A Two-step Adaptive Blind Source Separation for Machine Sound
Jiawen Li, LI Cong-xin · 2006
The difficulties of machine sounds failure diagnosis lie in the detected interested signal being polluted by many other noises which led to low signal noise ratio. In order to extract interested machine sounds, an efficient blind separation algorithm was presented. It first extracts the p largest eigenvalues of covariance matrix of observed signals by simple parallel adaptive principal component analysis preprocessing algorithm, and then estimates the p source by natural gradient algorithm. The output signals are always the p largest energy components of X. Its preprocessing and separation steps all exploit adaptive approach. The algorithm can deal with super-Gaussian, Gaussian and sub-Gaussian signal, has low computation complexity and is suitable for real-time application. Simulations show that it is feasible and effective for blind source separation of distorted machine sounds