POTENTIAL OF VARIATIONAL AUTO-ENCODER AS AN ALTERNATIVE TO A WELDING RESIDUAL STRESS PROFILE SAMPLING MODEL

Primary Water Stress Corrosion Cracking (PWSCC) is a significant degradation mechanism for probabilistic Leak-Before-Break evaluation. Welding residual stress profiles, which have a great influence on the evaluation results, are difficult to measure and have a large uncertainty due to very limited data. Therefore, constructing welding residual stress profiles by reflecting the actual characteristics is an important part in the probabilistic evaluation. This paper analyzed whether a generation model based on artificial neural networks could be an alternative to a sampling model of welding residual stress profiles. The network selected in this study is the Variational Auto-Encoder (VAE), which is well-known for learning the distribution of training data. Like the Auto-Encoder, the VAE is based on deep neural networks composed of an encoder and a decoder, but additionally learns the probability distribution of the input data. Systematically constructed sinusoidal wave training data sets were used to investigate the characteristics of the VAE. Amplitude and phase changes were made to generate the training data sets showing the self-balancing nature of the welding residual stress profile. The results showed that the generated data well represented the features of the training data, and demonstrated its potential as an alternative to a welding residual stress profile sampling model.