Multi-sensor fusion with optimized machine learning for non-destructive freshness monitoring of stored Korla fragrant pears

Traditional freshness assessment methods for Korla fragrant pears are limited by inefficiency, high cost, and reliance on subjective or destructive evaluations. This study proposes a non-destructive, multi-sensor fusion approach that integrates gas composition (O₂, CO₂, VOC), environmental parameters (temperature, humidity, PM_2.5, pressure), and dielectric properties (C, D, ε). Data were analyzed using machine learning models (BPNN, SVM, RF), optimized via genetic algorithms (GA) and particle swarm optimization (PSO). A dielectric-based grading standard (L₁–L₄) was also developed. Dielectric parameters correlated strongly with key freshness indicators (r = 0.86 for firmness; r = 0.88 for SSC). The PSO-SVM model with a gaussian kernel achieved 97.50 % accuracy and a 97.49 % F1-score, with 0.54 % deviation. Multi-source fusion significantly outperformed single-sensor models, as shown by the low accuracy (47.12 % ± 2.35 %) of gas-only data. A NET-based monitoring platform demonstrated 97.74 % data acquisition accuracy. These findings highlight a low-cost solution for real-time freshness monitoring, offering substantial improvements over traditional methods and supporting intelligent quality control in fruit storage.