Sliding mode observers for fault detection and isolation

This thesis describes the use of a class of sliding mode observers for fault detection and isolation purposes. Existing work has shown that the equivalent output error injection term associated with the sliding mode observer, which represents the average value of the nonlinear switched term (which induces and maintains the sliding motion), if properly scaled, yields accurate reconstructions of actuator faults. Existing observer design methods generate a certain class of observer gains, but do not utilise all degrees of freedom. In this thesis, a new method, exploiting this freedom is presented. The method uses Linear Matrix Inequalities and is easily implementable using standard software packages. New methods for accurately reconstructing sensor faults are also presented where appropriate filtering of certain measurable signals yields a fictitious system in which the original sensor faults are treated as actuator faults. Using the principles of actuator fault reconstruction in the existing work, sliding mode observers can be designed for the fictitious system to accurately reconstruct the sensor faults. This improves on the previous work where effectively only the steady state components of the sensor faults could be reconstructed. A new method using Linear Matrix Inequalities is presented, to synthesise observers which can robustly reconstruct faults in the presence of a class system of uncertainty, minimising the effect of the uncertainty on the fault reconstruction in an L2 sense. The robust fault reconstruction scheme is demonstrated by means of a case study, which is a nonlinear model of an aero-engine. System identification is used to obtain a linear model of the engine. An uncertainty representation is also obtained about which the observer is designed. The results from the case study show that the robust fault reconstruction scheme works and is effective.