Skip to Main Content
Fault detection is a part of every industrial engineer's brief, particularly in chemical plants. Traditional detection methods in this field have depended on limit checking of measurable output variables using standard SPC techniques; however, this approach is fraught with problems, notably: alarms are not raised until the fault is actually manifesting itself at the outputs; noise on the outputs may mask incipient faults; and many plants have too many variables to monitor them all. More recent fault detection schemes have tended to be model-based, using various types of model. The major problem is model identification, particularly in large plants with input and output variables. The approach described in this paper uses a combination of these two procedures. A statistical model is generated via partial least squares (PLS), a multivariate statistical modelling technique. Results from simulation studies on an EPSRC-funded benchmark plant, consisting of an overheads condenser and a reflux drum, are presented to illustrate the success of the approach. Standard SPC techniques are then used to detect simulated faults by analysis of the mismatch between the PLS model prediction and the original plant.