A method for process fault detection is presented, based on the integration of multiscale signal representation and scale-specific clustering-based diagnosis. Previous work has demonstrated the utility of our multiscale detection scheme applied to linear projection-based methods, such as PCA and Dynamic PCA. This work further demonstrates the use and method independence of the multiscale scheme by applying it to a nonlinear modeling method, namely Adaptive Resonance Theory-2. The multiscale ART-2 (MSART-2) algorithm detects a process change when one or more wavelet coefficients violate the similarity thresholds with respect to clusters of wavelet coefficients under normal process operation at that scale. In contrast to most other multiresolution schemes, this framework exploits clustering behavior of wavelet coefficients of multiple variables for the purpose of scale selection and feature extraction. By reconstructing the signal with only the relevant scales, MSART-2 can automatically extract the signal feature representing the abnormal operation under consideration. Illustrative examples as well as Monte Carlo bases for these claims via a comparative performance analysis over several case studies are provided. Comparison of average detection delays or run-lengths of MSART-2 with those of ART-2 for a variety of processes with different statistical characteristics is provided. Comparative results on real industrial case studies from a petrochemical process plant are also presented. Results indicate that MSART-2, as compared to ART-2, is a general approach that may be preferable for problems where it is necessary to detect all changes drawn from processes of various statistical characteristics. (C) 2004 American Institute of Chemical Engineers.