Transit Underdetermined Blind Source Separation Based on ICEEMDAN and FastICA for Switch Machine Locking Signal

It is the fundamental for achieving railway safety to ensure that the switch machine is fully locked. This article discusses a blind source separation method for locking the sound signal generated by the switch machine using the improved adaptive empirical mode decomposition (ICEEMDAN) and the fast independent component analysis algorithm (FastICA) in single channel. A new internal mode function component screening criterion is proposed based on the sound characteristics of locked signals. The ICEEMDAN method decomposes the one-dimensional collected sound signal into multiple IMFs. Then, based on the proposed screening criteria, select an IMF that meets both the correlation threshold and kurtosis threshold to form a new matrix. Finally, perform a negative entropy based FastICA on the matrix and select the component in the matrix that has the highest similarity to the locked signal spectrum as the separation result. A comparative study was conducted on various sound signal decomposition methods. The results react that the ICEEMDAN method selected in sound separation has better performance in indicators than other methods, including the spectral mean correlation coefficient (SMCC) and the kurtosis difference mean absolute error (KDMAE).