Restoration of defocused ferrograph images using a large kernel convolutional neural network

Abstract

Wear debris analysis is used for machine condition monitoring because it is capable of identifying wear related faults in machinery components. In particular, wear mechanisms can be explored based on debris shape and texture features. As current wear monitoring practices prefer rapid data interpretation, on-line ferrograph is favored where debris images are collected and processed on-site with an imaging apparatus. However, such images are often blurred due to lubricant contamination and limited optical field depth. This drawback makes the subsequent investigation of wear debris features a challenging task. To address this issue, a restoration method is developed to reduce the influence of out-of-focus blurring in on-line ferrograph images. Firstly, in order to model the out-of-focus blurring process, a degradation model is constructed with a convolutional neural network. Different from general deep learning structures, larger convolutional kernels are embedded in the network to more comprehensively characterize the blurring process. Next, training samples of different types of debris are collected in the laboratory. The neural network structure then is optimized to minimize the difference between restored images and their corresponding ground truths. The final restoration is conducted based on the trained network via a pixel-wise regression process. To evaluate the performance of the developed method, experiments are carried out with debris images collected from a gear-box test rig. Results have demonstrated that: (1) the proposed restoration strategy allows robust extractions of debris features in greater detail; (2) the developed method is more computationally efficient than traditional deblurring techniques, which enables it to be applicable in on-line machine health monitoring.