A feature extraction method based on probabilistic Principal components analysis and sampling importance resampling for bearing fault detection

The applications of monitoring the equipment online are often limited by the practical signal processing, limited by storage and transferring capacities. The efficiency is a key problem. Thus, a novel highly efficient feature extraction model for evaluating the equipment performance is proposed, which consists of the probabilistic Principal components analysis and the second generation wavelet analysis with the sampling importance resampling method. It starts by transforming raw signals into the wavelet domain by the second generation wavelet packet analysis. Then a sampling-importance resampling procedure is applied to reduce the redundancy and retain the distribution information. The obtained features are then fed into a probabilistic principal components analysis model to reduce the dimensionality. the proposed model is validated in a rolling element bearing test which shows that it is not only effective in diagnosis, but also may save the processing time of feature extraction, the data transfer bandwidth and the storage space.