비파괴검사 데이터의 머신러닝을 이용한 알루미늄 합금의 강도 추정 σ-구간 기반 데이터 증강 기법 적용

초록

Building on our previous work on the prediction of aluminum alloy strength using machine learning, this study introduces sigma-range-based data augmentation. As in the previous study, the inputs consisted of ultrasonic parameters-introducing longitudinal and shear velocities, an attenuation coefficient, and a relative nonlinearity parameter-and electrical conductivity obtained from eddy current testing, and the output was yield or tensile strength. A total of 705 datasets were experimentally obtained for aluminum 2xxx, 5xxx, 6xxx, and 7xxx alloys and examined. Although this data expanded from the approximately 400 datasets in previous studies, the data for reliable training was still limited. To overcome this issue, an sigma-range-based augmentation method was employed to expand the training data. This technique generates new data by adding random values within the range of k times the standard deviation to each parameter mean. Compared with the tensile test results, the minimum average prediction error was obtained at k = 0.7 for YS-with RMSE = 11.7 MPa and a 3.6% average error-and k = 0.5 for TS-with RMSE = 10.2 MPa and 2.4% average error. This represents a nearly two-fold improvement compared to the technique without augmentation, demonstrating improved prediction performance in data-limited environments.

키워드