A New Unsupervised Learning Approach for CWRU Bearing State Distinction

As one of the most relevant components in rotary machinery, ball bearings play an important role in diverse areas. To research bearing health state and remaining useful lifetime, several datasets have been developed. Among these datasets, Case Western Reserve University (CWRU) dataset is the most commonly used for bearing diagnosis. A large variety of approaches are applied on CWRU dataset and generating good even the tendency of perfect results. However, most of these approaches are based on supervised learning approaches and focus on classification of bearing faults. In this contribution, in difference to well-known existing approaches, an unsupervised approach combining autoencoder with k-mean is applied on the CWRU dataset. Firstly, the original data are segmented into proper parts. Segments in time domain are transformed to time-frequency domain by adjusting the window length and window function using Short-Time Fourier Transform (STFT), and an associated spectrogram is generated. Spectrogram features are extracted using autoencoder and clustered using K-mean. Various metrics are used to evaluate the performance of the proposed approach. All metrics values demonstrate that this approach could distinguish CWRU bearing from fault-free state to faulty state. As a new result, the requirement of related training datasets of the other approach is – for fault detection – no longer necessary in the future.