An End to-End Bearing Fault Diagnosis and Severity Assessment with Interpretable Deep Learning

Abstract

The induction motor (IM) consumes 85% of the electrical energy in industry applications. The crucial place occupied by the IM in the industry requires the installation of an efficient integrated machinery health management (MHM) process in order to improve the availability of this equipment and minimize the cost of maintenance. Characterized by a high potential of unsupervised feature learning, the Deep Learning (DL)-based MHM of IM has recently been introduced. However, it is still hard to interpret physically the features learned by most of the already implemented deep architectures. In this paper, an architecture called Deep-SincNet is implemented for bearing faults (BF) diagnosis task. The proposed scheme automatically learns the fault features from the raw vibration signals and accordingly finalizes the fault diagnosis process. As well as the features related to BF characteristic frequencies, the proposed architecture automatically extracts further interpretable features. The performance of the proposed approach is tested for two different databases. For several types of bearing faults, different levels of fault severity and variable loads, the Deep-SincNet provides high accuracy, a significant gain in implementation cost, faster convergence, and interpretable results