A Physics Based Deep Learning Approach for Cross Domain Bearing Fault Detection

Bearing fault detection across varying rotational speeds of induction motors (cross-domain) is difficult due to challenges associated with different data distributions under different operating modes. To build a fault detection model, it is required to capture the domain invariant features not influenced by speed variations. Existing methods for bearing fault detection exploit data-driven deep learning models without considering any structural information of bearing. This paper proposes a physics-informed deep learning approach for bearing fault detection across different working conditions of the motor. The proposed method utilizes the power spectrum of the envelope signal (PSES) from the vibration data as an input feature for 1D CNN deep learning network. The proposed method is compared with the Random Forest and 1D CNN with raw data input and evaluated on two public bearing datasets to demonstrate its superior performance and robustness to speed variation. The proposed method is able to predict cross-domain bearing faults with an average accuracy of 94.2% which is 23% higher than the existing state-of-the-art data-driven approach.