Research Of Soft Sensor Modeling Methods For Ultrasonic Chinese Medicine Extraction

Ultrasonic extraction technology is an important modern traditional Chinese medicine component extraction method.The existing technology cannot meet the online detection requirements of the extraction rate during the extraction process.In order to achieve online detection of the extraction rate during the process of dual-frequency ultrasonic extraction of puerarin,this work uses soft sensor modeling methods to establish soft sensors between secondary variables and the extraction rate.As a batch process,when there are obviously differences between different batches of the process,the soft sensor model established by traditional methods will show obvious performance degradation.To solve this problem,this paper uses a model adaptive method to improve the performance when the process changes from batch to batch,thereby extending the maintenance cycle of the model and reducing costs.Firstly,for the ultrasonic extraction process of traditional Chinese medicine with dual-frequency ultrasound,this study designed and conducted experiments on ultrasonic extraction of puerarin.Under different initial temperatures of the extract,experimental data from different batches were collected and labeled for establishing soft sensor models and verifying models performance.Secondly,this paper proposes an adaptive method based on just-in-time learning(JITL)for soft sensor modeling when batch changes occur in the ultrasonic extraction process of traditional Chinese medicine with dualfrequency ultrasound.This method uses a similarity index based on Euclidean distance to calculate the correlation between current queries and samples in the historical dataset.By selecting the most relevant dataset,local models are established online to achieve adaptive updates of models.This paper uses support vector regression(SVR)as the local model to establish a JITL-SVR soft sensor model,which improves model performance and solves the problem of model performance degradation caused by batch changes in ultrasonic extraction processes.Then,although JITL models improves model performance,it will bring some lag to predictions and reduce the real-time performance of models.To solve this problem,this paper proposes an adaptive method based on multi-head attention mechanism.This method uses multi-head attention mechanism to calculate the similarity between current queries and historical data in different feature spaces.By attention pooling,different weights are assigned to historical data according to their relevance with current queries,so that samples that are more similar to current queries have a greater impact on model predictions.This paper combines SVR with this method to establish an MHA-SVR soft sensor model,which can not only improve model performance when batch changes occur but also has better real-time performance than JITL and is more in line with online detection requirements for extraction rates.Finally,through simulation experiments,this paper compared and analyzed the performance indicators of two soft sensor models proposed in this paper with traditional SVR models and verified the effectiveness of this paper’s methods.