Abstract:

A potential function method with estimated parameter is proposed for blindness of scale parameter selecting of k-dimensional subspace method. Firstly, the reference value r₁ is estimated based on observed signals and the scale parameter σ₁ for potential function f is estimated based on r₁.When cluster planes are estimated by f, the reference value r₂ is estimated based on them and the scale parameter σ₂ for potential function g is estimated based on r₂.Thus, we obtain the estimation of mixing matrix. This method makes full use of the statistical properties of observed signals. The experiment shows that this method is reduced by 75% in matrix estimation error than the method before improved and is reduced by 1~2 orders of magnitude compared with other methods when there is no noise; the estimated precision is 8.7% higher and the accuracy rate of the source’s number estimation is twice higher.