Steel Material Microstructure Characterization using Knowledge Distillation Based Transformer Neural Networks for Data-Efficient Ultrasonic NDE System

Material microstructure characterization for texture recognition using ultrasonic testing has been widely used to examine the physical and structural integrity of materials. For system automation, Neural Network (NN) can be used to characterize the material texture accurately. However, training and deploying NNs requires substantial computational resources. In this study, we propose to use the knowledge distillation (KD) method and introduce a response-based teacher-student KD training framework to train NNs to find the optimal solution. In addition, a lightweight Ultrasonic Microstructure Characterization Transformer NN (TNN): UMCTNet, is proposed to recognize material textures using ultrasonic images. Training using the KD mechanism improves the data-efficiency of NN s by transferring knowledge from pre-trained NN models. In addition, TNN utilizes a simple network architecture with the attention mechanism resulting in reducing training and execution time. A data-efficient Ultrasonic Microstructure Characterization Convolutional NN: UMCCNet, is trained as the teacher model using ultrasonic images to distill the pre-trained knowledge into the UMCTNet with high accuracy and data-efficiency. To examine the results, an ultrasonic testbed platform was assembled, and C-scanning was created to acquire volumetric ultrasonic data from three different heat-treated steel blocks to train NNs. By applying the KD framework to train UMCTNet, we obtained the training and testing accuracy of 99.91 % and 99.27% respectively, and the highest image throughput of 192 images/seconds on testing to characterize steel material microstructures.