In chemical plants, soft sensors are used to estimate values of process variables that are difficult to measure in real time. Values estimated by soft sensors are employed to control plants rapidly. A crucial problem is degradation of predictive ability of soft sensor models. Due to changes of process states in plants such as catalyst performance loss, changes of raw materials and fouling on pipes, errors of soft sensors increase.
To reduce the model degradation, several types of adaptive soft sensors have been developed over many years. We previously categorized the types of the model degradation and those of adaptive soft sensors, and discuss the characteristics of adaptive soft sensors to use of the right type of adaptive soft sensors for the right type of the model degradation. The three main approaches to constructing adaptive soft sensors are the moving window (MW), just-in-time (JIT), and time difference (TD) approaches. As no all-round adaptive soft sensors exist, the selective use of TD, MW and JIT models could enable the predictive ability of soft sensors to increase dramatically. However, to improve prediction accuracy of each type of adaptive soft sensors contribute the improvement of overall performance in soft sensors.
We focus JIT models in this study. Although many researches exist on similarity such as distance and correlation and ensemble learning in JIT modelling, predictive ability of JIT models depends heavily on the quality of a database. Even when data in X-variables are similar in JIT modelling, data in a y-variable are not always similar due to process changes such as catalyst performance loss and sensor and process drift, which has harmful effects to JIT models.
Data measured in similar time would have the same relationship between X and y, but old data would have a different relationship between X and y even when data in X are similar since there happen process changes such as catalyst performance loss and sensor and process drift. Meanwhile, a model constructed with only recent data cannot adapt to rapid process changes when recent data have little variation.
We therefore propose to prepare multiple sub-datasets and construct a JIT model for each sub-dataset. Locally-weighted partial least squares (LWPLS) is focused in JIT modelling in this study. By constructing LWPLS models with different sub-datasets, the models would adapt to various process states and transition between them. In addition, contamination of data in which relationships between X and y are different can be prevented for each sub-dataset by separating old data and recent data. A final predicted y-value is calculated with weighted average of y-values predicted by multiple LWPLS models. The weights are given considering predictive ability of LWPLS models. The points are the way to divide sub-datasets and a final y-value estimated using multiple LWPLS models
Through case studies using simulated and real industrial datasets, it was confirmed that the use of the proposed method enables soft sensors to operate with high predictive accuracy.
See more of this Group/Topical: Computing and Systems Technology Division