From vibration to firmness assessment: Transfer-function modelling of intact kiwifruit utilizing machine learning

• A vibration-based, non-destructive system for fruit firmness prediction was developed • Integrated dynamic feature extraction with Neighborhood Component Analysis (NCA) • Identified rise time, overshoot, and damping coefficient as key predictive features • Optimized regression model achieved R² = 0.981 with near-zero bias and error • Outperformed neural networks, piezoelectric, and tactile-based prediction systems • Enables real-time post-harvest quality control, shelf-life optimization, and waste reduction Accurate, non-destructive assessment of fruit firmness is critical for post-harvest quality management and reducing food waste. This study presents a novel vibration-based predictive framework for kiwifruit firmness that integrates dynamic feature extraction with Neighborhood Component Analysis (NCA) for regression-based feature selection. Kiwifruit samples were subjected to 5-second vibrations, during which acceleration and angular velocity were recorded in three orthogonal directions. Analysis identified rise time (weight = 0.633), overshoot (weight = 0.474), and damping coefficient (weight = 0.360) as the most informative features for predictive modeling. These features were used to construct a regression model, optimized with a lambda parameter (λ = 0.040) to minimize the loss function. The model achieved RMSE = 0.0079, MSE = 6.35 × 10⁻⁵, and R² = 0.981, with a regression slope near 0.99 and a near-zero bias (∼0.00), demonstrating high fidelity between predicted and actual firmness values. Statistical evaluation using paired t-tests confirmed no significant difference between measured and predicted values (p > 0.05), validating the approach. The proposed method advances non-destructive firmness prediction by combining dynamic vibration analysis with feature-focused regression, providing improved precision and reduced systematic error compared to previous approaches such as neural networks, piezoelectric sensors, and tactile-based systems. Its practical applicability includes real-time post-harvest firmness monitoring, quality control, shelf-life optimization, and waste reduction. The integration of vibration-based feature extraction with NCA regression establishes a highly accurate, efficient, and transferable methodology that can be extended to other horticultural crops, supporting precision agriculture and automated quality monitoring systems.