Abstract: There are many non-linear phenomena in engineering applications, such as limit loop, resonate jumping phenomenon, range frequency depending phenomenon etc., and these phenomena are materially different characteristics of different non-linear systems. Threshold automatic regression model （TARMA） has been widely used in time series modeling because it has the character of generality and can explain the phenomena cited above with physics meaning. Evolutionary TARMA modeling algorithm is proposed which can overcome some limitations of traditional methods including H. Tong method, D.D.C method and local research method. The algorithm can also identify the type of model（linear or non-linear）, order number of model and some relevant parameters（threshold interval parameter, threshold parameter, and the corresponding parameters of ARMA model） etc. The experiments show that the algorithm is effective, global, self-adaptive and robust. Moreover, the models constructed are abundant because of the existence of randomness, so decision makers can select appropriate models to analyze time series or explain physically, which has the characteristics of automatization.