Abstract: A signal recovery method based on BP neural networks is proposed in this work to address the noise caused by water vapor absorption of terahertz waves in the environment. The method involves feature extraction on the amplitude of the collected signals in the frequency domain using principal component analysis (PCA), followed by the optimization of the back propagation (BP) neural network parameters using particle swarm optimization (PSO) algorithm. The results show that the BP neural network achieves the best signal recovery for internally validated explosives DMDP and RDX, with signal similarities of 0.995 and 0.999, respectively, the PCA combined with BP neural network (PCA-BP) achieves the signal similarity of 0.986 and 0.944, respectively, and the PSO-optimized PCA-BP (PSOPCA-BP) recovery achieves the similarities of 0.987 and 0.997, respectively. For externally validated explosives 345TNT and NQ, the BP neural network's recovery similarity scores are 0.273 and 0.278, respectively, PCA-BP shows improvements with scores of 0.944 and 0.985, respectively, and PSO-PCABP yields the best recovery results with similarities of 0.946 and 0.993. Additionally, by combining phase and amplitude information, the time-domain signals are obtained using the inverse Fourier transform, which proves that the proposed method is equally effective in improving time-domain data.

Key words: terahertz spectroscopy, water vapor noise removal, BP neural network, principal component analysis, particle swarm optimization