According to the equivalent mapping relation of decision variables,the constrained discrete optimization problems for gear transmission design are transformed into nonlinear constrained non- negative integer programming problems（ CNIPPs） and a discrete differential evolution（ DDE） algorithm is used to solve these problems. An index of average gene distance is introduced to evaluate quantitatively the population diversity. On this basis,this work presents an adaptive escape strategy in which an opposite- based learning operator is employed to generate new individuals to overcome the drawback that the basic DDE algorithm easily traps into local optimal regions for solving discrete optimization problems. Thus this study embeds the escape strategies in DDE algorithm,adopts feasibility rules to handle constraints,and forms to an escape DDE（ EDDE）algorithm for solving CNIPPs. The proposed EDDE algorithm is applied to approach a real case of gear transmission optimization and an index of relative comprehensive performance is presented to compare several algorithms on optimization performances. The experimental and analytical results show that this novel algorithm has good robustness and reliability and is better than compared ones in term of the comprehensive index. Furthermore,the obtained result is better than one of the published literature and the corresponding gear mass is decreased by27%.