Detection of moldy pear core based on the time-frequency analysis of acoustic vibration signals and multi-domain features fusion

Abstract

Mold core, as a serious internal defect, greatly affects fruit quality and the development of the pear industry. Due to its high contagions, it is thus desirable to implement early detection for moldy-core pears and / or remove pears with moldy core during sorting and grading. This study utilized the acoustic vibration non-destructive detection system to collect the acoustic vibration response signals. The acquired acoustic vibration response signals were converted the time-frequency images by the Short-time Fourier Transform (STFT) algorithm. The 14 time-domain statistical features T₁∼T₁₄ were extracted from the original signal curves by time-domain analysis method. The 7 frequency-domain statistical features F₁∼F₇ were extracted by frequency-domain analysis methods. The 15 texture features G₁∼G₁₅ were extracted from the STFT time-frequency images using gray level gradient correlation matrix (GLGCM) algorithm. Then, the Pearson correlation analysis was used to select the multi-domain sensitive features for discriminating pears with different moldy-core degrees. For the selected multi-domain sensitive features, the principal component analysis (PCA) was employed to convert the original high-dimensional dataset into a low dimensional representation. Finally, the processed single-domain and multi-domain fusion features were employed as inputs to construct the three classification models for identifying moldy pear core. The three classification models included partial least squares discriminant analysis (PLS-DA), least squares-support vector machine (LS-SVM), and extreme learning machine (ELM). The results indicated that the classification models constructed by the fused multi-domain features exhibited the higher discrimination accuracy for the three-categories pears. Among the constructed models, the ELM model achieved the optimal identification performance with an overall identification accuracy of 98.67 %. Specifically, the ELM model reached a 100 % classification accuracy for both healthy pears and pears with significant moldy-core (≥ 10 %), and a 96.15 % accuracy for pears with slight moldy-core (< 10 %). The overall accuracy, Recall, Precision, F₁, and Kappa coefficient of the constructed ELM model were up to 92 % in the external validation. Thus, the proposed method has excellent recognition capability for pears with different extents of moldy core and outperforms some traditional techniques both mentioned in this study and reported in other research.