A new method is proposed in this work to detect and isolate faults in a system consisting of multivariate normal data. The proposed method reduces the variable dimension via a T2 decomposition and combines multivariate monitoring and diagnosis in a unified approach. An adaptive regression adjusted (ARA) chart is introduced to utilize the diagnosis result in order to improve the monitoring scheme. Once an out-of-control alarm is detected, the fault isolation can be done directly. Copyright © 2007 John Wiley & Sons, Ltd.