To solve the problem that the BP network training algorithm is easy to fall into local minimal point, this paper designs a set of new filled function in view of its characteristic, which can be used to replace the objective function for search and improve the SPDS algorithm. The algorithm simulation test proves that, when the SPDS algorithm falls into local minimum point, the use of the filled function instead of objective function makes the algorithm avoid the fetters of minimum problem, and accelerates convergence to the global minimum point.