Journal of Pure and Applied Ultrasonics

K. Mohankumar ¹, M. H. Supriya ¹, P. R. Saseendran Pillai ¹

Affiliations
1 Department of Electronics, Cochin University of Science and Technology, Cochin-682022, Kerala, IN

Abstract

Higher order spectral analysis is becoming one of the key areas in marine acoustic signal processing, especially for the classification and identification of noise generating mechanisms. This paper investigates the feasibility of a target classifier with bicepstral features. Bicepstral coefficients are extracted from the noise data waveform after framing and windowing operations. These features are then used to train a support vector machine classifier. Various kernels are analyzed for the performance measure. It is found that SVM with a radial basis function kernel could achieve a success rate of 80.5%. The kernel parameters were chosen after performing grid search and cross validation. The result indicates that the bicepstral features could be utilized efficiently to form feature sets that could greatly aid in the classification process.

Keywords

Spectral Analysis, Bicepstral Coefficients, Vector Machines, Ocean Noise.

Full Text

   

Suraj Kamal ¹, Shameer K. Mohammad ¹, P. R. Saseendran Pillai ¹, M. H. Supriya ¹

Affiliations
1 Department of Electronics, Cochin University of Science and Technology, Cochin-682022, IN

Abstract

Sounds often do not occur in isolation rather as a mixture comprising of emanations from a multitude of sources which are simultaneously active in the acoustic scene. Sonars typically operate in turbid acoustic environments that are rendered by numerous noise sources, both man-made and natural. This poses immense challenges in the detection and classification of target signals as multiple sources are simultaneously active in the medium. The convolutive mixing found in the underwater ambience makes the situation more complex as they vary much with the scattering nature of the ambience in which they occur. In this paper, Blind Source Separation (BSS) technique has been used for retrieving a best representation of the original signal before they get mixed up in the channel. Independent Component Analysis (ICA) is the widely used realization method of BSS that relies on the information theoretic criteria of statistical independence of the sources to unmix the observed signals to yield the source components. Most of the implementations of ICA work only for linear mixtures, so that additional assumptions need to be incorporated for unmixing convolutive mixtures. Spatio-temporal extensions for ICA have been used for unmixing the convolutive mixtures. After estimating the source signals, the characteristic features like MFCC and LPCC are determined and used as input feature vector for a classifier based on VQ and Euclidean distance. The results clearly demonstrate the performance advantages obtained by ICA preprocessing.

Keywords

Blind Source Separation, Independent Component Analysis, Spatio-Temporal ICA, Target Classifier.

Full Text