Multi-sensor fusion combined with machine learning enables real-time freshness prediction of ‘Korla’ pear during storage

This study proposes an innovative hybrid sensor machine learning system for the objective and automated assessment of pear freshness, overcoming the limitations of traditional visual inspection methods in pear storage monitoring. The key control points in post-harvest handling were identified using hazard analysis and critical control points. A real-time monitoring system based on the Modbus-RTU protocol and Winform architecture was developed to obtain multi-source environmental data under storage conditions of 0 °C, 4 °C and 25 °C. Physicochemical indices were integrated with sensor data to construct predictive models using backpropagation neural network (BPNN), support vector machine (SVM), random forest (RF) and radial basis function (RBF) network algorithms. The results indicated strong correlations between microenvironment parameters and freshness indicators. Among the models, SVM demonstrated the best performance, with prediction accuracies of 96.67 % for firmness and 94.30 % for soluble solids content. It considerably outperformed the traditional Arrhenius equation method, which achieved an accuracy of 88.23 %. The monitoring system demonstrated high reliability, with data acquisition accuracy exceeding 99 % and robust operational stability. This study provides a non-destructive and efficient solution for pear storage management, effectively reducing the risk of quality deterioration and enhancing economic benefits.