Industrial processes must be well equipped with a variety of sensors to maintain a desired quality. However, some variables cannot be easily measured due to different causes, such as acquisition and/or maintenance costs and slow acquisition time. This situation leads to a lack of real-time information in the process, which could lead to lower quality in the final product. One of such processes is the debutanizer column, where butane content measurement is highly delayed. To enable online prediction of such variables, available information from the process can be used to estimate predictive models, known as soft sensors. To this end, data-driven techniques can be used, such as statistical and machine learning. However, such techniques usually take into account a single metric when estimating the models, and there are multiple factors that play an important role when designing a soft sensor, such as stability and accuracy. To cope with such a situation, this paper proposes a multi-objective optimization design procedure, where feature selection and ensemble member combination are performed. Therefore, the multi-objective differential evolution algorithm with spherical pruning (spMODE-II) is initially employed for building a pool of non-dominated linear support vector regression (SVR) models. Subsequently, the same evolutionary algorithm is applied for selecting the weights of the previously generated models in a weighted combination ensemble. In a final multi-criteria decision making stage, a preferred ensemble is selected using the preference ranking organization method for enrichment of evaluations (PROMETHEE). Results indicate that the proposed approach is able to produce a highly stable and accurate butane content soft sensor for the debutanizer column.
展开▼
