Multilayer Radial Basis Function Neural Network for Symbol Timing Recovery

In digital communication, synchronization between transmitter and receiver is essential for ensuring proper system performance. Error in the receiver symbol time sampling can significantly increase the bit error rate to unacceptable levels. In this paper, we propose a multilayer radial basis function neural network symbol-timing recovery (MRBFNN-STR). The proposed solution has been implemented for a 64-QAM (quadrature and amplitude modulation) system. Results show that the MRBFNN-STR improves the modulation error ratio up to 3.4 dB and reduces the bit error rate by almost one order of magnitude for 100 ppm (part per million) clock offset and signal to noise ratios above 25 dB compared to the classic widely used Gardner-Farrow’s approach. The MRBFNN is able to follow the system dynamics and to generalize, presenting good performance even when under operational situations not presented during the training phase (different clock offset, signal to noise ratio, etc.) and with lower-order modulation schemes, such as 32-QAM, 16-QAM, and QPSK (quadrature phase shift keying), without retraining. Due to the parallel nature of the MRBFNN architecture and the reduced complexity required for inference, it can be efficiently implemented in hardware and easily integrated into communication receivers, representing a feasible solution for receiver time synchronization.