Online monitoring of oil wear debris image based on CNN

Abstract

Image monitoring of oil wear particles is currently only applicable to microflows and is susceptible to bubble interference. This paper develops an optical oil-monitoring system that can be used for large-diameter pipes with high flow rates. A shallow and wide observation cell with an equivalent diameter of Φ5 mm is designed to allow a theoretical maximum monitoring flow rate of about 8 L/min, which is a significant improvement over current image monitoring of generally less than Φ2 mm pipes. A low-magnification (0.8X – 5X) stereoscopic microscope head is used to improve the field of view and depth of field, and a high-speed camera is used to increase the flow range that can be monitored. A set of experimental platforms is also constructed to produce bubbles and wear particles separately. Images of the wear particles and bubbles are then collected for subsequent training and verification of image classification algorithms. A motion object extraction algorithm based on background differences and the Otsu method is used to extract debris and bubble images, and a convolutional neural network (CNN) algorithm is used to distinguish between bubbles and debris. Compared with the traditional morphological feature extraction method, histogram of oriented gradient (HOG) feature extraction method, k-nearest neighbor (KNN) classification algorithm, and support vector machine (SVM) classification algorithm, the CNN algorithm eliminates the tedious process of feature extraction and selection, and has better classification results. The experimental results show that the system can effectively collect wear particle and bubble images and classify them, and the classification accuracy can reach 91.8%.