Vibroacoustic source separation using an improved cyclic Wiener filter

Noise radiated by rotating and reciprocating machines is often a mixture of multiple complex sources, the successful reduction of which is a field of intensive research. In this paper an advanced source separation approach is presented, based on cyclic Wiener filtering, which takes into account the cyclostationarity property of the signals. The aim of the Wiener filter is the separation of noisy measurements into their contributions from the N specific sources and the remaining “noise”. Traditionally this can be achieved by using reference signals which are strongly coherent with the sources of interest and uncorrelated with all the other interfering sources and the masking noise. The Wiener filter can be estimated using the raw signals or only their random part. Moreover, the filter can be underestimated if the Signal-to-Noise Ratio of the reference signals is low, thus leading to the paradox that the level of an extracted source contribution is higher than the overall level. In this study a general strategy is proposed in order to select over which part of the signals (raw or residual) should the filter be estimated. This strategy is based on the number of the available references and the expected number of sources and the link with the multivariable statistical regression. Moreover, in order to increase its robustness, it is proposed to estimate the cyclic Wiener filter using an additional constraint which imposes that the sum of the contributions of the periodic parts of each source equals the overall periodic part as is calculated by the synchronous averaging procedure. This produces a new estimator of the Wiener filter, obtained from a constrained least square optimization. The proposed method is applied on vibroacoustic signals captured on a test rig in order to quantify the contributions of “hydraulic noise” (originating mainly by four hydraulic pumps) and “mechanical noise” (originating from the various rotating parts of the engine).