Abstract: Because of the learning and training speed, the neural networks is not used in dynamic diagnosis with online selflearning in general. And the fault positioning is very important. In this paper, an observer method, is put forward to realize functions of fault isolation and part diagnosis, which is based on neural networks, built on mathematical model for the known part of the object, and combined with compensation net and the fault positioning technique. It is a deep knowledge method with fast convergence speed and good tracking character istics when training process is online.