A Low-Power Convolutional Neural Network Implemented in 40-nm CMOS Technology for Bearing Fault Diagnosis

This paper presents a machine learning-based method for diagnosing bearing faults in electric motors using a Convolutional Neural Network (CNN) that processes motor current signals. This approach, aiming for real-time detection with cost-effective hardware, employs DoReFa-Net for parameter quantization, reducing memory and computational needs, thus enhancing efficiency and affordability. The developed CNN hardware accelerator, implemented in 40-nm CMOS and incorporating power-gating technology, operates efficiently at 100 MHz with just 11.18 mW power consumption, achieving an impressive 95.93% accuracy in bearing health condition assessment. This innovation offers automatic, precise monitoring of motors, significantly cutting staffing costs and preventing failures due to severe faults.