Detecting the failure of a sensor in industrial processes is important to avoid the use of incorrect measurements. When a sensor fails the missing measurement are reconstructed, using the measurements of other sensors and inferring the missing or incorrect measurement. Although this technology has been developed more than 20 years ago, there are few commercial solutions available today. One of these few solutions uses principal component analysis, based on an algorithm originally developed by Qin and Li (1999). In this paper, this solution is applied to operating data from a minerals processing plant with persistent sensor problems Somewhat surprisingly, poor results are obtained, despite numerous attempts to improve reconstructability. Analysis indicates that the challenges are not about the algorithm but rather about choices that need to be made in the application of data-driven analysis tools to new data sets. These include data selection, filtering and interpreting which results are useful. It is suggested that together with any new algorithm presented. researchers should provide practical guidelines in choosing appropriate data, and any pre-processing that may be required.