ENBIS-18 in Nancy

2 – 25 September 2018; Ecoles des Mines, Nancy (France) Abstract submission: 20 December 2017 – 4 June 2018

A Multiresolution Framework for Building Industrial Soft Sensors

3 September 2018, 14:00 – 14:20


Submitted by
Tiago Rato
Tiago Rato (University of Coimbra), Marco Reis (University of Coimbra)
The quality features of industrial processes are typically obtained offline with a considerable delay and by resort to expensive equipment. Therefore, in order to make the quality assessment faster and decrease the experimental burden in the routine quality laboratories, a variety of data-driven soft sensors have been developed. These models are expected to capture the dominant relationships between the different process variables (predictors) and the relevant quality variables (responses), while accounting for their high-dimensional, dynamic and multiresolution structure. The first two characteristics are often addressed by standard latent variables models or variable selection algorithms. However, as current methodologies tacitly assume that all variables carry delocalized information about the process on exactly the same time period (i.e., they assume that the variables have the same resolution), the multiresolution aspect is usually disregarded. Furthermore, multiresolution is often confused with a multirate problem: multiresolution occurs when variables have different levels of granularity due to, for instance, automatic averaging operations over certain time windows; while a multirate structure is caused by the existence of different sampling rates, without changing the granularity of the variables.

The inconsistent use of the variables’ resolution limits the application of current soft sensor approaches to multiresolution data structures, namely their interpretational and predictive capabilities. Furthermore, even if the data is available at a single-resolution, it is not guaranteed that the native resolution of the predictors is the most appropriate for process modelling. Therefore, soft sensor methodologies must address not only the selection of the best subset of predictors to be included in the model, but also the optimum resolution to adopt for each predictor. For this purpose, novel feature selection algorithms are proposed for continuous and batch processes. The performance of the developed multiresolution soft sensors is comparatively assessed against their single-resolution counterparts. The results demonstrate that the optimized multiresolution soft sensors are bounded to be at least as good as current single-resolution methodologies and that they are almost always substantially better.

Return to programme