Data-driven Fatigue Crack Evaluation based on Wave Propagation Data

8 This project presents a machine learning based fatigue crack sensor 9 evaluation scheme without using a physical model. A supervised binary 10 classification problem is defined to distinguish damaged steel specimens 11 from intact specimens. Sets of time-series wave propagation data collected 12 by the fatigue crack sensor are utilized for the proposed scheme. In order to 13 apply supervised learning methods, different feature sets are defined based 14 on the sensor data using Discrete Fourier Transform (DFT). Two-third of 15 collected data are used as the training set while the rest are used as the test 16 set. The preliminary analysis shows that the features consisting of amplitude 17 distribution in frequency domain and the quadratic kernel function yield the 18 most desirable performance. The trained SVM model, however, performs 19 not very well for the test data set and needs to be improved, especially with 20 larger data sets to reduce high variance. Nevertheless, the proposed learning 21 scheme implies that the wave propagation data has some trend reflecting 22 damage state of a specimen, which shows a great potential of machine 23 learning techniques for fatigue crack analysis. 24 25