Data preprocessing for multiblock modelling - A systematization with new methods

摘要

With the advance of Industry 4.0, new data collectors are appearing at different points of the process generating blocks of data whose integrity should be preserved during data analysis. This is the scope of multiblock methods, whose potential has been recognized in several areas of application where they are becoming increasingly popular. Multiblock methods can be applied to a wide range of data-driven problems that practitioners face nowadays such as plant-wide process monitoring and diagnosis, process optimization and quality prediction of key product properties. These methods have the ability to find associations and interpretative connections between different data blocks from different sources and carrying complementary or overlapping information, as well as assessing the blocks' relative contributions to the final outcome. A critical stage in the application of multiblock methods is the selection of the appropriate preprocessing to apply to each block, before proceeding to the modelling. The preprocessing strategy can exponentiate the information extracted from the blocks and their mutual interactions or hide/mask/distort them if inappropriately done. In this article, we present a systematic workflow where both the intra-block and inter-block variation components are considered during preprocessing. We illustrate the application of the framework using two real case studies where a critical comparison is presented for the different preprocessing alternatives.