The USC Andrew and Erna Viterbi School of Engineering USC Signal and Image Processing Institute USC Ming Hsieh Department of Electrical and Computer Engineering University of Southern California

Technical Report USC-SIPI-272

“Advanced Signal Processing for Hydroacoustic Data of Internal Launchway Flows”

by Dae C, Shin, George A. Tsihrintzis and Chrysostomos L. Nikias

October 1994

In this report, we present the results from a number of signal processing approaches to analyzing hydroacoustic data measured in test internal launchway flows. The ultimate goal is to establish features of the data that allow their reliable classification into one of two categories, namely hydrophone data from a smooth section of the flow or from the section of the flow near a slot, as these are described in the report. We applied classification algorithms based on a variety of signal processing concepts, including modeling via second-order autoregressive equations, power spectrum- and polyspectra-based features, cepstral feature extraction based on third-order statistics, and modeling with symmetric, alpha-stable (S_S) distributions. In our analysis, we examined the transient flows (first half of each data set) separately from the steady-state flows (second half of each data set) and for each one we established reliable classification features. In particular, we found that the coefficients of a second-order autoregressive model as estimated via Gaussian maximum likelihood methods are features providing clear classification power between smooth and slot section flow data. However, part of the classification power is lost if the autoregressive model estimates are obtained via Yule-Walker equations or their generalizations. It is noteworthy, that improvement was observed in the Yule-Walker approach when fractional, lower-order moments were employed. The power spectrum was another feature that allowed data classification, as was the phase of the bispectrum. In particular, it was observed that wide main lobes in the power spectrum or significant low frequencies in the phase of the bispectrum corresponded to flow data from the slot section, while narrow main lobes in the power spectrum or significant high frequencies in the phase of the bispectrum corresponded to flow data from the smooth section. On the other hand, the magnitude of the bispectrum did not provide clear classification power. Similarly, the cepstral parameters did not provide clear classification power either. As a conclusion, high classification power was provided by the Gaussian maximum likelihood estimates of the parameters of a second-order autoregressive model and the phase of the data bispectrum.

To download the report in PDF format click here: USC-SIPI-272.pdf (6.6Mb)