Traditional engineering systems serve us in all areas of life. They include industrial production facilities (chemical plants, oil refineries, power stations, etc), transportation vehicles (airplanes, ships, automobiles), and household appliances (furnaces, washing machines, etc). Faults (malfunctions) in components of such systems jeopardize their function and integrity, resulting in economic loss, and may even threaten human life. Detecting the onset of faults, and determining their location, are important engineering tasks, usually performed with the help of computers.
An important class of fault detection and diagnosis methods utilizes the mathematical model of the monitored system; this is where our group has been active in the past 20 years. The basic idea of such methods is “analytical redundancy”; supplementing the physical sensors (measuring devices), the model is used as if it were a set of additional sensors. The physical measurements are compared to the ones expected on the basis of the model; any discrepancy, expressed as “residuals”, is the indication of possible faults. Fault detection is aimed at detecting the presence of faults from the residuals that are influenced also by nuisance effects, such as noise, disturbances and model errors. Fault isolation is aimed at determining the location of the fault, usually by some manipulation of the residuals. The analytical redundancy methods rely on modeling, analysis and design techniques borrowed from control and systems engineering.
Our theoretical work has concentrated on two main approaches:
- using the dynamic model of the system, mainly in the form of discrete transfer functions, to generate direct analytical redundancy relations;
- using a statistical model, obtained by principal component modeling, for the same.
Our main interest has been in fault isolation, via the development of residual manipulation techniques resulting in residual sets that exhibit structural or directional properties in response to various faults.
We had a major application project jointly with the General Motors Corporation. We developed a model-based method, and a set of models to support it, for the on-board detection and isolation of component faults affecting the car’s emission control system (“check engine” light on the dashboard). After some additional, in-house development, GM has been in the process of gradually introducing this method on its mass-produced cars.
We also did extensive simulation studies of our principal component methods on realistic models of large-scale chemical process systems.
Our work has been supported by the General Motors Corporation, the National Science Foundation, Virginia’s Center of Innovative Technology, and by the French Ministry of Research.