Analysis of the effect of irrelevant information for the design of soft sensors

Abstract Modern industrial production requires real-time control and monitoring of critical quality variables that are hard to measure, which leads to the development of soft sensors that infer the quality variables online. In the big-data era, massive data are collected from industrial processes, which might contain irrelevant information. In addition, large models, especially large deep neural networks, might contain redundant substructures that could also create irrelevant information. The negative effect of irrelevant information on soft sensors is rigorously investigated in this paper, which motivates the development of a new soft-sensing method based on information-bottleneck (IB) theory. This new method combines decoupled representation learning, irrelevant information elimination, and soft-sensor modelling into one model that can be optimised as a whole using a single cost function. The new soft-sensor model is in the form of a trade-off between model complexity and training accuracy, which results in structural risk minimisation and mitigates overfitting. Compressing the model complexity by removing irrelevant information increases the model performance of quality-variable inference on the testing data set. Using a case study on the sulphur-recovery unit, soft sensors were built with and without the proposed IB module. It is shown that adding the IB module improves R 2 from 55.6 % to 80.9 % and decreases the MSE from 0.039 to 0.026.