This paper searches the nearest training data set as the training data sub-set in fault characteristics data space, and applies the steepest descent method (back-propagation) to optimizing the parameters of the fuzzy rules in the local model. Then though simulation on Fisher's iris data set and comparison with adaptive neural fuzzy inference system (ANFIS) the average test error is reduced by 15% and the operation speed is increased by about 30%. After a fault characteristics data set from actual aeronautic engine test is put in this fault detection model system, the model system can accurately identify the three kinds of fault states existing in the engine, and the result shows that the fault diagnostic strategy is efficient and available for some fault diagnosis problems.