“…Data-driven condition monitoring and fault diagnosis of a bearing normally consist of data acquisition from the bearing, signal processing and data classification steps. However, due to several important factors, e.g., friction, clearance, and variable working conditions, the acquired vibration signals from these rolling bearings are non-linear and non-stationary, which makes extracting fault feature information a difficult task [ 3 , 9 , 10 , 11 , 12 , 13 , 14 ]. Specifically, when using popular feature extraction methods that analyze features from the time domain, frequency domain, or time-frequency domain, it is very difficult to identify the fault characteristics under variable working conditions [ 15 , 16 , 17 , 18 , 19 , 20 ].…”