An Optimized Bidirectional Long Short‐Term Memory Model Based on Hyperspectral Analysis of Protein Content in Milk Powder

Protein content is an important index in the assessment of dairy nutrition. As a crucial source of protein absorption in people's daily life, the quality of milk powder products not only has a deep impact on the development of the dairy industry, but also seriously damages the health of consumers. It is of great significance to find a faster and more accurate method for detecting milk protein content. This paper utilizes the chemical content of milk powder and hyperspectral data as independent variables. By comparing 14 kinds of preprocessing algorithms, the mean-centered (MC) method is selected to preprocess the data, and then the combined method of competitive adaptive reweighted sampling (CARS) and uninformative variable elimination (UVE) is used to screen the feature wavelength, so as to establish the model and learn the internal dynamic change law of the feature. Furthermore, the Attention mechanism was introduced to assign different weights to bidirectional long short-term memory (BiLSTM) hidden states through mapping weighting and learning parameter matrix. To reduce the loss of information and strengthen the influence of important information, at the same time, in order to solve the difficult problem of hyperparameter selection of the model, the whale optimization algorithm (WOA) is proposed to optimize the hyperparameter selection of the model. The test results showed that with WOA-BiLSTM-Attention model algorithm, the coefficient of determination (<i>R</i> ²) of 0.9975 and root mean square error (RMSEP) of 0.0337 in comparison with <i>R</i> ² and RMSEP values obtained from BiLSTM-Attention model algorithm, which were higher by 0.799% lower by 56.5%, respectively. This study provides algorithm support and theoretical basis for fast non-destructive testing based on deep learning algorithm to predict protein content in milk powder.