A deep learning-based approach for digital beamforming of antenna arrays

Project Title

A deep learning-based approach for digital beamforming of antenna arrays

Project Area of Specialization Information & Communication TechnologyProject Summary

Phased arrays and beamforming are integral components in modern digital communication systems. Currently beamforming based on analog techniques using computationally expensive electromagnetic simulators are in use. The main drawback of these analog techniques is slow processing for real-time applications. This problem can be solved to a large extent by using digital beamforming technique integrated with latest machine learning algorithms.     

In this project we will be developing digital beams using phased arrays and steer it by varying voltages and phases of each individual antennas integrated with machine learning, towards a desired target. We will propose a deep neural network (DNN) to utilize the radiation patterns as inputs and the amplitude and phase of the antenna elements as outputs.

Project Objectives

1. To design a deep learning-based methodology to synthesize the radiation pattern of array antennas
2. To construct a low complexity deep neural network (DNN) architecture for digital beamforming
3. The feasibility of the DNN-based methodology for strongly coupled MIMO antennas implemented inside a mobile phone

Project Implementation Method

We will use a deep neutral network (DNN) approach for the proposed system. We will use Python language for developing this system due to the support of powerful libraries. We will use python machine-learning libraries to build our machine-learning algorithm. We will use the google TensorFlow to develop the DNN for digital beamforming.

Table 1 Tools and Technologies for Proposed Project

Tools And Technologies	Tools	Version	Rationale
	Spyder	4.2.1	IDE
	Jupyter Notebook	6.2.0	IDE
	MS Word	2016	Documentation
	MS PowerPoint	2016	Presentation
	Technology	Version	Rationale
	Python	3.8	Programming language
	Python Libraries	-	Machine learning
	TensorFlow	2.0 or latest	For simulations
	CST	2017
	MATLAB	2015

Tools

And

Technologies

Spyder

Jupyter Notebook

MS Word

MS PowerPoint

Technology

Python

Python Libraries

TensorFlow

Benefits of the Project

• The proposed DNN-based approach for digital beamforming will be of reduced computational complexity from the conventional schemes for large antenna arrays.
•  The proposed DNN will utilize real and imaginary numbers instead of the phase value in order to enhance the ability of learning for an arbitrary antenna radiation pattern.
• The DNN-based methodology can be extended for strongly coupled MIMO antennas implemented inside a mobile phone.

Technical Details of Final Deliverable

Module 1: Extraction of Training Data

A four-element patch antenna array working in ISM band (2-4 GHz) will be designed in commercially available simulator CST (Computer Simulation Technology). To investigate the effect of deep learning, a high coupling between antenna elements in the array will be chosen with the distance between antennas being equal to 0.28 wavelength. This is narrower than 0.5 wavelength, typically used in array antennas. The number of radiation pattern cases per antenna will be 19, so the total number of radiation patterns for training data will be 1 x 19 x 19 x 19 = 6,859.

Module 2: Extraction of Validation Data

To verify whether the deep neural network is properly trained, validation data should not overlap with the training data. The total number of radiation patterns data for training data set will be 1 x 4 x 4 x 4 = 64.  The training and validation data set will be extracted with MATLAB and CST simulators.

Module 3: Neural Network Modeling and Deep Learning

The proposed DNN beamforming architecture will comprise of 5 layers as follows: The input layer consisting of 181 input data (according to observation angle value ranging from 0 to 180 degrees). The output layer will generate desired 4 amplitude and 4 phases for digital beamforming patterns. Three fully-connected hidden layers will connect both input and output neurons. For training proposed beamforming DNN, 6,859 training radiation patterns of Module 1 and 64 validation data of Module 2 will be used. Train and validation losses as a function of epoch will be used as bench marks to test the training of the proposed DNN beamforming network.     

Module 4: Beamforming Validation

To validate the DNN output data, one of the validation sets from Module 2 will be selected and compared with Module 3 output. The DNN output data should be similar to the output data of the validation set. The beamforming pattern (after deep learning) will be obtained through CST simulations by using the DNN output data. The two beamforming radiation patterns will be compared for validation.

Final Deliverable of the Project Software SystemCore Industry TelecommunicationOther IndustriesCore Technology Artificial Intelligence(AI)Other TechnologiesSustainable Development Goals Decent Work and Economic Growth, Industry, Innovation and InfrastructureRequired Resources

Elapsed time in (days or weeks or month or quarter) since start of the project	Milestone	Deliverable
Month 1	Critical Literature Review	Identification of research gaps
Month 2	Study and Implementation of selected beamforming algorithms	MATLAB codes
Month 3	Know-how of machine-learning tools and python platform	Generation of training and validation data sets
Month 4	Know-how of machine-learning tools and python platform	Generation of training and validation data sets
Month 5	Implementation of proposed DNN architecture for digital beamforming	DNN software module
Month 6	Implementation of proposed DNN architecture for digital beamforming	DNN software module
Month 7	Functional trained DNN module for desired beamforming	DNN software module with complex weights calculation capability
Month 8	Project testing and Integration	Project report and user manual