A Novel Transform Domain Based Hybrid Recurrent Neural Equaliser for Digital Communication Channel

摘要

Efficient neural network based adaptive equalisations for digital communication channels have been suggested in recent past. Recurrent neural network (RNN) exhibits better performance in nonlinear channel equalization problem. In this present work a hybrid model of recurrent neural equaliser configuration has been proposed where a Discrete Cosine Transform (DCT) block is embedded within the framework of a conventional RNN structure. The heterogeneous configuration on the RNN framework needs training and involves updation of the connection weights using the standard RTRL algorithm, which necessitates the determination of errors at the nodes of the RNN module. To circumvent this difficulty, an adhoc solution has been suggested to back propagate the output error through this heterogeneous configuration. Simulation study and bit-error-rate performance analysis of the proposed Recurrent Transform Cascaded (RTCS) equaliser for standard communication channel models show encouraging results.