Aiming to the characteristics of real passive sonar signal that broad-band, nonstationary and little knowledge of statistics. A fusion cost function is established which combined with the property of time delay structure and non-Parametric under broad-band convolution mixing model. Using a non-parametric kernel density estimation technique, the algorithm performs simultaneously the estimation of the unknown probability density functions of the source signals and the estimation of the unmixing matrix. The proposed algorithm uses the optimized unmixing matrix and estimated target source signals to achieve a bearing-energy spectrum in each frequency bin which accumulated to obtain the broad-band bearing-energy spectrum. Broad-band simulation results and real sea trial shows that the proposed method is close to the minimum variance distortionless response (MVDR) and multiple signal classification (MUSIC) methods in the aspect of bearing resolution and estimated accuracy, however, which play better in the aspect of weak target detection in which strong interference exists.