A Knowledge- and Data-Driven Soft Sensor Based on Deep Learning for Predicting the Deformation of an Air Preheater Rotor

In industrial processes, some important process variables cannot be measured directly by hardware sensors for technical or economic reasons. Soft sensors estimate these key variables using some other easily measured variables by building a mathematical model. A novel knowledge- and data-driven soft sensor is proposed in this paper to predict the deformation of an air preheater rotor in a thermal power plant boiler. Two submodels were constructed, including the knowledge-driven submodel, derived from all the available domain knowledge, and the data-driven submodel, constructed solely from the data. The two submodels were integrated with a mass balance model. A mathematical model based on technical expertise in predicting rotor deformation, named the Lab model, was used as the knowledge-driven submodel, and a deep learning model based on stacked autoencoders (SAE) was used as the data-driven submodel. To improve the performance of the model, the limited-memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) algorithm was adopted to optimize the SAE parameters. The experimental results demonstrate that, compared with the common knowledge-driven (KDM) and data-driven (DDM) models, the proposed Lab-stacked autoencoders (L-SAE) model is able to provide a higher predictive accuracy for the air preheater rotor deformation and inherits the advantages of both the KDM and DDM.