Anomaly Detection in High Mobility MDT Traces Through Self-Supervised Learning

Radio access network optimization is a critical task in current cellular systems. For this purpose, Minimization of Drive Test (MDT) functionality provides mobile operators with georeferenced network performance statistics to tune radio propagation models in re-planning tools. However, some samples in MDT traces contain critical location errors due to the user equipment's energy-saving, thus making MDT data filtering vital to guarantee an adequate performance of MDT-driven algorithms. Supervised Learning (SL) allows to train automatic systems for detecting abnormal MDT measurements by using a labeled dataset. Unfortunately, labeling MDT data is a labor-intensive task, that can be alleviated by using Self-Supervised Learning (SSL). This work presents a novel SSL method to detect MDT measurements with abnormal position information in road scenarios. For this purpose, a dataset is first labeled by combining unlabeled MDT traces from high-mobility users and freely available land use maps, and then an SL classifier is trained. Model assessment is carried out using MDT data collected in a live Long-Term Evolution (LTE) network. Performance analysis includes the comparison of six well-known SL algorithms and 3 different sets of input features aiming to improve model accuracy, generalizability, and explainability, respectively. Results show that considering predictors regarding positioning error increases model accuracy, whereas omitting this information allows to cover a wider range of terminals. Likewise, Shapley Additive exPlanations (SHAP) analysis proves that the use of high-level predictors significantly improves model explainability.