Real-time and explainable non-destructive nut classification using spike-triggered acoustic sensing

Rapid, non-destructive, and low-cost quality assessment methods are required for real-time nut sorting in automated production environments. Traditional inspection techniques, typically visual, manual, or destructive, are often time-consuming, costly, and difficult to integrate into automated production lines. To address these challenges, we propose an online classification system for walnut samples from Chandler and multiple Lara groups (e.g., distinct lots/grades), jointly predicting caliber and void by leveraging spike-triggered acoustic sensing and spectro-temporal signal descriptors. The proposed solution is based on a low-cost mechanical structure in which individual walnuts slide through a guided channel and impact a metallic plate, generating distinct acoustic signatures. A spike-detection strategy is applied to the continuous audio stream to isolate high-energy impact events, allowing for consistent and non-invasive acquisition of time-aligned segments. From these segments, a comprehensive set of spectral and perceptual features is extracted, including Fast Fourier Transform (FFT) energy bands and Mel-Frequency Cepstral Coefficients (MFCCs). These features are then used to train a Random Forest classifier offline, which is later deployed for real-time prediction of nut quality classes. Experimental results show strong classification performance, achieving an F1-score of 0.93 when distinguishing between Lara void and Lara caliber samples, assessing walnut quality. The system is well-suited for deployment both in real-world industrial settings and directly on-site in agricultural fields, addressing the growing demand for intelligent, non-invasive food quality monitoring technologies across the production chain.