MHA-SVR: An Adaptive Soft Sensor Based on Feature Interaction for Ultrasonic Phytomedicine Extraction.

Introduction: Ultrasonic extraction is a crucial technique for isolating active compounds from phytomedicine. However, as a batch process characterized by non-linearity and small sample size, it poses substantial challenges for real-time and prediction of extraction rates during the extraction of phytomedicinal. This work proposes an adaptive soft sensor for ultrasonic phytomedicine extraction.

Method: An adaptive soft sensor based on an attention mechanism was first proposed. The attention mechanism calculates correlations between samples and assigns weights based on their similarity to the current query. Support vector regression (SVR) is then used to construct the soft sensor for extraction rate measurement. To further enhance sample information analysis, multi-head attention is employed. This allows the model to calculate the similarity between current queries and historical data across different feature spaces, thus improving the modeling capabilities of the intrinsic data structure. Finally, a dual-frequency ultrasonic extraction experiment of puerarin is designed and conducted. The experimental data is collected and labeled from different batches under varying initial extraction temperatures. This data is used to establish the soft sensor model and compare its performance.

Results and discussion: The experimental results indicate that the proposed MHA-SVR model improved the coefficient of determination (R²) by 5.12% compared to the mainstream model and reduced the online prediction time by 88% compared to the JITL-SVR model. This work performance well exceeds the others while maintaining good real-time capabilities for the dual-frequency ultrasonic extraction of puerarin.

Conclusion: The multi-head attention and SVR-integrated soft sensor method proposed in this study effectively addresses the soft measurement challenges in online monitoring of multi-batch ultrasonic extraction processes. This approach demonstrates significant enhancement in extraction yield detection accuracy across varying batches and diverse initial operating conditions, thereby providing a robust technical solution for real-time quantification of extraction efficiency in botanical material processing.