Online identification of piston crack based on BOWHOG feature
                                                                        CHEN Rui, LIU Zhentao
                  (Power Machinery and Vehicular Engineering Institute, Zhejiang University, Hangzhou 310027, China)

Abstract: Aiming at solving the problem of online crack identification in fatigue simulation bench test, the machine learning method based on BOWHOG feature was applied and improved. The HOG feature extraction method was used to extract crack features of various shapes, and the word bag model and k means clustering were used to generate a characterization dictionary for cracks. The directional retention was improved by the enhanced training method of rotation and symmetry and the dimensional retention was improved by image pyramid. Then the image region of interest was extracted by the joint gray level distribution image difference method and the online recognition of the piston crack in the complex environment was realized by support vector machines. The results indicate that the improved BOWHOG algorithm has an accuracy of 84.1%, showing better size and rotation retention, consuming less computing resources, and significantly improving the accuracy and robustness of crack identification algorithm limited to few training samples.

Key words: piston; gray gradient histogram; bag of words model; feature extraction; support vector machine