Chemical process operations an rarely steady and often vary with time. Slow and persistent process drifts associated with gross errors make data rectification difficult. In this study, a multivariate statistical process control method for detecting abnormal process conditions, which is based on the posterior mode estimator, was proposed. It first decomposed the observation equation of a state-space model into a trend component, an irregular component, and a gross error component with recursive transition equations. A penalized least-square method was then applied to derive the maximized likelihood and unbiased estimation of the measurements. A precise prior-assigned model was not needed for the estimator, and it could be implemented on-line easily. Statistical methods based on Hotelling＇s T-2 distance, Wilks Lambda criterion, simultaneous confidence intervals, and measurement tests were adopted to detect abnormal drifts in process operations. For performing these tests, a simple method of introducing a forgetting estimation on the steady-state operation values was suggested to enhance the robustness property and to treat the effects of non-steady data measurements. An on-line monitoring system, underlying the desire for real-time analysis and monitoring in chemical or petrochemical plants, was developed for detecting process upsets and drifts. The promising performance of the proposed method in detecting after-burning hazards has finally been demonstrated by a continuous catalyst regenerator in comparison with the results suggested by a senior process engineer for practical operations.