Nonlinear techniques in telecommunications

The identification and compensation of unwanted nonlinearities are required in many telecommunications systems in order to improve the system performances.

1. Volterra models Nonlinearities representation is very important because the compensation technique depends on the model used. The Volterra filter is widely used to model unwanted nonlinearities and provides also compensations techniques.
The major drawback of the Volterra technique is due to the computational complexity required to implement the model and the compensation technique.
On the other hand, an efficient compensator requires accurate nonlinear system identification based on complex Volterra estimators.
In practice, because the nonlinearities order is unknown, adaptive algorithms are suitable to estimate Volterra kernels and to construct the nonlinear model.
Fast kernel estimation techniques are investigated in a large number of papers, in order to construct higher order models that give accurate representations.
2. Neural networks equalizers The multiple quadrature amplitude modulated (M-QAM) signals, more efficient in transmission from the spectral point of view, have known an expanding research interest. The M-QAM signals are severely affected by the nonlinear distortions, because they have a variable envelope modulation. To compensate these unwanted distortions neural networks equalizers for complex signals have been developed which are straightforward extensions from the real counterparts, obtained by replacing the relevant parameters with complex values. Recently, the RBF-NN have received considerable attention, since the MLP network is plagued by long training times and may be trapped in bad local minima. The RBF-NN is able to approximate any arbitrary nonlinear function in the complex multi-dimensional space with a reduced calculus complexity comparative to other NN. The RBF-NN often provides a faster and more robust solution to the equalization problem than the MLP . In addition, the RBF-NN equalizer has a structure similar to the optimal Bayesian symbol decision equalizer, so its performance is better than the MLP equalizer performance.

Decision regions of an RBF-NN complex equalizer in a complex data space, which have strong nonlinear decisions boundaries


Selected Publications:


Intelligent Signal Processing Centre
Copyright © 2010