Machine Learning Based Dynamic TTI Bundling for LTE VoIP Uplink Coverage Improvement

To decrease the end-to-end delays, Long-Term Evolution (LTE) radio access networks have considerably smaller Transmission Time Intervals (TTI) compared to WCDMA networks. In order to increase the uplink cell coverage over Voice Over LTE (VoLTE) calls for cell-edge User Equipments (UE), TTI Bundling (TTIB) was introduced in release 8 of the 3rd Generation Partnership Project (3GPP). It is based on the principle that the same packet with different redundancy versions (RVs) is transmitted in consecutive TTIs increasing the probability of successful packet decoding. The bundle size in TTI Bundling is always fixed to 4 without considering the UEs' dynamic channel conditions. For UEs that face better channel conditions, the Evolved Node B (eNB) might be able to decode the data packet with fewer RVs than all 4. UEs present in the cell edge need to transmit with a higher power compared to cell near UEs, and adding multiple redundancy version transmission, without considering the channel conditions of the UE, will not only add to the power requirement of the UE but also impact the number of Resource blocks (RBs) available to non-TTI Bundling UEs. In this work, we propose a Machine learning based approach for computing the optimal bundle size of each UE based on its current channel condition. We implement two sequential models, where the first model predicts the channel conditions of the UE for the next transmission. This prediction is then fed to the second model which classifies the prediction into a bundle size ranging from 1 to 4. Using this predicted bundle size for the TTI Bundling UE, our results show a $\mathbf{2 6. 6 7 \%}$ reduction in UE's power consumption, and a $\mathbf{2 9. 4 4 \%}$ improvement in the system throughput.