Prediction Method of Tangerine Peel Drying Moisture Ratio Based on KAN-BiLSTM and Multimodal Feature Fusion
Tangerine peel, rich in moisture (75–90%) and medicinal value, requires drying to prevent spoilage and extend shelf life. Traditional heat pump drying often causes uneven airflow, leading to inconsistent drying and nutrient loss, compromising product quality and storage stability. In this study, a p...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2025-05-01
|
| Series: | Applied Sciences |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2076-3417/15/11/6130 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Tangerine peel, rich in moisture (75–90%) and medicinal value, requires drying to prevent spoilage and extend shelf life. Traditional heat pump drying often causes uneven airflow, leading to inconsistent drying and nutrient loss, compromising product quality and storage stability. In this study, a prediction model of drying moisture ratio of tangerine peel based on Kolmogorov–Arnold network bidirectional long short-term memory (KAN-BiLSTM) and multimodal feature fusion is proposed. A pre-trained visual geometry group U-shaped network (VGG-UNet) is employed to segment tangerine peel images and extract color, contour, and texture features, while airflow distribution is simulated using finite element analysis (FEA) to obtain spatial location information. These multimodal features are fused and input into a KAN-BiLSTM model, where the KAN layer enhances nonlinear feature representation and a multi-head attention (MHA) mechanism highlights critical temporal and spatial features to improve prediction accuracy. Experimental validation was conducted on a dataset comprising 432 tangerine peel samples collected across six drying batches over a 480 min period, with image acquisition and mass measurement performed every 20 min. The results showed that the pre-trained VGG-UNet achieved a mean intersection over union (MIoU) of 93.58%, outperforming the untrained model by 9.41%. Incorporating spatial features improved the coefficient of determination (R<sup>2</sup>) of the time series model by 0.08 ± 0.04. The proposed KAN-BiLSTM model achieved a mean absolute error (MAE) of 0.024 and R<sup>2</sup> of 0.9908, significantly surpassing baseline models such as BiLSTM (R<sup>2</sup> = 0.9049, MAE = 0.0476) and LSTM (R<sup>2</sup> = 0.8306, MAE = 0.0766), demonstrating superior performance in moisture ratio prediction. |
|---|---|
| ISSN: | 2076-3417 |