Robust data-driven soft sensor based on iteratively weighted least squares support vector regression optimized by the cuckoo optimization algorithm

• A robust data-driven LSSVR-based soft sensor for debutanizer is designed.
• We optimized the soft sensor model by Cuckoo Optimization Algorithm (COA).
• Iteratively weighted LSSVR improves the robustness of the LSSVR-based soft sensors.
• The IWLSSVR-based soft sensor is more robust than robust ANFIS-based soft sensor.

In process industries, use of the data-driven soft sensors for the purpose of process control and monitoring has gained much popularity. Data-driven soft sensors infer the process quality variables from the available historical process data. A considerable amount of process data such as pressures, temperatures, etc., are measured routinely and stored permanently. However, the quality of these data often varies. Measurement noises and data outliers are the most common effects which lead to poor quality of process data. Application of standard statistical techniques to operate data may lead to model deterioration due to contaminating observations. Therefore, the objective of this paper is to present a robust approach for the development of data-driven soft sensors. In this paper, the modeling method that is used to develop soft sensor is a combination of Nonlinear Auto Regressive with eXogenous inputs (NARX) structure with Least Squares Support Vector Regression (LSSVR). The LSSVRs' parameters are optimized by a new evolutionary optimization technique known as the Cuckoo Optimization Algorithm (COA). Then in order to make the soft sensor robust against the data outliers and noises especially the long tail noises, a new approach is proposed. The proposed method is based on the Iteratively Weighted LSSVR (IWLSSVR) which uses the Myriad weighting function. The proposed approach was applied to the prediction of the n-butane (C4) concentration in a debutanizer column unit. The technique was consequently compared against the conventional LSSVR algorithm which is based on the quadratic loss function. It turns out that reweighting the LSSVR estimate using the Myriad weight function improves the performance of the LSSVR-based soft sensor when noises and outliers exist in the measured data. The designed robust soft sensor is also compared with another robust soft sensor which is recently developed based on the Adaptive Neuro-Fuzzy Inference System (ANFIS) optimized by the Particle Swarm Optimization (PSO). The simulation results show that the designed IWLSSVR-based soft sensor is more robust when the measured data have some impurities.