Development of freshness monitoring models for grass carp cutting products at multiple storage temperatures and construction of a real-time prediction platform

Spoilage of aquatic products cause economic losses and health risks, highlighting the need for real-time freshness monitoring technologies. In this study, four value-added grass carp cuts (dorsal, abdomen, belly, and tail) were stored at five temperatures (0, 3, 6, 9, and 12 °C) for 0-10 days. Five freshness indicators were measured, and six machine learning models were applied for prediction. The best-performing model for each indicator was ensembled into a real-time prediction platform. Results showed that the optimal models for predicting TVC, TVB-N, K value, overall sensory acceptability, and sensory evaluation score were radial basis function neural network (RBFNN), backpropagation neural network, gradient boosting machine, RBFNN, and support vector regression. These models achieved high prediction accuracy, with R 2 values ranging from 0.9552 to 0.9734. TVC modeled by RBFNN yielded the best validation performance with relative errors ranging from 0.00 to 0.21. The real-time platform completed freshness prediction within 0.06 s and classified samples into "Fresh", "Less Fresh" or "Spoiled" levels. Overall, this study proposed an innovative multi-cut, multi-indicator, and multi-temperature modeling strategy. Combined with the practical implementation of an intelligent digital platform, it provides a robust, non-destructive, and data-driven tool for quality evaluation and management in the aquatic supply chain. • Freshness changes of grass carp cuts were monitored under multiple temperatures. • TVC, TVB-N, K value, and sensory indicators were used to assess freshness. • Six machine learning models were constructed and compared to predict freshness. • A real-time platform was ensembled for freshness and shelf life prediction.