Synthetic Channel Generation using Generative Adversarial Networks

Bangshaab, Mads

4. term

Signal Processing and Computing, Master

2020

2020-06-03

54 pages

In this report Generative Adversarial Networks (GAN) are designed and implemented with the purpose of generating synthetic responses of a radio channel. An investigation of relevant theory of neural networks and specifically GANs is conducted and based on this, choices with regards to architectures are made. For the purpose of training and evaluating the GANs, synthetic channel response data is generated based on a stochastic model. From this data relevant statistical properties are computed as to form a basis for evaluation. Two different realisations of a GAN are then designed based on the theory reviewed and relevant design choices made regarding the training data. These two realisations are based on a FCNN and a CNN respectively. Through testing it is seen, that the DCGAN has the best performance. The DCGAN is then evaluated with regards to the robustness of the network with regards to noise added to the training data and the size of the data set. Based on these findings, the model used for the DCGAN is then trained anew with another data set containing real measurements, where both training data is limited and noise are included in the measurements. Through testing of the GAN it is found, that though not achieving perfect results, the use of GANs for stochastic radio channel response generation is feasible.

In this report Generative Adversarial Networks (GAN) are designed and implemented with the purpose of generating synthetic responses of a radio channel. An investigation of relevant theory of neural networks and specifically GANs is conducted and based on this, choices with regards to architectures are made. For the purpose of training and evaluating the GANs, synthetic channel response data is generated based on a stochastic model. From this data relevant statistical properties are computed as to form a basis for evaluation. Two different realisations of a GAN are then designed based on the theory reviewed and relevant design choices made regarding the training data. These two realisations are based on a FCNN and a CNN respectively. Through testing it is seen, that the DCGAN has the best performance. The DCGAN is then evaluated with regards to the robustness of the network with regards to noise added to the training data and the size of the data set. Based on these findings, the model used for the DCGAN is then trained anew with another data set containing real measurements, where both training data is limited and noise are included in the measurements. Through testing of the GAN it is found, that though not achieving perfect results, the use of GANs for stochastic radio channel response generation is feasible.

Deep learning ; GAN ; Machine Learning ; Supervised Learning

Download
View record in AAU Student Projects