Impact of Data Grouping on the Multivariate Analysis of Several Concrete Plants

Abstract

Multivariate analysis can be used to study industrial process data exhibiting collinearity between variables. Such data can often be collected into conceptually meaningful groups or blocks. While data blocks may appear intuitive (e.g., raw material properties vs. process parameters), such blocking is sometimes much more subjective. The novelty of this work lies in the investigation of the impact of data blocking on the subsequent analysis. To our knowledge, no such investigation can be found in the literature. To fill this gap, we analyze the impact of grouping data from 10 Canadian concrete plants in which multiple blocking alternatives are considered. The analysis is performed via principal component analysis (PCA) to reduce the dimensionality of the matrix and also via consensus principal component analysis (CPCA). The data grouping options are as follows: (1) all data combined into a single block, (2) grouped according to the factory, (3) grouped according to parameter type, and (4) grouped according to parameter type within each factory. The results show that the grouping strategy alters the conclusion by emphasizing specific aspects of the data. While some grouping options emphasized seasonal variations, others emphasized other characteristics in the data, such as step changes in processing regimes or the significant impact of the raw materials’ moisture on the process. As such, it appears relevant to consider multiple blocking options when analyzing complex datasets. Doing so will give the analyst a better understanding of overarching trends and more subtle characteristics of the dataset.