Detection of Sleeping Cells in Self-Organizing Cellular Networks: An Adversarial Auto-Encoder Method

Automatic fault management is one of important components in self-organizing networks to mitigate and recover from failures of problematic cells. As a special case of cell outage, sleeping cell, which provides degraded service for subscribers without triggering alarms, brings challenges to fault detection. Machine learning provides effective tools for such a detection task. However, traditional algorithms likely constitute biased classifiers when there are imbalance input measurements. Also, misclassification is unavoidable and existing schemes are insensitive to the different costs generated by different misclassifications. To address these issues, we propose a novel method to learn from multi-class imbalance measurements, through combining a modified adversarial autoencoder (AAE) and cost sensitive support vector machine (SVM). Specifically, the proposed method uses AAE to generate more data for minority classes and the loss function of AAE is modified to promote the stability of model training. Then, the cost sensitive SVM is utilized to classify synthetic balanced samples, assigning different costs for varying classification results. Experiment results, evaluated by effective metrics, reveal that the proposed algorithm can improve the detection performance of imbalance sleeping cell data and show superior results compared with state-of-the-art schemes.