Abstract: The performance of neural networks applied in structural health monitoring(SHM) is carefully discussed by using probability tool. Then two types of errors of the network are distinguished. And the further study shows that there are 3 parameters that can influence the performance of the network and the influence varies with the structural serving time. A SHM-oriented mistake curbing strategy is presented to agilely curb the two types of errors on demand of SHM. Then it is applied in designing Back-Propagation neural network. The new error energy function is defined and the way setting the value of mistake curbing coefficient is suggested in the consideration of long time monitoring. Further more, the weight updating rule is derived.