Feature Engineering and Classification of Elastic Waves from Partial Wave Simulations of Active Sonar Targets

Active sonar target classification continues to be a challenging task due to the many complex processes occurring within the ocean. There have been numerous feature extraction algorithms and machine learning techniques developed and tested on both simulated and experimental data to aid in this solution. This work compares the time-frequency representations of spectrograms and scalograms. The scalogram representation is hypothesised to provide increased classification due to the wavelet transform ability to dilate and contract, thereby capturing various time and frequency resolutions while the spectrogram resolution is fixed. The data are simulated impulse responses from solid spheres of four different types of materials at various radii excited by a monostatic plane wave. Feature extraction of the impulse responses is performed on the time-frequency domain representations. Classification is performed using a 10-fold crossvalidated support vector machine. The scalogram representation demonstrated an increased classification of 26% when compared to the spectrogram representation. Results are reported as overall classification accuracy and confusion matrices.