Abstract:
Bearing is a critical machine element whose fault-free health status is crucial to the reliable operation of the machine. In recent times, machine learning-based approaches of training models on huge amount of measured data are adopted to monitor the bearing health status. However, collecting a large number of complete fault samples in an industrial set-up is very expensive thereby fostering the need to introduce methods that have excellent classification and clustering performance without depending on an extensive amount of training data. To address this, a dynamic time warping (DTW)-based signal similarity measuring approach is proposed. In this paper, a reference signal corresponding to a particular bearing health class is generated, which is then warped by DTW on any test bearing signals to reveal the inherent similarity by calculating their cumulative distances. The distance measure is used to classify the bearing health. It is observed that DTW alone cannot obtain acceptable results when employed to handle complex bearing signals because of the presence of measurement noise and unwanted interfering components. An enhancement is proposed by integrating DTW-based similarity search with spectral kurtosis (SK)-based demodulation for enhanced classification. The proposed method is validated using the Case Western Reserve University (CWRU) bearing dataset.