Automated 3D ferrograph image analysis for similar particle identification with the knowledge-embedded double-CNN model

Ferrograph-based wear debris analysis (WDA) provides essential information for the root cause analysis of wear failures. However, this technique has been hampered as an intelligent approach by two problems: lack of fault particle samples and conflicts among redundant features. To address this issue, a knowledge-embedded double-CNN model is proposed to identify two representative fault particles: fatigue and severe sliding particles, by using the 3D topographical information. First, a non-parametric CNN network model is constructed with a 2D height map of 3D particle surfaces. The convolution kernels are evaluated to determine identification errors due to the small number of samples. In the refinement stage, four efficient kernels are extracted via the image similarity with the labeled images, which are created based on the physical wear mechanism of the two types of particles. Furthermore, an improved CNN network with six parallel convolution layers is established to handle the feature maps of these kernels for objective particle identification. The proposed model is trained by 20 groups of fault particles and further verified with 10 groups of shuffled particle samples and the network visualization. Validation experiments reveal that discriminative features have contributed to accurately identify all tested fatigue and severe sliding particles.