Naslov
Performance Comparison of Blind Source Separation Algorithms for Nonstationary Signals

Autori
Saulig, Nicoletta ; Milanović, Željka ; Mauša, Goran

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
Proceedings of International Conference on Innovative Technologies (IN-TECH 2015) / - , 2015

Skup
International Conference on Innovative Technologies (IN-TECH 2015)

Mjesto i datum
Dubrovnik, Hrvatska, 09.09.2015. - 11.09.2015

Vrsta sudjelovanja
Poster

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
Hoshen-Kopelman; K-means; Time-Frequency; Spectrogram; noise; nonstationary signal; multicomponent

Sažetak
In this paper two methods for blind source separation and extraction of nonstationary signals are compared. Based on the requirements of the real life applications we have selected four performance criteria ; namely the algorithm execution time, available resource usage, entirety of extracted components, and the possibility of real time implementation. We have considered one time-frequency based peak detection and extraction algorithm, and a connected componentsalgorithm for cluster labeling, adapted for time-frequency analysis. The sensitivity of first proposed algorithm is highly dependent of the definition of the initial denoising threshold and its execution time is significantly larger than the second proposed algorithm. However, this algorithm allows the implementation in real time. Algorithm for cluster labeling used in this paper isa Bounded Knapsack Problem(BKP) based connected components algorithm, applied after the K-means based denoising method. Thisalgorithmis not suitable for implementation in real time, but allows complete component extraction, uses minimal computer resources and has lower execution time than the time-frequency based method. Six times shorter execution time is obtained for the connected components algorithm than time-frequency based one. All proposed algorithms are tested on simulated noiseless and noisy nonstationary signals embedded in Additive White Gaussian noise (AWGN) for different Signal-to-Noise ratios (SNR). It is also shown that K-means based algorithm performs adaptive denoising on generated spectrograms which optimally removes noise when compared to the time-frequency based method.All simulations are performed on a 3.2GHz quad core computer.

Izvorni jezik
Engleski

Znanstvena područja
Elektrotehnika, Računarstvo

POVEZANOST RADA