Autonomous assessment of delamination in laminated composites using deep learning and data augmentation

Deep learning models can autonomously learn discriminative features from the data; however, insufficient training data often limit their use. This paper proposes a synthetic data augmentation strategy to alleviate the issue of limited data and employ deep learning models for the autonomous assessment of delamination in laminated composites. Contrary to the existing techniques of image data augmentation (rotation, cropping) and time-series data augmentation (adding noise, windowing), the proposed approach brings out additional information during data augmentation through the variation of loading conditions without the need for further experiments. The approach was thoroughly validated and verified in time and frequency domains using various types of experimental vibration testing. The experimentally measured data of 45-time series was augmented to 4,545-time series, resulting in a more rigorous delamination assessment in laminated composites. The proposed approach is autonomous and does not require human-engineered statistical features while using a small amount of measured data. In addition, the approach would assist in tackling the issue of imbalanced data from the healthy and faulty states of laminated composite.