Automatic data-driven spectral analysis based on a multi-estimator approach

Abstract

In signal processing, spectral analysis is widely used but, whereas computing the power spectral density (PSD) by Fourier approaches is relatively easy, its analysis and reading are much more demanding especially for spectrally rich signals. This paper presents an original method which automatically picks out and estimates the relevant spectral structures of an unknown random stationary process, embedded in an unknown non-white Gaussian noise. First, a statistical hypothesis test is applied to each local maximum value of the estimated PSD to detect the potential spectral peaks of interest. Second, an original feature space is proposed for classifying and characterizing the detected structures. Then, one key idea of the proposed strategy is to use not only one spectral estimator but to combine the results of different ones, taking benefits of their good properties. Therefore the detection and classification steps are applied to different spectral estimations. A last fusion step outputs a complete attribute vector, including a confidence index, for each detected structure. Another key idea of this data-driven approach is that all parameters are automatically set up without a priori knowledge. This approach is fully adapted to the preventive maintenance of complex systems, as illustrated in the paper.