An Adaptive Genetic Algorithm Optimizes Double-Hidden Layer BPNN for Rapid Detection of Moisture Content of Green Tea in Processing

Abstract

Moisture content (${\text{M}}_{\text{C}}$) plays a crucial role in evaluating the quality of tea processing. However, the current automated production line for green tea heavily relies on manual methods to determine ${\text{M}}_{\text{C}}$, which leads to low productivity and inadequate automation. Therefore, there is an urgent need for a fast, accurate, and convenient ${\text{M}}_{\text{C}}$ detection method. In this study, near-infrared spectroscopy (NIRS) data were collected from seven stages of green tea processing and preprocessed using various techniques, such as Savitzky-Golay (SG) and detrend (DT), to reduce spectral noise. Subsequently, feature variables of the preprocessed spectral data were selected using full-band principal component analysis (PCA) and competitive adaptive reweighted sampling (CARS). Afterwards, prediction models for ${\text{M}}_{\text{C}}$ of green tea were developed using partial least squares regression (PLSR) and back-propagation neural network (BPNN). To address the convergence speed and local optima issues of BPNN, the study proposes an adaptive probabilistic genetic algorithm (AGA) to optimize the initial weights and thresholds of BPNN, including single and double-hidden layers, respectively. The results demonstrate that the double-hidden SG-DT-PCA-AGA-BPNN model outperforms the single-hidden layer model, achieving a high correlation coefficient ($R_P$) of 0.994 and a low root mean square error (RMSEP) of 1.01%. This study highlights the effectiveness of increasing the number of hidden layers and using AGA to optimize the initial thresholds and weights of BPNN in improving the prediction accuracy. Furthermore, it provides a new approach to implement ${\text{M}}_{\text{C}}$ detection technology in green tea processing.