Physics inspired sparse deconvolution network for dispersion compensation of terahertz wave

Terahertz (THz) technique, as a rising nondestructive testing (NDT) approach, has emerged great potentials in the nondestructive assessment of various non-metallic materials due to its high time resolution and penetrability. However, the dispersion effect of THz wave during propagation process can cause the broadening, overlap and distortion of THz response signal, which brings a significant challenge for accurate extraction of time-of-flight (ToF) information of structure. The present methods mainly rely on traditional signal processing techniques to improve the estimation accuracy of ToF from the signal, rather than focus on the effects of dispersion on detection resolution. Also, such methods are usually limited by the sophisticated manual interventions and the requirements of prior knowledge of structure, which may fail for complex signals from multilayer structure. Therefore, in this work, a physics inspired sparse deconvolution network is specially proposed for the dispersion compensation of the signal. The core aims to utilize the sparse deconvolution network to address the sparse inverse problem in THz testing, and establish the mapping relationship between dispersive signal and sparse impulse vector corresponding to ToF. Initially, an improved physical transmission model is established based on the dispersion characteristics of the wave. Subsequently, the sparse deconvolution network is specially constructed to obtain ToF information based on THz detection mechanism. Then, the dispersion compensation process can be performed by convoluting the estimated ToF with the THz reference signal. During this process, to improve the dispersion compensation performance, the progressive curriculum weighted loss (PCW-Loss) and the amplitude-position joint error (APJE) metric is designed for the optimal performance. Finally, a series of numerical simulations and experiments, including the single-layer and multi-layer structures, are implemented to verify the effectiveness of the proposed method. Overall, the proposed method is the first attempt to solve the dispersion issue of the signal by sparse deconvolution network, and can provide a novel insight and solution for the accurate and automatic ToF estimation in various THz NDT scenarios.