In recent years, several specific advances in the study of chaotic processes have been made which appear to have immediate applicability to signal processing. This paper describes two applications of one of these advances, nonlinear modeling, to signal detection & classification, in particular for short‐lived or transient or signals. The first method uses the coefficients from an adaptively fit model as a set of features for signal detection and classification. In the second method, a library of predictive nonlinear dynamic equations is used as a filter bank, and statistics on the prediction residuals are used to form feature vectors for input data segments. These feature vectors provide a mechanism for detecting and classifying model transients at signal‐to‐noise ratios as low as −10 dB, even when the generating dynamics of the transient signals are not present in the filter bank. The second method and some validating experiments are described in detail. ©1996 American Institute of Physics.