The diagnosis scheme compares actual measurements with predicted nominal values of process variables that characterize normal operation. This comparison process is termed fault detection. In processes that operate in steady state, nominal values can often be retrieved from design specifications or documentation created by process engineers. To account for the effects of noise and measurement inaccuracies, a margin of error is added to the nominal values to increase robustness and avoid false alarms [22]. However, this decreases sensitivity, which is acceptable provided the delayed detection does not result in dramatic errors. For systems that typically operate in steady state modes, design documents often specify the upper and lower limits on nominal values of all system parameters and measured variables.
For systems whose normal operational modes include transients and dynamic behaviors, nominal values of process variables are harder to obtain. A fairly accurate process model is required to run in parallel with the process. Given the same initial state and the same input as the process the simulation mechanism should predict the process output in normal operation. In reality, approximations in the models and drift in the system may result in the estimated state vector slowly deviating from the actual system values. To prevent this, an observer mechanism shown in Fig. 2 can be used to estimate and make corrections to the estimated state vector. A critical issue when applying this scheme to obtain nominal values is the model adaptation rate, especially in case of incipient faults. If this rate is too fast, the model quickly adapts to changes in the system variables due to faults and generates nominal values that do not indicate a deviation.
Figure 2: A general observer scheme.