Your Location Within Site:
Blind source separation
Blind source separation has advanced to provide powerful algorithms that allow separation of signals without prior knowledge of the nature of the signals, interferers, noise or the mixing process. This capability can be combined with prior knowledge of the signals, such as probability density function, time or frequency sparsity, to separate signals previously considered irretrievable corrupted.
Mixed signals of virtually any sort can be separated without prior knowledge of the signals. Additional knowledge can assist separation.
Recent advances in signal processing and machine learning have seen the emergence of a common framework for several traditionally rather disjoint fields. Notably, methods from adaptive filtering, blind source separation and machine learning can now be unified allowing innovations from any one to provide solutions to other fields and providing wider application areas.
This has provided enhanced capability and performance in:
- filtering to reduce possibly unknown corruption on wanted signals,
- separation of signals, or signals from noise, interference,
- separation of signals into independent component signals representing underlying unobservable influences,
- classification of signals, samples, images, sequences.
Our research focuses on extension of the signal processing theory behind these technologies as well as on methods of application. Emphasis lies in convolutive mixtures, underdetermined systems (where the observable signals are outnumbered by the unobservable) and in the convergence of adaptive algorithms, separation and pattern recognition for estimation, tracking, separation and classification. This work is complemented by research and development of novel multi-channels sensors where the signals are difficult to acquire.
Our current application areas in implementation include reverberant wireless propagation, reverberant sonar signals, separation of the sound of foetal heartbeat from the much louder sounds from the mother and applications to EEG for monitoring the brain activity of at-risk neonates.
Industrial Research Limited has capability to apply these techniques to development of algorithms for many applications. Our expertise for implementation in software, DSP[?] and FPGA[?] provides a capability to transfer these technologies into industrial applications.