Abstract: The optimum design of the self-propelled combine gearbox showed some deficiencies such as too much calculation and too many charts. A new algorithm was proposed which improves the training of neural networks. Different from previous approaches, this new approach focuses on the samples, emphasizes particularly on parameter adjustment of networks. Via fuzzy deduction, adjustment of the samples and classified training, a better self-adaptive training performance was achieved. After learning and training of BP Artificial Neural Networks(ANN), single-input and double-output structures in the form of 1-8-2, 1-6-2, and 1-4-2 were adopted. According to practical effectiveness, 1-6-2 structure with high identification precision as the final BP ANN form is selected, and the identification precision of BP ANN is perfectly high. This approach based on artificial neural networks to design the combine gearbox, gave a mathematical description. The method establishing gearbox design model by neural networks was analyzed. Study showes that the neural networks model can reduce the frequency of system analysis, improve the precision of the model to a great degree, meet the requirements of calculation, and then find out an improved scheme from design space.