Real-time freeze point prediction using multirate measurements in the blending process

In blending processes, real-time monitoring of product properties is crucial for maintaining quality and optimizing operational efficiency. However, properties such as the freeze point are typically measured using slow and expensive laboratory tests. To enable real-time monitoring, analyzers are developed based on these laboratory measurements. Additionally, there are certain compounds whose freeze point is less than $-70{}^{\circ }\text{C}$, which are beyond the detection limits of conventional laboratory techniques. This paper introduces a framework that combines the expectation–maximization algorithm with particle-filtering to estimate the freeze point of a compound used in the fuel-blending process, where conventional laboratory methods struggle to provide measurements. The method integrates multirate data, by combining high-frequency sensor data with low-frequency laboratory measurements, to estimate the freeze point. The soft sensor parameters are then identified using the estimated freeze point and directly measured input features such as the true boiling point. The proposed model allows estimation of the freeze point, particularly for components whose properties are not readily measurable using standard laboratory techniques. The proposed approach is compared against two other approaches: (1) a estimation using only high-frequency sensor data and (2) a estimation using only slow laboratory measurements. The soft sensor developed using the proposed framework reduces dependence on offline testing, providing a cost-effective and operationally viable alternative, while validation with industrial data confirms its applicability and effectiveness in real time, achieving an $R^2$ value of 0.4074 that demonstrates reasonable predictive performance under industrial conditions.