Investigation of Transfer Learning Method for Motor Fault Detection

Industry 4.0 is propelling modern industries forward due to its reliability, stability, and performance. Electric motors (EMs) are utilized in multiple industries for their efficiency, precise speed and torque control, and robustness. Detecting faults in motors at an early stage is crucial to ensure maximum productivity. Recently, DL has been implemented as a data-driven approach for detecting faults in motors. However, due to the limited availability of labeled fault data, the performance of the DL model is constrained. This issue is addressed by leveraging transfer learning (TL), which uses knowledge from a larger source domain to improve performance in a smaller target domain. In this paper, a multiple fault detection (FD) model for EMs is proposed by combining the ideas of deep convolutional neural networks (CNNs) and TL. A one-dimensional signal-to-image conversion technique is suggested for converting the vibration signal to images, and an Inception-ResNet-v2-inspired FD model is proposed for detecting bearing faults in the motor. The proposed method achieved a mean accuracy of more than 99%.