(183c) Mutual Information-Based Feature Selection for Enhanced Soft Sensor Performance in Petrochemical Processes

The safe and profitable operation of any process operation is highly relied on the real time information from the process, which is obtained from huge number of sensors across the facility. Due to the special situation of chemical process, sensors are facing challenges from all kinds of unfriendly working environments, such as high temperature and high corrosive medium, making the online measurements subject to hardware failure and malfunction, even impossible at certain extreme situation. Soft-sensor technique is therefor applied to either cross-validate the measurements, or provide an alternative to facilitate the operation.

The idea of soft-sensoring is to extract the correlation among measurements, mapping one group to another. The performance of soft-sensor not only depends on the mapping algorithm, but also the measurements for modeling. Chemical process measurements are abundant and often redundant, which make the soft-sensoring both feasible and challenging. The choice of input measurements is crucial. For this purpose, methods such as normalized Mutual Information (MI) and Maximal Information Coefficient (MIC) have been developed, along with the application of the Maximum Relevance and Minimum Redundancy (mRMR) algorithm for variable selection. However, existing methods still face limitations such as inadequate handling of common redundancy and inefficient ranking of variable importance.

In this work, a mutual information-based relevance-redundancy algorithm is introduced for feature selection, in which the relevance and redundancy among process measurements are evaluated through a comprehensive correlation function and ranked according to their importance using greedy search to obtain the optimal measurement set. The algorithm has been validated in both bench mark process and a delayed coking process at a petrochemical enterprise using classical machine learning algorithms to estimate the outlet temperature of the cracking furnace online. The results demonstrate significant reductions in both mean absolute error and root mean square error compared to other methods, confirming the importance of choice of feature selection in soft-sensoring performance, other than the contribution of algorithms. The findings suggest that this approach holds great promise for improving the reliability and efficiency of petrochemical production processes.