A Low-Complexity LSTM Network to Realize Multibeam Beamforming
Hansaka Aluvihare, Carina Shanahan, Sirani M. Perera, S. Sivasankar, Umesha Kumarasiri, Arjuna Madanayake, Xianqi Li · 2024
Massive data structures can be embedded in the form of weight matrices, enabling to design of neural networks with low-complexity learning algorithms. These data can be organized in rows of matrices, containing progressive phase shifts for a specific beam, along with input and output vectors consisting of time-domain signals. In our previous work, we have identified that multi-beam beamformers based on true-time-delays (TTDs) can be mathematically formulated as the elements of delay Vandermonde matrices (DVM). Thus, by adopting a frequency domain variable, we can express TTDs of time delay data in terms of elements of the DVM. Learning from prior work, we propose to present a low-complexity neural network to realize multibeam beamforming leveraging a novel LSTM network. The goal of our work is to reduce the complexity of the multibeam beamforming algorithm from ${\mathcal{O}}({N^2}L)$ to ${\mathcal{O}}({N^s}L)$, where 1 < s < 2, by imposing factorization of the DVM in an LSTM network having L layers.