John C. Mossing
Mohammad A. Karim
Dept. of Elec. Eng., Univ. of Dayton, 300 College Park, Dayton, OH 45469
Acoustic sensors can be used to passively detect, track, and identify non-line-of-sight sound sources. The sound sources discussed in this paper contain strong harmonic components and range from being ``pseudo'' stationary to transient in nature. This research project investigates the use of various time-frequency analysis techniques for the purpose of selecting features to be extracted from the acoustic signatures of motorized sound sources. Acoustic data were filtered and digitized using a commercially available analog-digital convertor. The short time Fourier transform (STFT) was used as the initial identifier, followed by narrow-band peak detection algorithms used to select peaks above a user defined SNR. When sufficient ``nonsmeared'' peaks exist, these peaks are used to generate a set of harmonically related features which are input to a bayesian quadratic classifier. When the signal became more ``nonstationary'' and the use of the STFT resulted in excessive ``smearing,'' wavelet and Wigner-Ville analysis technique were employed. Results are presented highlighting the trade-offs of these time-frequency analysis techniques. Additional results are presented discussing the harmonic relationship algorithm employed to normalize the frequency shifts (due to rotor/engine RPM fluctuations) which exist in this type of signature data.