Wear particle classification using genetic programming evolved features

This paper explores the feasibility of applying genetic programming (GP) to classify wear particles. A marking threshold filter is proposed to preprocess ferrographic images before optimising the feature space of wear particles using GP. Subsequently, evolved features by GP are quantitatively evaluated by the Fisher criterion and distance fitness function, and clustering performance is evaluated qualitatively. The evolved features are compared with a conventional feature set as the inputs to support vector machines, probabilistic neural networks, and k-nearest neighbour. Results demonstrated that the evolved features indicated a significant improvement in classification accuracy and robustness compared with conventional features. Finally, 3 typical wear particles, sliding, cutting, and oxidative, are successfully classified.