A study on depth classification of defects by machine learning based on hyper-parameter search. (15th February 2022)

• Record Type: Journal Article
• Authors: Chen, Haoze; Zhang, Zhijie; Yin, Wuliang; Zhao, Chenyang; Wang, Fengxiang; Li, Yanfeng
• Is Part Of: Measurement. Volume 189(2022)
• Publication Date: 2022-02-15
• Subjects: Hyper-parameters search -- Defect depth classification -- K-nearest neighbor -- Support vector machine -- Random forest; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.110660
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: This paper explores the effect of different defect depths on heat transfer. Hyper-parameter search is used for defect depth classification. Three different machine learning algorithms are applied to defect identification. It can achieve 100% accuracy of defect depth classification within 0.63 s. Abstract: To overcome the low efficiency of crack depth detection of steel, we explored for the first time the method based on hyper-parameters search in the field of defect depth classification. And the effect of different defect depths on the heat transfer to the metal surface during heating and cooling process was analyzed. Moreover, we de-noise the infrared thermal images by median filtering algorithm. Then we propose two time-series temperature features: the crossing temperature feature and the temperature difference feature, and compared their robustness. We perform hyper-parameter search by grid search and random search, for KNN, SVM and random forest. Experiments prove that the temperature difference feature is effective in this study. The KNN based on grid search can achieve 100% accuracy. The SVM has the highest classification efficiency, that based on grid search and random search can achieve 100% classification accuracy in 0.63 s and 0.78 s, respectively.