Optimized CNN model for identifying similar 3D wear particles in few samples

Typical wear particles can be considered as distinctive indicators of on-going wear faults in machines. However, the small number of samples has limited the identification accuracy for similar fault particles. Besides, the three-dimensional (3D) characterization of wear particles may face huge challenges due to excessive surface parameters. Focusing on the problems of high similarity particles and few samples, a CNN-based particle classification method is developed with an example set of fatigue and severe sliding particles, which are the product of severe wear of machines. For data reduction, imaged 3D particle surfaces are firstly converted into 2D depth maps without losing surface information. Virtual fault particle images are then synthesized using a Conditional Generative Adversarial Networks (CGAN), according to the particle generation mechanism and distinctive particle features. Furthermore, a non-parametric particle identification model is established with the optimization of the CNN structures and training method, and the network is further optimized with the particle image standardization and the network visualization. Validation experiments reveal that the proposed method can accurately identify all tested fatigue and severe sliding particles with their typical characteristics.