Implementation of a neural network based bicepstral classifier for marine noise sources

The higher order statistics (HOS) is being widely used for analyzing real world signals as it can reveal information about nonlinear signal generating mechanisms, which is otherwise impossible with conventional analysis or methods. The identification of signatures of such nonlinearities could be an important step towards the analysis of the signals, especially for classification problems. Applications of conventional techniques like cepstrum have been explored in a variety of areas including audio processing, speech processing, geophysics, radar and sonar signal processing. However, the cepstral analysis may fail in the presence of additive noises. Since higher order techniques like bispectrum and the trispectrum conserve the phase characteristic of the wavelet it is evident that the cepstrum derived from these polyspectra will also conserve phase information. This paper investigates the feasibility of realizing an intelligent classifier for marine noise signals, with the help of artificial neural networks, using higher order cepstral features. The results show that the bicepstrum analysis technique can effectively be used for extracting the features of marine noise sources, thereby providing a potential feature extraction method for classification problems.