A hybrid convolutional neural network for intelligent wear particle classification. (October 2019)

• Record Type: Journal Article
• Authors: Peng, Yeping; Cai, Junhao; Wu, Tonghai; Cao, Guangzhong; Kwok, Ngaiming; Zhou, Shengxi; Peng, Zhongxiao
• Is Part Of: Tribology international. Volume 138(2019)
• Page Start: 166
• Page End: 173
• Publication Date: 2019-10
• Subjects: Wear particle classification -- Convolution neural network -- Transfer learning -- Support vector machine; Tribology -- Periodicals; 621.89
• DOI: 10.1016/j.triboint.2019.05.029
• Languages: English
• ISSNs: 0301-679X

Abstract

Abstract: For the purpose of automatic wear debris classification, a hybrid convolution neural network (CNN) is used with transfer learning (TL) and support vector machine (SVM) to identify four types of wear debris including cutting, sphere, fatigue and severe sliding particles. Experimental results indicate that image features extracted from the CNN is more distinguishable than that acquired from the local binary pattern, the histogram of oriented gradients and the color-based methods. The classification accuracy and efficiency of the proposed hybrid CNN with TL and SVM is also higher than that of the CNN, the CNN with TL, and the CNN with SVM. This work provides an effective solution for automatic wear debris identification applicable for machine wear mechanism analysis. Highlights: A hybrid CNN model is developed for automatic wear particle classification. The classification model contains particle feature extraction and recognition, which improves the analysis accuracy and efficiency. Weights and biases of the CNN are initialized with the ImageNet trained parameters to improve generalization. A new wear particle classifier is built by augmenting the CNN and SVM to increase the classification accuracy.