Performance analysis of network coding based on deep learning and cooperative relay system in wireless communication

Abstract

In this thesis, we aim to improve the performance of network coding based on deep learning and cooperative relay system in wireless communication. Two network coding techniques are considered. First, a random linear network coding linearly combines input signals at nodes in a network with a randomly generated coding coefficient to encode packet as new output information. In this case, in the conventional random linear network coding transmission scheme, each node includes a generated coding coefficient and an encoded packet in one frame and transmits them together. However, since the coding coefficients must be transmitted together, there is a problem that a high transmission overhead is generated and the network load increases. Second, a physical-layer network coding is an effective network coding technique that exploits an interference of the mixed signals that occurs naturally when electromagnetic waves arrived simultaneously. However, an underlying core problem is the performance degradation of demodulation for high-order modulation. Therefore, a separated random linear network coding based on the cooperative relay system was considered, and a deep neural network, which is the core of artificial intelligence technology, was applied to physical-layer network coding. In the first part, we suggest separated random linear network coding based on cooperative medium access control. The proposed scheme can reduce the retransmission overhead by separately transmitting the coding coefficients and the encoded packet. To confirm this, the average retransmission bit and retransmission rate are analyzed. We can notice that the retransmission rate performance can be obtained up to 3dB compared to conventional schemes. In addition, the proposed scheme has disadvantageous in terms of network throughput because it additionally uses cooperative control signals for separate transmission of coding coefficients and encoded packets. However, the cooperative control signal has a very small length compared to payload, so that the throughput degradation is very small. In the second part, we analyze the performance of separated random linear network coding with a outdated channel model. In the practical relay environment, when separate random linear network coding is used, a time delay generates outdated channel state information due to the separate transmission of coding coefficients and encoded packets. Therefore, the outage performance and cooperative diversity of the proposed scheme are analyzed in consideration of the correlation between the coding coefficient transmission channel and the encoded packet transmission channel. As a result of analyzing the performance of the proposed scheme in the outdated channel model, it can be seen that the difference of outage probability performance according to the correlation coefficient is about 1dB. In addition, despite the outdated channel state information, it was confirmed that the full cooperative diversity gain can be obtained. Finally, a physical-layer network coding based on deep neural network is proposed. A convolutional autoecoder for physical-layer network coding was designed. The constellation mapping and demapping of symbols at each node is determined adaptively through the deep learning technique, such that the bit error rate performance is improved for high-order modulation. Simulation results verify the advantages of the proposed scheme over the conventional physical-layer network coding for various modulation types.