Research on Self-Learning Flaw Detection in Ultrasonic C-Scan Non-Destructive Testing

The ultrasonic C-scan technique is widely utilized for corrosion monitoring of metal plates and pipelines due to its ability to generate comprehensible images of flaws. However, the application of this technique has been challenging because of the large-scale and unbalanced signals reflected from the tested objects. In this paper, we present a method for identifying ultrasonic response signals of defects using an ultrasonic detection with contrastive learning (UD-CL) framework. This self-learning framework is used to unsupervisedly learn latent representations of echo signals obtained from ultrasonic C-scans. The algorithm addresses both the imbalance in the distribution of raw echo data and the lack of sufficient annotated ultrasonic signals. The UD-CL method consists of two components: a feature extractor based on a convolutional network and a learnable nonlinear transformation layer, which are used to obtain an efficient latent representation and perform contrastive optimization, respectively. We demonstrate the effectiveness of our approach by applying it to an aluminum plate with various types of flaws. The results show that the UD-CL outperforms previous deep learning techniques in terms of accuracy in identifying the location and category of flaws when trained using only 10% of the labels.