Direction of arrival estimation of wideband signals based on Gauss-Cauchy mutation particle filter

This paper proposes a novel broadband signal direction-of-arrival (DOA) estimation algorithm based on a Gauss-Cauchy mutation particle filter (GCPF). To overcome the performance limitations of conventional subspace-based and sparse-recovery methods—such as the incoherent signal-subspace method (ISM), coherent signal-subspace method (CSM), and sparse Bayesian learning (SBL)—in challenging scenarios characterized by few snapshots, low signal-to-noise ratio (SNR), or coherent sources, the proposed approach fundamentally reformulates the DOA estimation problem from a sequential Bayesian tracking perspective. Rather than relying on accurate estimation and eigen-decomposition of the covariance matrix—a process that demands numerous snapshots and is sensitive to errors under data-limited conditions—the algorithm directly processes the frequency-domain observations using a sequential Monte Carlo framework. The core innovation of this work is the introduction of an adaptive Gauss-Cauchy hybrid mutation resampling mechanism, which systematically combats the intrinsic issues of particle degeneracy and sample impoverishment in conventional particle filters. This mechanism intelligently differentiates particle treatment based on their importance weights: Gaussian mutation, with its concentrated probability density, is applied to high-weight particles to refine their positions through local exploitation, whereas heavy-tailed Cauchy mutation is employed for low-weight particles to facilitate global exploration of the state space and prevent convergence to local optima. By dynamically balancing exploration and exploitation, the proposed resampling strategy preserves particle diversity and improves the overall robustness and accuracy of the filter. Furthermore, the method operates without requiring pre-estimated initial angle values or prior knowledge of source characteristics, and it inherently handles coherent signals due to its direct likelihood-based updating rule. Extensive simulation experiments demonstrate that the GCPF algorithm exhibits superior performance in estimation accuracy, angular resolution, and operational robustness. Under extremely challenging conditions using only 100 snapshots and an SNR of 0 dB, its RMSE is reduced by approximately 50 % compared to the ISM method; for coherent sources with an angular spacing of 10 °, the detection probability reaches approximately over 30 %, significantly outperforming traditional ISM, CSM, and SBL methods. Furthermore, the GCPF algorithm achieves higher computational efficiency, with an average runtime of only 0.11 s, which is about 31–56 % faster than the comparative methods.