The influence of water vapor absorption is inevitable in optical path from the THz source to detector without nitrogen gas, resulting signal distortion at the tail of THz pulse. In the frequency spectrum, the unwanted water-vapor absorption lines will emerge, making it difficult to identify the possible specific absorption lines of the measured sample near the water-vapor absorption lines. To eliminate water-vapor influences can greatly expend the applications of THz-TDS system in the most common open-air environment. In this paper, we used a Support-Vector-Machine algorithm (SVM) in the recent advanced machine learning technology to recover the actual THz signal by removing the influence of water vapor degradation. The learning and prediction of the water vapor absorption effects is completed by iterative training process of the SVM algorithm. After the SVM model is built, we found that it can effectively eliminate the fluctuations of the THz-TDS signal obtained in the open-air environment, thus the corresponding water-vapor absorption peaks in the frequency spectrum are greatly suppressed. We also compared the signal recovering ability of our SVM algorithm with traditional Back- Propagation(BP) neural network algorithm with the same training data as well as the same training time and found that the SVM algorithm outperforms the traditional BP neural network algorithm. To furtherly verify the generalization ability of our SVM model, the THz signals measured under different humidity are set as the inputs of the SVM model. It turns out that the SVM algorithm can still effectively eliminate the water-vapor absorption effects.
|