for low-cost, production-scale bearing fault diagnosis in industrial IoT and predictive maintenance applications.