SM-GMVAE: An intelligent evaluation model for defect depth based on few ultrasonic signals

Ultrasonic non-destructive detection is widely used for recognition and estimation of structural defects. Deep learning, especially deep neural network (DNN) has become a research hotspot for defect automated evaluation. Nonetheless, most current models are based on supervised learning approaches. To improve the performance of model, more data is needed to train model. Unfortunately, the collection of data in industrial scenarios is often limited and data labeling is also a time-consuming and labor-intensive task. In order to overcome this problem, This paper proposed a novel Similarity Metric Gaussian Mixture Variational Auto-Encoder model (SM-GMVAE) that combines few-shot learning and non-destructive testing techniques to evaluate defect depth with limited data. This model is designed into two modules: feature extraction (FE) module and similarity metric (SM) module. The FE module is designed to extract the feature of defect signal via the Variational Auto-Encoder (VAE). The SM module is used to measure the similarity of two defect signal based on the Gaussian Mixture Model (GMM). Moreover, sparse filtering techniques are used to enhance the fused features in the SM module. To validate proposed model, several specimens containing defects of different depths were produced. We construct the defect dataset based on defective ultrasound detection signals and several case studies on this datasets. The results demonstrate that the GMM and sparse filtering techniques used in our model can improve the model evaluation accuracy, and the proposed model outperforms other models.