Continuous DOA estimation via tensorized LSTM

To effectively utilize the correlation information of the signals that received by sensor arrays, continuous direction-of-arrival(DOA) estimation methods based on Long Short-Term Memory (LSTM) networks have been proposed. Nonetheless, existing LSTM-based DOA estimation methods typically vectorize the received signals as input to the network, resulting in a large number of trainable parameters and a substantial computational burden. To address these challenges, this paper introduces a tensorized LSTM network method for two-dimensional continuous DOA estimation. In this approach, we employ temporal tensor signal modeling for dynamic targets and design the corresponding tensorized LSTM network that reduces the number of network parameters while effectively capturing angle-time domain information. Simulation results demonstrate that the proposed tensorized LSTM network not only significantly reduces the number of network parameters, but also achieves a higher DOA estimation accuracy as well as an improved generalization.