This paper describes a technique for on-line process fault detection and diagnosis based upon the statistical analysis of process data. A principal component model is developed from nominal operating data and is used to reduce the dimensionality of the measured process information. Current process performance is monitored through plots of the squared prediction error and the principal component scores. The fault signatures for various process malfunctions are located through multivariate statistical analysis of on-line monitored data. In order to identify fault directions in measurement space, a principal component analysis is carried out for each data set corresponding to different process malfunctions. Fault diagnosis is then achieved by comparing the direction of the current on-line measurements with those of a database of known trajectories of previously identified faults. The fault whose direction is most closely aligned with the current data direction is identified to be the most likely fault to have occurred and is taken as the diagnosis result. The technique is straightforward to implement and can be used to complement current fault diagnosis approaches. Application of the proposed technique to the on-line diagnosis of equipment malfunctions in a continuous stirred tank reactor (CSTR) system illustrates the power of the approach.