BearingNAS Enables In-Sensor Fault Diagnosis on Microcontrollers

The ability to perform intelligent fault diagnosis directly within sensors offers significant advantages for industrial applications, particularly for critical components like bearings. This paper introduces BearingNAS, a Hardware-Aware Neural Architecture Search (HW-NAS) framework designed to facilitate this "in-sensor" processing, even on highly resource-constrained micro-budgets (e.g., 4-8 kiB RAM, 16-32 kiB Flash). BearingNAS frames the architecture search as a constrained optimization problem, specifically targeting these extreme resource limitations. To make the process accessible and cost-effective, it employs a lightweight, derivative-free search strategy that can run entirely on a standard laptop CPU, converging in less than an hour. The framework utilizes a single data-flow search space with a decaying kernel growth formulation to prevent parameter explosion during the search. The framework was evaluated using the Case Western Reserve University (CWRU) bearing benchmark, optimizing architectures for various STMicroelectronics microcontrollers, including the LSM6DSO16IS Intelligent Sensor Processing Unit (ISPU). The best resulting in-sensor architecture achieved a highly competitive diagnostic accuracy of 99.50% on the ISPU, demonstrating the practical viability of shifting machine learning workloads directly into sensor packages for low-cost, production-scale fault diagnosis.