Acoustic-vibration detection of moldy pear core: a novel approach combining image coding and hybrid deep learning

Abstract

The common acoustic-vibration detection technology for pears with moldy core had weak adaptive extraction ability and low accuracy in distinguishing early moldy pear cores. To address this issue, this study proposed an "single-point excitation and dual-point sensing" acoustic-vibration method combined with the image coding and hybrid deep learning model. The self-designed acoustic-vibration testing system was used to obtain the acoustic-vibration signals of pears. The acoustic-vibration signals were transformed into the feature images using the improved image encoding methods. Subsequently, the DSC-SqueezeNet feature extractor improved by the deep separable convolution (DSC) and the vision transformation (ViT) feature extractor was used to adaptively extract the deep features related to the moldy pear core. Then, the t-SNE method was used to perform the qualitative and quantitative clustering analysis on the deep features extracted by the two feature extractors. The quantitative analysis results showed that the deep features extracted by DSC-SqueezeNet from multi-scale asymmetric recurrence plots (MARP) feature images outperformed those extracted by ViT in terms of inter-class separation (DB = 0.1981) and intra-class compactness (CH = 4409.7842), and were more effective in characterizing the detailed information in the response signal images of the three pear categories. Finally, the deep features extracted from the optimal feature images were input into the two classifiers: multi-layer perceptron (MLP) and extreme gradient boosting (XGBoost). The hyperparameters of MLP and XGBoost classifiers were optimized by the Optuna algorithm to construct the DS-OPMLP model and DS-OPXG model for identifying the pears with moldy core. The classification results indicated that the DS-OPXG discrimination model established by the multi-scale asymmetric recurrence plots (MARP) as input has the higher classification accuracy of 95.83 % on the test set, which was 5.56 percentage points higher than that of the DS-OPMLP model. The sensitivity and specificity for the mildly moldy pear core were 96.15 % and 95.65 %, respectively. Also, the ablation experiments were executed to verify the effectiveness of the fusion of DSC and XGBoost. The DS-OPXG model only distinguished one mild moldy core pear as significant moldy core pear for the external validation set. The classification accuracy was 90.48 % for healthy pears, 97.44 % for the pears with mild moldy core and 86.49 % for the pears with significant moldy core. Therefore, the DS-OPXG model constructed by the MARP and hybrid deep learning can effectively capture the detail information of the acoustic-vibration signal and further improve the discrimination accuracy for the pears with the mild moldy core. This holds great significance for the practical application of acoustic-vibration detection of internal diseases in pears.