Fault Detection and Analysis
Algorithmic methods to detect faults in instrumentation of process automation systems have been developed and tested in laboratory process equipped with standard sensors and actuators. Faults which are hidden in the sense that instrument signals still stay normal are under main interest. Methods under studying are based either on heuristic reasoning or analytic redundancy by detecting changes in estimation error signals of specific non-linear filters.
In the 80's and early 90's, the laboratory contributed fault detection and analysis based on utilization of state estimator or observers. The changes in stochastical characteristics of the residual signal reveals different kind of faults. The faults can be detected with stochastical tests in reliable way.
Also knowledge based methods have been developed. Expert system based solutions were developed in the late 80's. G2 software system which utilized temporal logic were developed in the middle of 90's.
In the late 90's, neural networks have been applied to fault diagnosis. The orthogonal signals representations, calculated on-line with the corresponding linear filters, are effective feature vectors for recognition of fault situations in batch type processes. This Wiener-NN approach can be utilized also by implementing parallel estimators for key state variables, which base on information from different subsystems; cooling system, outgas measurements and pH-control. The process events can be seen in redundant way in all these sensor groups for these subsystems. The redundancy is utilized in fault detection and analysis, see Nonlinear Wiener-NN models for dynamic systems.