Emadaldin Sh Khoram-Nejad - PhD student, Acoustics Research Laboratory, Mechanical Engineering Department, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran
Abdolreza Ohadi - Professor, Acoustics Research Laboratory, Mechanical Engineering Department, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran

This paper aimed to detect the crack location in a shaft-disk-cracked blade (SDCB) system. An open crack with three different depths and three distances from the blade root is considered. A finite element (FE) simulation of the SDCB system rotating at ۳۰۰۰ rpm was done in the MSC Adams as commercial software, simulating the systems' dynamics. The results are compared with those of another FE software and the assumed mode method (AMM) to validate the modeling procedure. Feature extraction and selection always have crucial roles in classification problems. Therefore, this paper tries to show the benefit of automatic feature selection compared to manual feature selection to detect the crack location. For this purpose, the performance of an ensemble classifier using manual feature selection is compared to a deep multi-layer perceptron (MLP)-based classifier, which selects the best features automatically and determines the crack location. The ensemble classifier includes the support vector machine (SVM), k-nearest neighbor (k-NN), and decision tree classifiers. Then, the sensitivity of the proposed crack detection models to different measuring points, one at the bearing location and another near the cracked blade tip, is studied. The best location for sensors to detect cracks more accurately is determined during this comparison. The results indicate that the best accuracy of crack detection is achieved when using the MLP-based classifier and the signals of the displacement of the cracked blade tip. This accuracy is ۹۷.۱۴%, and the existence of a crack is detected without an error (۱۰۰% accuracy).

Crack detection; Feature selection; Assumed mode method; Multi-layer percep-tron; Ensemble classifier