THzMINet: A Terahertz Model-Data-Driven Interpretable Neural Network for Thickness Measurement of Thermal Barrier Coatings
Fengshan Sun, Binghua Cao, Mengbao Fan, Bo Ye, Lin Liu
Abstract
The complex microstructure of thermal barrier coatings (TBCs) and varying lift-off distance bring a great challenge for accurate terahertz thickness measurements. Available methods to address this challenge are recognized as data-driven or model-driven. However, data-driven approaches are dependent on massive samples, whereas model-driven methods are subject to generate unreliable results. Here, a terahertz model-data-driven interpretable neural network (THzMINet) is developed to measure the TBC thickness. First, an improved terahertz analytical model is formulated to generate simulated signals as the main part of training dataset. Then, a time stream, a frequency stream, and a division layer are separately proposed to constrain the calculation of THzMINet to be the same as the terahertz physics, followed by enabling the online measurement of refractive index, time-of-flight, and thickness. Meanwhile, a novel loss is built to deal with the imbalanced model-data training dataset via an adaptive weight. Finally, the performances of THzMINet are tested by actual TBC specimens, and it allows for accurate and reliable thickness measurements.