Restricted Research - Award List, Note/Discussion Page
Fiscal Year: 2023
54 University of North Texas (141942)
Principal Investigator: Huang,Yan
Total Amount of Contract, Award, or Gift (Annual before 2011): $ 3,763,760
Exceeds $250,000 (Is it flagged?): Yes
Start and End Dates: 1/1/22 - 3/31/23
Restricted Research: YES
Academic Discipline: Computer Science & Engineering
Department, Center, School, or Institute: College of Engineering
Title of Contract, Award, or Gift: Cognitive Distributed Sensing in Congested Radio Frequency Environments
Name of Granting or Contracting Agency/Entity:
Northeastern University
CFDA Link: DOD
12.431
Program Title:
none
CFDA Linked: Basic Scientific Research
Note:
The objective of this research is to design a power-efficient mmWave hybrid beamforming system with phased array antenna, and to develop deep learning-based beam alignment methods for swarm coordination and distributed sensing. In this project, we will undertake an integrative co-design approach through coherent integration of expertise in phased-array antennas, hybrid beamforming, communications, software defined radio (SDR), and deep learning. Hybrid beamforming is a promising solution to combine the technical advantages of digital and analog beamforming architectures while keeping the cost manageable for many practical applications. In the analog domain, a phased array relies on variable phase shifters connected to array elements to control the antenna’s radiation pattern. In general, the more reconfigurable phase states each array element has, the higher beamforming performance the array exhibits, albeit at the expense of system complexity, implementation cost, and power consumption. It was reported that using phase quantization helps reduce power consumption of a hybrid beamforming system compared to using continuous phase shifts 1. Our past work has also shown that using a 2-bit quantization for the variable phase states of phased-array elements does not sacrifice much in term of beamforming performance while significantly reducing the complexity and cost of the array 2. The main technical challenge is how to design 2-bit phase-shifting array elements having wideband operation and using low-loss switching mechanism with minimum number of electronic switches. In the digital domain, fully digital beamforming is typically associated with high cost, high complexity, and large energy consumption because each antenna element is backed by an independent RF chain consisting of a full-fledged transceiver. With hybrid beamforming structure, it is possible to achieve simultaneous multi-beam multi-channel communication with a small number of digital transceiver chains. The main design challenge is in providing sufficient power and computing efficiency to enable integration of such design in small formfactor mobile platforms such as mobile user devices, swarm robotics, UAVs, etc. In this research, we will use low-power software defined radio platform to design and implement hybrid beamforming system 3, 4. We will perform the evaluation experiments at KRI ECUAS Lab in collaboration with Northeastern University research team. In dynamic swarm environment, accurate beam alignment between two moving peers is critical and challenging as the angles of the two moving agents must be tracked reliably and continuously. Due to the high computational complexity, the traditional approaches such as Kalman filter and particle filtering do not scale well in a highly mobile environment. Recent research has attempted to use recurrent neural network with Long-short Term Memory (LSTM) to learn to predict the angle of departure between a fixed base station and a mobile user 5. However, in a high mobility swarm environment, both agents move in a space with elevation and obstacles, introducing additional spatial and temporal diversity (6, 7). We propose to capture this diversity with attention based deep learning models. Recent advancement in generative pre-training models such as OpenAI GPT3 has achieved high performance in language understanding. And those models have shown success in many other domains. The core of these models is to use a self-supervised architecture and attentions for prediction.
Discussion: No discussion notes