To address the difficulty and low accuracy of the non-destructive detection of thermal growth oxide (TGO) thickness, a detection method combining terahertz time-domain spectroscopy and a deep learning model is proposed. By establishing a simulation model of the thermal barrier coating, the terahertz time-domain spectral data of the simulation and physical sample are obtained, and the deep learning model is optimized to improve the accuracy of TGO thickness detection. When processing the simulation data, the average determination coefficient of the model is 0.934, whereas when processing the physical sample data, the average determination coefficient is 0.857, indicating a high degree of fitting. For the detection results of TGO thickness in the range of 3?10 μm, the mean relative error of the model is less than 10%, indicating a high detection accuracy. This method effectively captures the complex relationship between TGO thickness and terahertz time-domain spectral signals, and provides technical support for the life assessment and failure prediction of thermal barrier coatings.