Generation and classification models of ultrasonic signals in aged cast austenitic stainless steel (CASS)

Abstract

In this study, we attempted to generate new ultrasonic signals from thermally aged CASS specimens using a statistical-based model to augment the experimental data. A variational auto-encoder (VAE) model is adopted to generate the ultrasonic signals. To verify the quality of the generated data (signals), two ML algorithms (k-nearest neighbor and multi-layer perceptron algorithms) were used for classification with newly generated data for training and experimental data for testing. The results showed that new ultrasonic signals could be generated effectively with a small amount of experimental data, with the newly generated data showing features approximating those of the experimental data, indicating high classification accuracy rate for ultrasonic signals from aged CASS specimens. In addition, a sampling method for the generated model is suggested.