The main purpose of fault detection in batch process monitoring is to identify batches displaying atypical behavior in comparison to normal operating data. The current growth in the number of measurable variables due to process automation yields datasets in which the number of variables is much larger than the number of batches. That may compromise the performance of Multiway Principal Component Analysis (MPCA), which is the most popular quality control approach used in batch processes. To overcome that, new strategies to handle high-dimensional datasets become necessary. In this paper we propose the Pareto Variable Selection (PVS) - MPCA method to monitor batch processes described by high-dimensional datasets. The main idea of PVS-MPCA is to select process variables that promote the best classification of production batches in conforming or non-conforming classes, prior to the construction of T-2 and Q control charts used to monitor batch performance. Our proposition was applied to a real dataset from a chocolate conching batch operation and compared to classical MPCA-based monitoring. PVS-MPCA promoted a reduction of 85.18% in false alarm rate retaining only 5 unfolded variables, in opposition to 2,864 unfolded variables used in classical MPCA. The missed detection rate was null, ensuring that only conforming batches were released to the production line. (C) 2019 Elsevier Ltd. All rights reserved.