Multiple classification without model order estimation

A novel approach to weighting eigen-subspaces is proposed which leads to a pseudo-spectrum with super-resolution performance similar to the MUSIC algorithm without the need for model-order detection. The method is shown to be applicable to a type of problem encountered in active carrier-wave sonar direction of arrival estimation where the SNR is high and the targets are discrete point targets and few in number. The method forms a covariance matrix with subsequent eigendecomposition as in MUSIC. Similarly, a pseudo-spectrum is formed as a reciprocal of a weighted reconstruction. Weights are used for all eigenvectors, contrasting with MUSIC & weighted subspace methods which typically use only the noise subspace. Further the weights proposed here include functions based on the nature of each individual eigenvector which can be related to the sparsity of the components of the eigenvector. The presented results of digital simulation demonstrate a performance comparable to MUSIC & MUSIC + MDL for a range of signal/target positions, relative strengths & SNR values. Further, application of the algorithm to the real sonar data shows promising improvement in target localisation.