Addressing Data Sparsity with GANs for Multi-fault Diagnosing in Emerging Cellular Networks

Abstract

Data-driven machine learning is considered a means to address the paramount challenge of timely fault diagnosis in modern and futuristic ultra-dense and highly complex mobile networks. Whereas diagnosing multiple faults in the network at the same time remains an open challenge. In this context, the data sparsity is hindering the potential of machine learning to address such issues. In this work, we have proposed a data augmentation scheme comprising Pix2Pix Generative Adversarial Network (GAN) and a customized loss function never used before, to address the data sparsity challenge in Minimization of Drive Tests (MDT) data. Our proposed unique augmentation scheme generates images of MDT coverage maps with Peak signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) values of 25 and 0.97 respectively, which are significantly higher than those achieved without our customized loss function. The performance of data augmentation scheme used is further evaluated with a Convolutional Neural Network (CNN) model for simultaneously detecting most commonly occurring network faults, such as antenna up-tilt, antenna down-tilt, transmission power degradation, and cell outage. The CNN applied on the data generated from the 1% of the MDT data with the proposed augmentation scheme has lead to a gain of 550% in the detection of all classes, including the four faults and cell with normal behavior, as compared to when it is applied on the data generated without our customized loss function.