Terahertz time-domain spectroscopy imaging combined with convolutional long short-term memory neural network for In situ moisture reduction prediction of particle foods during pulsed fluidized bed drying
Yuqiao Ren, Xianglu Zhu, Brijesh K. Tiwari, Da‐Wen Sun
Abstract
Fluidized bed drying (FBD), as a novel dehydration method, has been widely applied in drying various heat-sensitive high-value food products, in which the pulsed fluidized bed drying (PFBD) technique addresses the ineffective fluidized limitation of the traditional FBD when dealing with cohesive products. However, achieving in situ real-time monitoring of dynamic product moisture content (MC) change during both FBD and PFBD without sampling is still challenging. This paper presents an innovative integration of a whisk-broom terahertz time-domain spectroscopy (THz-TDS) imaging system with a fluidized bed dryer, facilitating in situ, real-time monitoring of MC reduction of products without process pausing for sampling. The reliability, and robustness of the established system and prediction models were tested by drying peppercorn and sweetcorn under various drying settings, including different temperatures, flow patterns, and durations. Implementing the convolutional neural networks (CNN) and long short-term memory (LSTM) for correlating terahertz time-domain image data with reference datasets in this research provided reliable MC reduction prediction results for the mixture dataset from peppercorn and sweetcorn drying. The successful application of this combined sensing and analytical method presents a promising approach for enhancing industrial processing control and expands the utility of THz-TDS in the agriculture and food industrial fields.