Manifold Learning Using Linear Local Tangent Space Alignment (LLTSA) Algorithm for Noise Removal in Wavelet Filtered Vibration Signal

A denoising procedure is proposed to remove both out-band and in-band noise for extraction of weak bursts in signal obtained from defective bearing. Energy of continuous wavelet scalogram is computed and the band having higher energy is selected to remove the out-band noise. Signals of selected band are brought together to form a high-dimensional waveform feature space. Further, low dimensional waveform manifold is formed using linear local tangent space alignment (LLTSA) algorithm to remove in-band noise. A criterion, entitled as frequency factor is also proposed to determine the optimum neighbour size of LLTSA. The two complicated conditions are chosen to demonstrate the effectiveness of the technique in the extraction of bursts in the noisy situations. A significant improvement in the signal to noise ratio is observed when in-band noise is removed using manifold learning by LLTSA algorithm. The experimental result reveals the success of the proposed denoising procedure in extraction of defect features, even in the case of noisy condition.