A fast, new classification system for engineering surfaces has been developed and used to detect the incidence of wear. The novel feature of this system is that the classification can be performed without the need for any surface parameters. In addition, only normal working conditions data (the target class) are required to train a classifier. In this system, first, a surface to be classified is represented by a set of dissimilarity measures (e.g. differences in surface height) calculated between the unclassified surface and already pre-classified surfaces belonging to the target class. The representation set of measures is then used to assign a surface into the target class or reject as an outlier. Outliers are anomalies or faulty conditions that can be ill-defined, undersampled data or even unknown data. However, several problems still remain to be solved before the approach can be used as a fully functioning pattern recognition system for the applications in machine condition monitoring. This includes difficulties associated with selecting a right size of the representation set and building an accurate one-class classifier. These problems have been addressed in this study and analysis results for unworn and worn surfaces have been also presented. It was found that (i) skewness, correlation, principle component analysis dimensionality and boundary descriptor are well suited for selecting the representation set and (ii) a combiner of the Parzen and support vector data description (SVDD) classifiers with the median rule gives better classification results than single classifiers (i.e. Gaussian density, mixture of Gaussian densities, Parzen density and SVDD classifiers). (c) 2006 Elsevier Ltd. All rights reserved.