Anomaly detection of High-Mobility MDT Traces Through Self-Supervised Learning.

Abstract

Radio access network optimization is one of the most critical tasks in cellular systems. For this purpose, Minimization of Drive Test (MDT) functionality provides mobile operators with geolocated network performance statistics to tune radio propagation models in replanning tools. However, MDT traces contain critical location errors due to energy-saving modes, which require filtering out wrong samples to guarantee an adequate performance of MDT-driven algorithms. The design of such a classifier detecting valid measurements can be automated by training a supervised learning model with a labeled dataset. Unfortunately, labeling MDT data is a labor-intensive process. In this context, self-supervised learning (SSL) arises as a promising solution to automate labeling of MDT measurements compared to rulebased solutions. This work presents a novel SSL method to filter MDT measurements in road scenarios by combining user mobility traces constructed with unlabeled MDT data and freely available land-use maps. Once labeled, measurements are used to train a supervised learning model. To this end, a proper set of handcrafted features is first derived. Model assessment is carried out over real MDT data collected in a live Long-Term Evolution (LTE) network. Performance analysis includes well-known supervised models, such as Support-Vector Machine, Random Forest, k-Nearest Neighbors and Multi-Layer Perceptron. Results show that all models perform better in MDT measurements including positioning accuracy information. Nevertheless, it is shown that models without this feature can still be used obtaining reliable results and more generalizable models.