Motor Bearing Fault Classification using Laser Sensor and Light Weight CNN

Riccardo Pretto, Aveen Dayal, Yonas Oghbasilasie Haile, José Amendola, Linga Reddy Cenkeramaddi, Kjell G. Robbersmyr, Ajit Jha

Abstract: Among the several faults, bearing fault is considered one of the most common and major problems in the motor. In this paper, we present end-to-end motor bearing fault detection and classification employing laser, wavelet transformation and light weight convolution neural network (CNN). The proposed method has three stages: first, raw vibration signatures from the motor shaft vibration operating with healthy and faulty bearing is captured using laser in point and measure fashion. Thereafter, these time domain signals are converted to time-frequency plane using wavelet transform. Finally, the features are extracted from that plane and used to train a light weight CNN model to detect and classify the bearing fault. Experimental findings demonstrates that the laser based method in tandem with CNN model can classify the bearing fault with 99 % accuracy. Further, the proposed model gives similar performance compared with standard models - SqueezeNet, and MobilNet with added benefits of much smaller model parameters and model size which enables efficient deployment on the edge.