A multi-task deep learning model for rolling element-bearing diagnostics

Abstract

The rolling element bearing is the predominant source of failures in rotating machinery. Therefore, detecting the corresponding faults, predicting their locations and measuring their severity is of immense importance. Classical intelligent diagnostics approaches rely on feature extraction methods followed by a classification model. Recently, deep learning models improved the fault classification accuracy by learning a suitable representation directly from the raw sensor data. In this work, we present a novel multi-task deep convolutional neural network trained end-to-end on raw vibration data to learn a shared representation for fault isolation and fault size evaluation. The proposed model architecture is constructed by stacking blocks of convolution layers, pooling layers, and batch normalization layers followed by a regression head and a classification head. Extensive experiments show that the proposed approach produces a superior performance to other existing methods and generalizes well to fault sizes not present in the training set.