The paper presents a novel method of initial weights optimization method in Multi-Layer Perceptron Network(MLPN). Firstly, the sample sets should be transformed by K-L Transform. Secondly, use K-L Converting Matrix to initialize the weights between input and hidden layer. Thirdly the MLPN is trained by BP algorithm, and the convergence speed of MLPN is improved evidently. The ultimate test shows the new algorithm is suitable for the situation of low-dimensional data.