Fault detection approach for systems involving soft sensors

This paper develops a new approach for fault detection which involves soft sensors for process monitoring. Unlike existing approaches, which compare current measurements, or linear combinations thereof, to values of these measurements representing normal operations, the methodology presented here deals directly with the state estimates that need to be monitored. The advantage of such an approach is that the effect of abnormal process conditions on the state variables can be directly observed and that it is possible to include nonlinear relationships between measurements and states. At the same time, this type of approach has the drawback that the variances of the unmeasured states are not equal to the variances of the actual process variables due to the use of a soft sensor. However, for many popular soft sensor techniques, such as Kalman filters and related approaches, it is possible to compute variances of the predicted states that correspond to normal operating conditions. This paper presents a general framework for using soft sensors for process monitoring, i.e., soft sensor design and computation of the statistics that represent normal operating conditions, and illustrates this framework in three specific applications. It should be pointed out that the contribution of this work does not lie with the soft sensor design or the computation of the statistics itself as either part has individually already been addressed in the existing literature. However, the authors are not aware of any studies where both tasks are combined for process monitoring, which forms the contribution of this work.

► Fault boundaries for estimated variables of dynamic systems are determined.
► Derivation and discussion includes technique for nonlinear systems.
► Case studies provide illustration of method and results.