Machine Learning-Assisted THz Technology for Nondestructive Testing in the Tobacco Industry

Abnormal moisture content and impurity contamination are core threats to product quality in the tobacco industry. Rapid, accurate, and non-destructive detection is therefore essential for product optimization. Terahertz time-domain spectroscopy (THz-TDS) holds great potential for tobacco detection owing to its non-invasiveness and high penetration capability. However, the effect of laser pulse width on THz radiation characteristics remains to be verified and the challenge of rapid signal analysis remains unsolved. In this work, a free-space THz-TDS system was constructed, which is driven by two compact Ti:sapphire lasers (with pulse widths of 20 fs and 45 fs, respectively) and enables rapid switching of laser sources. Experimental results indicate that the THz wave generated by the 20 fs laser achieves a broader bandwidth (2.5 THz) and a higher center frequency (0.45 THz). Further analysis reveals a strong linear correlation between the THz signal amplitude and the moisture content of tobacco leaves, while distinct waveform differences are observed between tobacco leaves and impurities. By integrating machine learning algorithms, high-precision impurity classification (with an accuracy of 99.9%) and rapid moisture prediction (with a mean absolute error of 0.28%) are realized. The entire signal measurement and analysis process can be completed within 2 seconds. This work provides technical verification for the intelligent upgrading of tobacco detection systems and demonstrates the enormous potential of THz technology in the tobacco industry and even the broader agricultural sector.