Doctor of Philosophy
Electrical and Computer Engineering
Ultra-dense heterogeneous network (HetNet), in which densified small cells overlaying the conventional macro-cells, is a promising technique for the fifth-generation (5G) mobile network. The dense and multi-tier network architecture is able to support the extensive data traffic and diverse quality of service (QoS) but meanwhile arises several challenges especially on the interference coordination and resource management. In this thesis, three novel network schemes are proposed to achieve intelligent and efficient operation based on the deep learning-enabled network awareness. Both optimization and deep learning methods are developed to achieve intelligent and efficient resource allocation in these proposed network schemes.
To improve the cost and energy efficiency of ultra-dense HetNets, a hotspot prediction based virtual small cell (VSC) network is proposed. A VSC is formed only when the traffic volume and user density are extremely high. We leverage the feature extraction capabilities of deep learning techniques and exploit a long-short term memory (LSTM) neural network to predict potential hotspots and form VSC. Large-scale antenna array enabled hybrid beamforming is also adaptively adjusted for highly directional transmission to cover these VSCs. Within each VSC, one user equipment (UE) is selected as a cell head (CH), which collects the intra-cell traffic using the unlicensed band and relays the aggregated traffic to the macro-cell base station (MBS) in the licensed band. The inter-cell interference can thus be reduced, and the spectrum efficiency can be improved. Numerical results show that proposed VSCs can reduce $55\%$ power consumption in comparison with traditional small cells.
In addition to the smart VSCs deployment, a novel multi-dimensional intelligent multiple access (MD-IMA) scheme is also proposed to achieve stringent and diverse QoS of emerging 5G applications with disparate resource constraints. Multiple access (MA) schemes in multi-dimensional resources are adaptively scheduled to accommodate dynamic QoS requirements and network states. The MD-IMA learns the integrated-quality-of-system-experience (I-QoSE) by monitoring and predicting QoS through the LSTM neural network. The resource allocation in the MD-IMA scheme is formulated as an optimization problem to maximize the I-QoSE as well as minimize the non-orthogonality (NO) in view of implementation constraints. In order to solve this problem, both model-based optimization algorithms and model-free deep reinforcement learning (DRL) approaches are utilized. Simulation results demonstrate that the achievable I-QoSE gain of MD-IMA over traditional MA is $15\%$ - $18\%$.
In the final part of the thesis, a Software-Defined Networking (SDN) enabled 5G-vehicle ad hoc networks (VANET) is designed to support the growing vehicle-generated data traffic. In this integrated architecture, to reduce the signaling overhead, vehicles are clustered under the coordination of SDN and one vehicle in each cluster is selected as a gateway to aggregate intra-cluster traffic. To ensure the capacity of the trunk-link between the gateway and macro base station, a Non-orthogonal Multiplexed Modulation (NOMM) scheme is proposed to split aggregated data stream into multi-layers and use sparse spreading code to partially superpose the modulated symbols on several resource blocks. The simulation results show that the energy efficiency performance of proposed NOMM is around 1.5-2 times than that of the typical orthogonal transmission scheme.
Summary for Lay Audience
5G network becomes ultra-densified and heterogeneous to support the explosive data traffic and diverse QoS requirements. However, the densified and multi-tier architectures bring new challenges, especially on the interference coordination and resource management. In order to address these challenges in 5G ultra-dense HetNets, three novel network schemes are proposed to achieve intelligent and efficient operation based on the deep learning-enabled network awareness and multi-dimensional resource allocation.
First of all, to improve the energy efficiency of ultra-dense HetNets, a hotspot prediction based virtual small cell (VSC) operation scheme is proposed. The VSCs are formed only when the areas are hotspots, which can be predicted through deep learning technology. Then, a large-scale antenna array enabled highly directional beamforming scheme is adaptively designed to cover these VSCs. Within each VSC, one user equipment is selected as a cell head to collect the intra-cell information and relays the aggregated traffic to the macro-cell base station.
Moreover, in order to fully utilize the resource available in 5G ultra-dense networks, a novel multi-dimensional intelligent multiple access (MD-IMA) technique is developed to adaptively select the multiple access (MA) schemes. The proposed MD-IMA technique learns the overall system requirements and then adaptively multiplexes co-existing devices in multi-dimensional resources to meet the real-time QoS requirements. The resource allocation problem of the MD-IMA system is formulated as an optimization problem to maximize the overall network performance as well as to minimize receiver complexity. Both model-based optimization algorithms and model-free deep reinforcement learning-enabled approaches are proposed to solve this optimization problem.
Finally, to support growing vehicle-generated data traffic in 5G-vehicle ad hoc networks (VANET), a Software-Defined Networking (SDN) enabled 5G-VANET is presented. In this integrated architecture, SDN can provide a global view to adaptively cluster vehicles only when needed. In order to reduce the signaling overhead, one vehicle in each cluster is selected as a cell head to support the aggregated traffic. To ensure the capacity of the trunk link between the gateway and base station, a new modulation scheme is also proposed to effectively aggregate the trunk link traffic.
Liu, Yanan, "Intelligent and Efficient Ultra-Dense Heterogeneous Networks for 5G and Beyond" (2020). Electronic Thesis and Dissertation Repository. 6959.
Available for download on Saturday, October 30, 2021