Abstract：There is a very expansive application foreground for prediction model of terrane crack depth for combined perforating-fracturing in oil well. Simple prediction model is very easy to achieve by ANN method with BP algorithm. However，the prediction credibility keeps unknown in condition that learning samples change after learning. So how to deal with the changing becomes an important issue. Much effort is made to approach a dynamic model that can adjust itself when samples partially change. Based on the investigation of ANN model with BP algorithm，an abecedarian upgrade strategy and algorithm of single hidden-layer BP networks are given in this paper. This strategy includes three key parts，classified learning with the basis of sample unit comparability judgment after sample space changes，self-adaptive arithmetic of classified sample appearance probability during learning and self-adaptive control method of network scale. Firstly，learning samples changes partially，and new samples will be sorted by the different aims of upgrade. Secondly，the upgrade learning starts from current state of the network，and network scale might be adjusted to suit the new sample space. Finally，with less upgrade learning steps，the network can adapt to the new sample space，without a whole second learning course. In order to test this strategy，the upgradeable self-adaptive prediction model of terrane crack depth for combined perforating-fracturing is established with the satisfaction of precision. All samples come from systemic numerical FEM analysis with FINAL under the condition of blast load. This model can adjust its scale and the weight matrix rapidly on the self-adaptive rule，in allusion to the local alteration of sample space. To the problem of sample changing in the ANN model，all of these practically offer a new way with single-hidden-layer BP algorithm. This self-adaptive upgrade strategy can be applied to other modeling courses.
