Roll contact fatigue defect recognition using computer vision and deep convolutional neural networks with transfer learning

Abstract

An end-to-end machine learning approach for classifying rolling contact fatigue (RCF) defects utilizing defect images is presented and evaluated. The core component of this approach is the use of a fine-tuned AlexNet architecture (FT-AlexNet), which is a well-known pre-trained deep Convolutional Neural Network (DCNN). Through comparing the FT-AlexNet method with two classical two-step classification methods that include a feature extraction step and then train a classifier, it was found that the FT-AlexNet could not only avoid the need of additional steps and variability involved in selection of feature extraction methods and classification strategies and parameters, but also obtain the comparatively better classification accuracy and generalization ability. In addition, the 'black box' working principle of FT-AlexNet was analyzed through visualization, which displayed its robustness to noise and background interference to some degree. However, it was also found that the FT-AlexNet architecture, although improved compared to the more traditional methods, was not as accurate for the identification of micro defects for cases with substantial variation in the image background.