Detecção de decisões precipitadas de Handover em Redes LTE utilizando Redes Neurais Recorrentes

Abstract

The handover decision algorithms based on the Received Reference Signal Power (RSRP) are sensitive to measurement errors and oscillation of the user's geographic position, causing signaling overload and interruption of services provided by the mobile network. However, the use of intelligent algorithms based on Recurrent Neural Networks (RNN) offer a solution to mitigate handover failures. This work validated the implementation of two RNN models, Long Short-Term Memory (LSTM) and Gate Recurrent Unit (GRU), for detecting hasty handovers decisions in the Long Term Evolution (LTE) network. The implemented neural network models estimate the probability of unnecessary handover decision, using as input the LTE Key Performance Indicators (KPI) and mobile device location. The results of the experiments show that the method used for classification did not obtain optimal results for detection of handover classification, obtaining precision of 96.12% and accuracy of 78% using the GRU recurrent layer. However, the metrics obtained are adequate within what is expected from a classification scenario using RNN networks in an unbalanced database.