Sheet-break is a long standing problem in the pulp and paper industry. This study is concerned with the analysis of process data to diagnose causes of sheet-breaks and therefore significant down times. PCA was used to model the process and a combined index based on the Hotelling's T-2 and Squared Prediction Error (SPE) was developed as a sheet-break detection indicator. As the process is subject to external disturbances, changes and frequent interruptions, pre-processing of the data played an important role in getting consistent results. We used several novel techniques for data selection, scaling and modelling. The models were validated using a large validation data set with known fault conditions. The developed model, data visualization tool and engineering judgement was used for off-line diagnosis of root causes of sheet-breaks. Several operational changes were recommended and implemented on the process resulting in significantly reduced sheet-breaks. Key Performance Indicators calculated before and after the changes shows the significant economic gain as, a result of this 'data-mining' project.