Analysis of transients on basis of identification of signal generative structure: Even unto chaos

A novel approach to the analysis of transient signals is proposed and demonstrated. This approach is based on the use of a neural net to estimate the future values of a transient time sequence. In the course of training the net, it learns a model of the process which is the source of the transient signal. This type of analysis is similar to the act of parsing in syntactic pattern recognition, and the act of synthesizing a model of the process is similar to grammar inference in syntactic pattern recognition, except that in the present case the signal values are numeric and are continuously valued. The approach is demonstrated with use of a nonlinear dynamic model as the signal generator. With this model, it is possible to move from simple signal environments to chaotic regions. Numerical results are presented to illustrate highly accurate predictive estimates of the signal source output