DEEP LEARNING BASED BEARING FAULT DETECTION WITH MULTI-RATE SIGNAL PROCESSING FOR DATASET FROM VARIOUS SHAFT ROTATIONAL SPEEDS

This thesis research begins with the bearing fault frequency analysis, leading to the bearing characteristic frequency zones which provide significant information for signal feature extraction. Then, a fusion algorithm is proposed to extract a principal signal from x-, y-, and z- channel sensor recordings with signal enhancement in terms of signal-to-noise power ratio. To tackle the dataset from the different shaft rotational speeds, a multi-rate signal processing technique is proposed to address the challenge of bearing fault detection by regulating the dataset to the reference shaft speed. With the regulated dataset, the one-dimensional frequency spectral features, which become independent of the shaft speed, can be extracted for a one-dimensional deep learning network.