Methodology for quality risk prediction for milk powder production plants with domain-knowledge-involved serial neural networks

Abstract

In dairy enterprises, predicting product quality attributes that are influenced by operating parameters is a major task. To reduce quality loss in production, a prediction-based quality control method is proposed in this study. In particular, a serial neural network was designed, and an innovative quality risk prediction methodology based on the integration of SNN and domain knowledge was created. The methodology involves three steps: (1) the processing steps at each unit operation are mapped to a layer of a back propagation network, (2) the branch networks are connected by key quality attributes, and (3) the model is trained with preprocessed data. The experiment was conducted based on milk powder production, demonstrating that the proposed methodology has a higher accuracy and shorter response time compared with those of existing methods. In addition, the practical value of the prediction methodology in actual dairy companies was discussed.