Fault Detection Based on Gaussian Process Models

The traditional model-based fault detection and isolation (FDI) rely on tacit assumption that the validity of process model is out of question for current process data. However, it happens in practice that a process might end up in regions for which the underlying model is not validated. This can result in false alarms. To avoid the risk, a validity index is suggested as a measure of confidence assigned to the detection results. This measure is based on estimated distance of the current process data from data employed in the learning set. If the former is close to the latter, the detector output will be assigned high confidence. It is shown in this paper that the idea can be realized in a rather natural way when Gaussian Process models are used to describe the input-output behavior of nonlinear dynamic systems. This is a relatively recent modelling approach, which is probably for the first time adopted to the FDI problem domain. For a particular case of sensor faults it results in a statistical test in the form of a weighted sum of prediction errors. The effectiveness of the test is demonstrated on a simulated pH process.