Deep Learning-based Microwave Network Fault Prediction Model and Analysis

This research focuses on creating deep learning-based fault occurrence time prediction models using Gated Recurrent Unit (GRU) and Long Short-Term Memory (LSTM) algorithms to predict fault occurrences and prevent system failures for microwave communication networks. Identification of cost-cutting opportunities and assigning maintenance teams to diagnose problems ensures stable telecom services and efficient fault alarm data collection for Ethiotelecom. The trained the models using two data-split strategies 70/30 and 80/20 to assess their consistency and compare their performance. The Universal Network Management System (UNMS) provided five months’ worth of failure history data for this purpose. We used the following common error measures to assess prediction accuracy: Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). According to the findings, the Gated Recurrent Unit (GRU) model was not as effective as the Long Short-Term Memory (LSTM) model. Our findings confirm that both deep learning models effectively predict telecom alarm events. GRU shows $\mathbf{0. 2 5 9}$ MAE on average and LSTM model has predicted the fault occurrence time with $\mathbf{0. 2 5 8 9 4 5}$ MAE.