Few-shot ultrasonic defect classification in insulation materials using finite element simulation and Siamese neural networks

Abstract

High-performing models for structural defect classification and intelligent recognition generally require extensive, high-quality datasets. However, acquiring authentic defect data is usually costly. This paper investigates an intelligent defect classification model based on a Siamese neural network, utilizing extensive finite element simulation data supplemented by a limited set of experimental data. This study focuses on basin-type insulator materials with internal circular hole defects. For each of the 11 different defect sizes and locations, 300 simulation samples were generated and embedded with experimental noise. An ultrasound signal was measured for each defect type and paired with the corresponding simulation data to form the training set. A Siamese neural network was constructed to extract common features from both the experimental and simulation datasets. The model was tested using experimental data from various defect specimens, resulting in a classification accuracy exceeding 95 % on the two-dimensional reconstructed experimental data. The research showed that the accuracy of the model with a single input experimental sample is comparable to the supervised learning results with 80-shot setting. This indicates that the Siamese neural network can effectively learn common features between simulated and experimental data, achieving satisfactory classification performance even with a small number of experimental samples.