Uppskattning av omgivningstemperatur för basbandsenheter med hjälp av maskininlärning

Accurate estimation of ambient temperature supports the efficient and reliable operation of baseband units (BBUs), which perform core digital signal processing tasks in modern mobile networks. In centralized or Cloud-RAN architectures, environmental factors such as heatwaves or irregular airflow patterns may, in extreme cases, impact system performance. If the system mistakenly interprets temperature-induced anomalies as hardware failures, it will trigger false alarms, resulting in unnecessary maintenance. Thus by enabling real-time temperature monitoring, ambient temperature estimation helps reduce false alarms and contributes to long-term system reliability. While traditional physical sensors are commonly used for temperature monitoring, their integration into large-scale and complex products can be challenging due to cost, placement limitations, and scalability issues. To address this, we proposed a machine learning based approach to indirectly estimate BBU ambient temperature using readily available internal data and operational parameters, offering a flexible, scalable, and cost-effective solution over dedicated external sensors. The study leverages supervised learning methods, including Random Forest, XGBoost, Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN), trained on data collected under controlled environmental conditions. A group-aware evaluation framework is adopted to ensure robust generalization across unseen hardware configurations. Results show that the best-performing model is MLP, which achieves a mean absolute error (MAE) of 0.83 °C, a root mean square error (RMSE) of 0.94 °C and an R² of 0.99, confirming the feasibility of data-driven modeling for temperature estimation. According to the results, this method has the potential to improve fault classification, reduce unnecessary maintenance, and inform more effective thermal control strategies, contributing to more sustainable and cost-effective network operations. Future work should aim to improve cross-configuration generalization and explore real-world deployment in field conditions.