Broadband ML estimation under model order uncertainty

The number of signals plays a crucial role in array processing. The performance of most direction finding algorithms relies strongly on a correctly specified number of signals. When this information is not available, conventional approaches apply information theoretic criteria or multiple hypothesis tests to simultaneously estimate model order and parameter. These methods are usually computationally intensive, since ML estimates are required for a hierarchy of nested models. In the previous work, we proposed a computationally efficient solution to avoid this full search procedure and demonstrated its feasibility by extensive simulations. Here we extend to broadband data, and address issues unique to the broadband case. Our max-search approach computes ML estimates only for the maximally hypothesized number of signals, and selects relevant components through hypothesis testing. Another novelty of this work is the reduction of indistinguishable components caused by overparameterization. Our approach is based on the rank of the estimated steering matrix. Numerical experiments show that despite an unknown number of signals, the proposed method achieves comparable estimation and detection accuracy as standard methods, but at much lower computational expense.