Plant-based meat extrusion is a complex multi-stage process involving dynamic interactions among raw materials, operational parameters, and resulting product structure. Maintaining consistent product quality requires precise control over extrusion conditions, which remains challenging given the process's nonlinear behavior. Conventional physics-based models often struggle to accurately capture these complexities, while existing structure evaluation methods are labor-intensive and costly. Moreover, the lack of real-time feedback mechanisms hinders adaptive process control, limits the implementation of automated quality assurance, and contributes to product variability. To address these limitations, this study presents unsupervised and semi-supervised computer vision (CV) algorithms for automated texture assessment, integrated with machine learning-based extrusion models. These models capture the relationships between operational parameters and structural attributes, such as fibrousness and porosity. The proposed semi-supervised CV approach demonstrated significantly higher accuracy and a lower structure-scoring error than the unsupervised method. In integrated time series modeling, both XGBoost and LightGBM effectively captured the dynamics of the structure score, die temperature, and specific mechanical energy, with LightGBM performing slightly better across most metrics. By establishing dynamic links between process conditions and product structure, the proposed framework enables real-time monitoring and optimization of plant-based meat extrusion, enhancing process control and product consistency. • LSTM and GRU models capture high-frequency extrusion dynamics for optimization. • Unsupervised and semi-supervised computer vision improve structure quantification. • Integrated structure scores with process data for dynamic structure evolution prediction. • XGBoost and LightGBM facilitate structure prediction despite small dataset constraints.
Bagheri et al. (Wed,) studied this question.