Abstract : This document presents research work done with the ETSN department in Supélec, Rennes campus, about reduction of the non-linear effects in an OFDM modulation using neural networks.
First the documents starts with an introduction to digital communications, and particularly to the OFDM modulation. Today several norms use this technique, mainly because the channel equalization is simple to make, and as such data can be sent more efficiently on multipath channels. However the temporal OFDM signal is sensitive to the power amplifier non-linearities and several techniques have been studied to limit these effects.
Then the neural networks are presented, with their use in function approximation. After having described the two most usual neural network architectures, the higher-order networks, such as the RPN, are introduced. The learning techniques of these networks are also described.
In the different compensators studied in this document, the neural network is put in the receiver, after the channel equalization. Its goal is to correct the received signal to compensate the non-linear effects. First the neural network is in the frequency domain. In a 4 carrier OFDM system with a QAM-16 modulation, an SSPA amplifier, a 0dB back-off and for a bit error rate of 10-2, the neural corrector brings a 1.5dB gain on the signal to noise ratio. But difficulties appear with a higher number of carriers.
For this reason we simplified the neural networks by putting it in the temporal domain. This technique is closer to other solutions already proposed in monocarrier systems, but a few differences are still present in the channel equalization and the type of function that the neural network must realize. A compensator with a RPN network has shown good performance, even with a higher number of carriers. We mesured a 8dB gain in a 48 carrier OFDM system with a QAM-16 modulation, an SSPA amplifier and a 0dB back-off. In these conditions, the systems allows to divide the amplifier power, and therefore the power consumption, by 4, while still having the same transmission quality.
The temporal RPN compensator is then simulated in a multipath channel, to show that the compensation is still efficient in a severe channel. A comparision with another comensator is also presented. Our compensator is less complex, to the detriment of the performance.
This document ends with prospects than can extend the work done in this thesis.