Abstract: Offline rule extraction for the T-S fuzzy systems usually gives a fixed number of fuzzy rules, which make it a bottleneck for revealing the complexity of nonlinear systems. Thus, due to a growing and pruning strategy of the neural network, in this paper the fuzzy rules are extracted from real-time data and their number is adjusted online by the impact degree of one local model, such that the rules vary with the system dynamically and more precisely reflect the character of nonlinear systems. Furthermore, the accuracy of the T-S model is guaranteed by the parameter learning based on a competitive extended Kalman filter (EKF). The entire algorithm presents a completely online identification of the T-S model and gains a structural and parameter adaptability. An example for CSTR identification illustrates its good performance.