Soft Sensor Modeling For Multi-rate Data Based On Stacked Autoencoder In The Hydrocracking Process

Hydrocracking is an important production process in the refining petrochemical industry.Product quality information is an important foundation for process monitoring,control and optimization in hydrocracking.The quality variables are usually measured by offline methods with large lag,while soft sensor technology provides a practical method for real-time prediction of quality variables by building a model between process variables and quality variables.Hydrocracking process contains multiple products,and each product has multiple quality variables,which also have different sampling characteristics.Most of the traditional soft sensor models only use the labeled samples to predict a single quality variable,which may cause problems like waste of unlabeled samples,excessive number of models and total parameters,heavy equipment burden,and cannot model the relationship between different quality variables hindering the implementation of soft sensor in production.With its outstanding performance in unsupervised feature extraction,large-scale data processing,scalability,stacked autoencoder provides an important basic model for effectively solving the above problems.Hence,this paper taking the data in hydrocracking process with multiple sampling rates as research object,studies the modeling of multi-rate data based on stacked autoencoder.The main innovative research work of this paper is as follows:(1)Aiming at the waste of numerous unlabeled samples information caused by unequal-rate input/output data in hydrocracking process,a multiple-input single-output soft sensor based on stacked attention autoencoder is proposed.Through semi-supervised strategy,unlabeled and labeled data are used for pre-training and construction of soft sensor model.And the attention module is designed to extract quality related features for soft sensor accurate modeling.Simulation results verify that the proposed method can effectively improve the prediction accuracy of single quality variables.(2)Aiming at the calculation invalid of back propagation and model bias caused by multi-rate quality variables in hydrocracking process,a multiple-input multiple-output soft sensor based on multi-rate stacked autoencoder is proposed.The method innovatively designs the network structure as shared network and quality-specific network,which are used to extract the common and characteristic features of multi-rate quality variables respectively.Thus,the correlation relationship between multirate quality variables is modeled,and the model training is completed through pre-training,global fine-tuning and sub-network individual finetuning.Simulation results verify that the proposed method can reduce the total amount of model parameters and improve the prediction performance.(3)Aiming at the difficulty of model integration and learning caused by irregular-sampling quality variables in hydrocracking process,a multiple-input multiple-output soft sensor based on irregular-rate stacked autoencoder is proposed.The method completes model training through three steps of pre-training the base network,generating the sub-networks’ mask matrixes and alternately fine-tuning the sub-networks in parallel,which realizes the integration of irregular-sampling quality variables and meanwhile extracts more fine-grained common and characteristic features.Simulation results verify the effectiveness of the proposed method for irregular-sampling quality variables modeling.Figure 30,Table 13,Reference 85.