A reduced-rank spectral approach to multi-sensor enhancement of large numbers of sinusoids in unknown noise fields

The enhancement of sinusoids in adverse noisy environments finds applications in many fields. Among the numerous methods that have been proposed to perform this task, few can actually operate without any information on neither the sinusoids nor the noise characteristics. In this paper, we propose a non-parametric spectral approach inspired by the adaptive line enhancer (ALE) principle, which has the capability of detecting and enhancing an arbitrary large number of tones in additive noise of unknown characteristics. Into addition and contrary to the classical ALE, it can take advantage of the extra information provided by multiple sensors so as to significantly increase its performance. The idea essentially consists in a reduced-rank regression between the measured signals and their delayed versions at each frequency of interest, and its implementation is carried out within a general “spectral estimation” framework which encompasses classical estimators such as the WOSA, the multi-taper, and the lag-window. This allows the design of enhancement filters with improved frequency selectivity. The rationale of the method is supported by a statistical performance analysis, and its efficiency is eventually demonstrated on actual industrial signals.