A two-layer ensemble learning framework for data-driven soft sensor of the diesel attributes in an industrial hydrocracking process

In the hydrocracking process, it is of great significance to timely measure the product attributes for real-time process control and optimization. However, they are often very difficult to measure online due to technical and economical limitations. To this end, soft sensor is introduced to predict product attributes through easy-to-measure process variables, with the advantages of low cost, fast response, and ease of maintenance. In this paper, a two-layer ensemble learning framework is developed for soft sensing of three diesel attributes in an industrial hydrocracking process. In this modeling framework, the process variables are first divided into subspace blocks according to process topological structure to capture the local behaviors of different production cells. Then, to overcome the weak generalization ability of a single calibration model with specific hypothesis, different regression learners are constructed on each variable subblock to increase the model diversity. At last, individual models are fused to improve the prediction performance and generalization ability of soft sensor models. The effectiveness and flexibility of the proposed ensemble learning method is validated on a real industrial hydrocracking process.