Identifying Customer Usage Patterns and System Thermal Behavior in 5G Baseband Hardware : An Evaluation of Feature-Based and Model-Based Time Series Clustering

Sammanfattning

The increasing complexity of 5G mobile networks and their hardware infrastructure creates a growing need to understand real-world usage patterns in network components. This thesis explores how time series clustering can be used to identify patterns and behaviors in thermal and power measurements from Ericsson baseband products, which are core components of 5G networks. Two clustering strategies are investigated: model-based clustering using Hidden Markov Models, and feature-based clustering using extracted statistical features. The study uses one year of multivariate time series data from 212 basebands, including temperature, power, and fan speed measurements. Each sequence was modeled individually with a Hidden Markov Model for the model-based approach, and transformed into a vector of statistical features for the feature-based approach. Clustering was performed using K-medoids with appropriate distance metrics, and the resulting clusters were evaluated using performance scores, internal cluster analysis, and interpretation supported by domain expertise. The results show that model-based clustering with a Hidden Markov Model yields more clear and interpretable clusters than the feature-based approach. The clusters primarily reflect patterns in the overall thermal and power load of the basebands, as well as abnormal system behaviors. The study also highlights challenges such as high dimensionality and limited data when using feature-based clustering, emphasizing that careful feature engineering and selection is a crucial and non-trivial step that impacts clustering results. These findings suggest that time series clustering, particularly model-based methods, can be a valuable tool for analyzing behavior in telecommunication hardware such as basebands, with potential applications in anomaly detection and performance monitoring.