Assessment of energy-efficient wireless network using autoencoders with unsupervised deep learning

Abstract

The propagation of wireless networks in e-business applications demands efficient and robust anomaly detection techniques to ensure data security and reliable communication. A conventional autoencoder is learned to efficiently compress data as an unsupervised neural network within an encoding process. The autoencoder can learn to rebuild the information from the compact model so that the dissimilarity of the reconstructed to the original data is on least. Conventional wireless communication techniques are developed to deliver consistent data transmit through an impaired channel of the transferred signals. This work presents an autoencoder model for an end-to-end information bits communication system over a wireless channel with reliable transmission. The Autoencoder for energy-efficient wireless networks is defined with a pair of (channel uses number, and input bits number). For both the transmitter (encoder) and receiver (decoder), the developed network architecture includes only two fully connected layers. The transmitter includes two fully connected layers and the input layer allows a vector (one-hot) with M length. Several normalized autoencoders are compared for the learned constellations to unit average power and unit energy. The AWGN channel layer is connected after the transmitter (encoder) layers. The training progress result demonstrated that the validation loss remains slowly declining, while the validation accuracy rapidly achieves a value larger than 92%. The block error rate (BLER) outcome demonstrated that the autoencoders could be trained and learned as a modulation scheme and joint coding in an unsupervised method.