Sensor-fault-diagnosis using inverse dynamic systems

Abstract

The principle of observer-based sensor fault detection and isolation is extended by making use of inverse dynamic systems. Residual signals are generated by taking the observation error as an input to the left inverse of the observer which in turn has the desired residual signals as outputs. In this way the residuals ideally respond to their associated faults without any delay, and perfect fault isolation is achieved at the same time. However, it is generally difficult to find a stable inverse to an FDI-observer, especially when the plant is unstable or exhibits nonminimum-phase behaviour. Since the resulting equations for the construction of an inverse are strongly nonlinear, a genetic optimization algorithm was applied to solve the problem. It is demonstrated how both the observer-eigenstructure and the observer inverse can be optimized simultaneously using genetic optimization algorithms. In order to illustrate its applicability, the method is applied to an industrial turbo-charged combustion engine power plant.