Denoising Diffusion Model for DOA Estimation

Deep learning methods for robust Direction-of-Arrival (DOA) estimation are often limited by their scenario-specific training. They struggle to generalize to unseen scenarios like a variable number of sources, necessitating huge datasets and training separate networks. In this paper, we first derive the source-number decoupling property of signals, then propose a denoising diffusion model that reformulates the DOA estimation problem as a signal denoising task. The proposed model is trained on signals with a fixed number of sources, yet can generalize to scenarios with a variable number of sources. Simulation results demonstrate that the proposed method achieves superior estimation performance and robustness in scenarios with a variable number of sources, low SNRs and few snapshots under the same data scale.