A warning system for detection of excessive vibration in a centrifuge system of a product treatment plant is built using a database of past faults and an equivalent amount of normal operating null data. A logistic Partial Least Squares (PLS) model is derived using wavelet coefficients to approximately decorrelate the time series data. This model provides a baseline to evaluate any improvement through kernel methods. The kernel paradigm is introduced from a Bayesian perspective and used to develop a detector with significantly less false positives and missed detections. (C) 2009 Elsevier Ltd. All rights reserved.