Knowledge-guided CNN Model for Similar 3D Wear Debris Identification with Small Number of Samples

Wear debris analysis (WDA) enables the provision of essential information towards the monitoring of machine fault diagnosis and the analysis of wear mechanism. However, this experience-based technology has not yet been automated for the identification of similar particle types due to the small number of samples and highly-dispersed features. To address this problem, a knowledge-guided convolutional neural network (KG-CNN) model is developed to focus on two representative severe wear particles: fatigue and severe sliding particles that have highly similar contours but weakly discriminative surfaces. The height images of particle surfaces are adopted as the initial objective. Characterized by typical particle features, the empirical WDA knowledge is represented into the feature-marked images, and further automatically learned by a U-Net-based knowledge extraction network. By weighting with the U-Net output, a knowledge-guided particle classification network is constructed to identify similar particles under a small number of samples. With this methodology, the empirical WDA knowledge is transferred to guide the classification network for locating the discriminative features in particle height images. Thirty sets of fatigue and severe sliding particles are acquired from wear tests as the training and testing samples. For verification, the network kernel is visualized to trace the particle feature propagation in the classification. Experimental results reveal that the proposed method can accurately identify fault particles that acquired from wear tests.