A carregar...
Publicação
Deep learning for reliable communication optimization on autonomous vehicles
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 1.58 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Reliable network coverage prediction is essential to ensure continuous communication in autonomous vehicles, particularly in dense urban environments where buildings and other structures can significantly degrade signal quality. This dissertation proposes a hybrid deep learning architecture designed to accurately predict the Reference Signal Received Power (RSRP) in specific urban locations. The developed model integrates three types of information: semantically encoded urban images derived from OpenStreetMap (OSM), tabular network metrics such as RSRQ (Reference Signal Received Quality), SNR (Signal-to-Noise Ratio), CQI (Channel Quality Indicator) and serving cell distance, as well as a physics-based path loss estimation based on the 3GPP TR 38.901 standard. Each data modality is processed by a specialised neural network, CNNs for spatial features and DNNs for tabular data before being fused into a unified regression model. The approach was evaluated using real 4G LTE data collected in Cork, Ireland, following a 10-fold cross-validation scheme with strict modality alignment. Results show an average Root Mean Square Error (RMSE) of 2.77 dB with a standard deviation of ± 0.07 dB, outperforming purely physical models (e.g., 14.13 dB RMSE) and unimodal deep learning baselines (e.g., 3.81 dB with tabular-only inputs). Additional tests demonstrated spatial generalisation (RMSE ranging from 2.73 dB to 3.10 dB across regions), robustness to noise (e.g., 3.21 dB under degraded RSSI), and high sensitivity to relevant features (e.g., performance drop to 2.85 dB without CQI). The overall performance reached 2.87 dB RMSE, 1.98 dB MAE, and R 2 = 0.93. By fusing physical modelling, geospatial context, and radio metrics, this hybrid approach proves effective and robust, offering a promising tool for autonomous vehicle applications and urban cellular network planning.
Descrição
Tese de Mestrado, Engenharia Informática, 2025, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Deep Learning Autonomous Vehicles RSRP Prediction OpenStreetMap (OSM) Hybrid Learning