The first application of improved empirical mode decomposition in vibro-acoustic detecting for sub-healthy pears with core browning

The pear with browning not spreading over the core line is called sub-healthy fruit, which maintains normal commodity value but has the risk of flesh disorder. Timely identification of the sub-healthy fruit could facilitate control measures of storage and marketing. When the vibro-acoustic approach is employed for this detecting task, the signals of sub-healthy pears are little different from those of healthy and slightly diseased pears. Therefore, accurately identifying sub-healthy pears is rather challenging. With our improved empirical mode decomposition (EMD) algorithm, the pear signal was decomposed into a series of components, from which the characteristic components were selected to reconstruct the signal. Then, we have used the vision transformer (ViT) to extract depth features from images of the reconstructed signals converted by five visual encoding methods and selected the features with better clustering ability to construct deep-shallow detection models. For the neighboring categories signals in similar waveforms, the top two importance components can be used as characteristic components to reconstruct the signal with high discrimination. The extracted deep features from the recurrence plot images of the reconstructed signals have the best clustering ability, and they are input to the logistic regression (LR) classifier to construct the ViT-LR model. This model has achieved the best classification performance with an accuracy of 93.30 %, F 1 score of 93.25 %, Kappa coefficient of 90.92 %, and Matthews correlation coefficient of 89.97 %; the model still has good generalization performance through external validation. This study confirms that the signal decomposition method can effectively increase the accuracy of pear core browning detection, and has the potential to be used for early or slight internal disease detection of other fruits and vegetables.