On a method for testing ICA based blind source separation algorithm performance applicable in audio-based on-load tap changer diagnostics

The biggest challenges in using Blind Source Separation (BSS) algorithms, such as those based on Independent Component Analysis (ICA), arise from the inability to determine the reliability of the resulting independent components (signals). In sensitive areas, such as machinery diagnostics, such uncertainties could also have a negative impact on decision-making processes. For that reason, any additional confirmation that yields a better understanding of BSS algorithm capabilities and the issues that may arise from using this method in solving audio-based diagnostic problems is desirable. In this paper, the focus is placed on On Load Tap Changer (OLTC) audio-based diagnostics. The dominant audio signals that mix with the carrier of the useful diagnostic material, in this case, express stationary character. Given the fact that the targeted OLTC audio fingerprint usually represents a highly non-stationary signal that appears only in a certain period when compared to these interferences, it is possible to develop a source separation method based on a simple modeling approach. For that purpose, in this paper, a non-linear latest square curve fitting method was used for the extraction of the OLTC audio fingerprint, which was then used as a reference for testing the source separation efficiency of several different ICA algorithms.