Integration of machine learning algorithms and empirical formula-driven data augmentation for freshness prediction of bighead carp cutting products

Traditional chemical methods for evaluating fish freshness are time-consuming and destructive, underscoring the need for rapid and non-invasive predictive models. In this study, random forest (RF), backpropagation neural network (BPNN), and long short-term memory (LSTM) models were employed to predict muscle freshness of four bighead carp cuts during storage at different temperatures (0, 3, 6, 12, 24℃). A total of 108 data points were collected for TVC, TVB-N, K-value, and sensory evaluation indicators. BPNN outperformed both RF and LSTM in TVC, TVB-N and sensory evaluation, with R 2 in testing sets being 0.8857 0.9998, and 0.9312, respectively. Data augmentation using exponential decay function (EDF) and Arrhenius function (AF) improved performance for all models. EDF-LSTM was the best for predicting TVC in eye muscle, with an average validation error of 19.22%. AF-RF provided the best predictions for K-value in dorsal, belly, and tail muscles, with errors of 17.27%, 15.93%, and 15.53%. EDF-RF and AF-LSTM were optimal for sensory evaluation in eye and dorsal muscles, with errors of 11.01% and 21.52%. These findings demonstrate that integrating machine learning with data augmentation offers a promising approach for non-destructive freshness prediction in fish cuts across a range of storage temperatures. • Predicting freshness of four bighead carp cuts using machine learning models. • Collected datasets involving TVC, TVBN, K-value, and sensory evaluation indicators. • TVC was the most reliable input dataset for model to predict freshness. • The average relative error for TVC prediction of TM using RF was less than 5%.