Análisis de la aplicación de auto codificadores en sistemas de comunicaciones para minimizar el efecto del ruido

Abstract

Conventional communication systems use processing blocks to perform specific tasks, such as encoding or modulating input signals. Recently, Deep Learning (DL) theory has inspired programmers and researchers in the area of digital communications to apply it to these traditional systems with the aim of making them adaptable to different environments, reducing the complexity in the design of blocks that perform specific tasks. In this context, the main objective of this work is to develop an Autoencoder (AE), which is the application of an unsupervised learning artificial neural network, with an encoding-decoding structure similar to that of a conventional communication system. The model presented in Chapters 4 and 5 is developed using the TensorFlow libraries and functional API. The implemented encoding was based on the following modulation schemes: BPSK, QPSK, 8-PSK, 16-PSK, 32-PSK, 4-QAM, 16-QAM, 32-QAM, 64-QAM, and 128-QAM. Through training, the model aims to efficiently reduce the error between the original information and the reconstructed information after passing through a noisy AWGN channel. To optimize the training, the early stopping technique is incorporated, which halts the training when the loss function values stop improving. To verify the functionality of the AE, constellation diagrams, Block Error Rate (BLER) and Bit Error Rate (BER) plots were obtained for each modulation scheme. These graphs were analy- zed and compared with the typical curves of the aforementioned digital modulations. The results of the autoencoder indicate greater precision and efficiency in reconstructing the transmitted data, consequently leading to a significant improvement in information transmission.