A Low-complexity Structured Neural Network Approach to Intelligently Realize Wideband Multi-beam Beamformers
Document Type
Presentation
Location
COAS: Math Conference Room
Start Date
18-4-2025 10:30 AM
End Date
18-4-2025 10:55 AM
Description
In wireless communication, true-time-delay (TTD) wideband multi-beam beamformers could be utilized to realize squint-free FFT-based beamformers via delay Vandermonde matrix (DVM)-based beams. Building upon the DVM-based approaches, we present a structured neural network (StNN) to intelligently realize TTD wideband multi-beam beamformers. The StNN architecture is designed based on structured and sparse weight matrices and sub-weight matrices, reducing the computational complexity in conventional neural networks from O(M2 L) to O(pLM log M), where M represents the number of nodes per layer, p < < M denotes the number of submatrices within each layer, and L is the total number of layers.
We present numerical simulations of beamformed signals in the 24 GHz to 32 GHz frequency range to validate the StNN model's feasibility. The simulations show that the StNN lowers complexity while remaining consistent with theoretical results and achieving accuracy comparable to fully connected neural networks, making it suitable for low-complexity intelligent systems to realize TTD wideband multi-beam beamformers.
This is a joint work with Xianqi Li, Arjuna Madanayake, and Sirani M. Perera
A Low-complexity Structured Neural Network Approach to Intelligently Realize Wideband Multi-beam Beamformers
COAS: Math Conference Room
In wireless communication, true-time-delay (TTD) wideband multi-beam beamformers could be utilized to realize squint-free FFT-based beamformers via delay Vandermonde matrix (DVM)-based beams. Building upon the DVM-based approaches, we present a structured neural network (StNN) to intelligently realize TTD wideband multi-beam beamformers. The StNN architecture is designed based on structured and sparse weight matrices and sub-weight matrices, reducing the computational complexity in conventional neural networks from O(M2 L) to O(pLM log M), where M represents the number of nodes per layer, p < < M denotes the number of submatrices within each layer, and L is the total number of layers.
We present numerical simulations of beamformed signals in the 24 GHz to 32 GHz frequency range to validate the StNN model's feasibility. The simulations show that the StNN lowers complexity while remaining consistent with theoretical results and achieving accuracy comparable to fully connected neural networks, making it suitable for low-complexity intelligent systems to realize TTD wideband multi-beam beamformers.
This is a joint work with Xianqi Li, Arjuna Madanayake, and Sirani M. Perera