Abstract: A novel method used in the fault diagnosis for diesel engines is presented. The method is based on soft computing theory that fuses together fuzzy theory, neural network and genetic algorithm, and combines itself with the Yinyang compensatory theory. In this way, shortcomings such as an excessive multiplicity of rules, or slow convergence when employing the fuzzy logic or neural network separately can be overcome. The inference system selected here is FNNKD (fuzzy neural network with knowledge discovery). Although the use of this system takes a longer training time, the parameter obtained after training are optimal in a global field. When using this system, accuracy in recognition will increase and the complexity in calculation will decrease. Finally, its feasibility is verified in the experimental study of vibration in gear box and bearing faults of a CA10B automobile.