Name Detection and Fuzzy Classification of Transient Signals in the Time – Frequency Plane
Funding Reference FCT - POSI/32708/CPS/2000
Dates 2000|2004

New techniques for passive detection and classification of underwater acoustic transient signals are developed and tested. The focus is on theoretical and algorithmic aspects so as to achieve an acceptable compromise between optimality and robustness to model mismatches, and computational efficiency. Several classes of transients are considered, e.g., man made, mammals’ signatures, and spiky noise generated by hydrothermal vents, covering a diversity of frequency bands and physically appropriate source models (deterministic and stochastic). The methods here proposed match the temporal non-stationary of transient signals. Observation noise is assumed either Gaussian or non-Gaussian impulsive.

Data representation involves filtering and sampling the received signal, followed by a linear decomposition using the discrete wavelet transforms with compactly supported short duration filters. Assuming that a delayed signal is correctly represented by its delayed coefficients, the process described is optimized, yielding the best compromise between performance and computational complexity. This implies choosing the observation intervals, the sampling frequencies, the likelihood test rates for real-time processing, and the design of the mother wavelets. Sub optimum processors are also developed for multipath ambient, assuming random multipath attenuations and delays. The proposed approach can increase the robustness of the resulting detector, requiring much less computations than the generalized likelihood ratio test.

Translating classical detection techniques to the Time-Frequency (TF) plane does not produce better detection statistics. However, working in the TF plane provides a significant advantage: more powerful pre and post processing allow operation in lower SNR’s (<-5 dB). By adjusting the TF kernel, distinct sub optimal detectors result and the best suited for each specific transient can be selected. Additional complexity, due to bidimensional correlation, is combated using a generalized distribution, representing the transient as a delta distribution. This square root the computational cost. The design of that distribution, being trivial for polynomial phase signals, is generalized to accommodate the transients considered.

The performance of the proposed techniques is evaluated based on the theoretical analysis of the algorithms developed and/or on computer experiments driven by simulated and real data.

Research Groups Signal and Image Processing Group (SIPG)
Project Partners Escola Naval (Portuguese Navy School)
ISR/IST Responsible
Victor Barroso
[1] Francisco Garcia, Isabel Lourtie, "Estimation of Locally Covariance Matrices from Data", Proc. ICASSP2003 - IEEE International Conference on Acoustics, Speech and Signal Processing, Hong Kong, 2003 - PDF