Memory-Saving Gridless Direction-of-Arrival Estimation Based on Distributed Deep Learning

As the scale of antenna array increases, centralized methods based on deep learning (DL) encounter significant memory constraints on devices. To address this issue, we propose a distributed DL-based framework for gridless direction-of-arrival (DOA) estimation, which substantially alleviates memory constraints for devices operating in large-scale antenna scenarios. We employ an overlapped subarray selection strategy that partitions the complete array into multiple subarrays, allowing for partial array elements overlapped between adjacent subarrays. This strategy effectively compensates for the loss of cross-correlation information between subarrays, thereby enhancing the precision of DOA estimation. Within this framework, each subarray is paired with an independent subprocessor responsible for compressing the received data and transmitting the results to a fusion center. The fusion center leverages a graph neural network (GNN) to effectively extract DOA estimation information from complex datasets. This framework treats DOA estimation as a regression task, leveraging Toeplitz prior to achieve high-precision gridless DOA estimation through postprocessing. Additionally, we introduce a hybrid data-driven and model-based framework that significantly reduces computation time while ensuring the accuracy of DOA estimation, making it particularly suitable for real-time applications. Simulation results demonstrate that our proposed distributed DL methods achieve DOA estimation accuracy comparable to that of centralized DL methods, while exhibiting lower time complexity and reduced memory requirements for devices.