Transient sound signal analysis for watermelon ripeness detection using HHT and NMF

Abstract

Harvesting watermelon at an inappropriate time can significantly impact its quality and flavor. To ensure rapid, reliable, and nondestructive determination of watermelon ripeness, this study focuses on analyzing the tapping sound of watermelons at various ripeness levels. The tapping sound, characterized as a transient acoustic signal, exhibits consistent resonance properties but varying frequency features across ripeness stages. A sound processing method was developed by integrating Nonnegative Matrix Factorization (NMF) filtering and Root-Mean-Square (RMS) normalization. Frequency characteristics and variations in watermelon tapping sounds were analyzed using the Hilbert-Huang Transform (HHT) and NMF-based feature extraction. Machine learning models, including Support Vector Machine (SVM), HHT combined with SVM (HHT + SVM), and NMF combined with SVM (NMF + SVM), were employed to classify watermelons of different ripeness levels. Experimental results, based on 100 samples each of unripe, ripe, and overripe watermelons, showed a gradual decrease in the average frequency distribution of tapping sounds from unripe to overripe stages. The classification accuracy of watermelon ripeness using SVM alone was 62.78 %, which improved to 74.44 % with HHT + SVM and further increased to 92.22 % with NMF + SVM. These findings demonstrate that feature extraction methods based on NMF and HHT effectively capture the frequency characteristics and time-decay properties of transient acoustic signals. This study offers an efficient and practical method for acoustic nondestructive detection of watermelon ripeness, providing a novel approach for processing transient and abrupt sound signals with broad potential applications.