A new adaptive variable neural network is developed according to the inner product feature of sample space.This method does not need to iterate learning.It features that its self-learning velocity is fast,accuracy is high,and it is able to change self-structure on work according to producing new sample.It has been applied to a type of aeroengine start system,its accuracy of failure diagnosis is prior to the original one,and the maintenance work is cut down.Moveover it has capability for real-time failure diagnosis,which opens up bright prospect for its propagation.