Continuous monitoring of moisture loss of beef, beetroot, and banana slices during microwave vacuum dehydration by using THz-TDS combined with transformer-based neural network

Abstract

Microwave vacuum dehydration (MVD) has emerged as a preferred alternative to conventional methods such as hot air drying, offering faster dehydration rates while better preserving product quality. Despite these advantages, challenges remain in implementing effective real-time monitoring systems and accurate dehydration prediction methods during the MVD process. This study investigated the feasibility of using terahertz time-domain spectroscopy (THz-TDS) to continuously monitor the drying kinetics of beef, beetroot, and banana slices during MVD without interrupting the dehydration process. Polytetrafluoroethylene (PTFE) was demonstrated as the most suitable airhose material among polyethene (PE), PTFE, and quartz with the highest transmittance of 0.824. Using the deep learning model of transformer-based neural network (TbNN) introduces the self-attention mechanisms to extract features at characteristic frequencies. It successfully correlated THz-TDS transmittance data with actual moisture loss of samples with a prediction accuracy of 0.96, which shows excellent generalisation capability of this TbNN model on such a small dataset size. Besides, the calibration strategy successfully improves the accuracy from 0.94 to 0.96, with a regression coefficient of R = 0.98328. The integration of these sensing and analytical techniques offers a valuable framework for improving industrial processing control while broadening the applications of THz-TDS technology across agricultural and food production sectors.