Model based methods for sensor fault-tolerant control of rail vehicle traction
This thesis explores the application of modem fault-detection methods to electric rail traction drives. Such drives consist of three main components, induction motors, power inverters and the control system. The power electronics are relatively simple so the scope for fault-tolerance is limited, whilst fault-detection techniques for induction motors are already well developed. There is however scope for work on the instrumentation. The thesis concentrates on the use of model-based techniques to produce a torque and flux estimator for an induction motor which is tolerant to intermittent sensor disconnections. The motors are controlled on torque and flux, these cannot be measured directly and are estimated from measurements of the applied voltages and the resulting currents. The existing estimator has poor steady-state performance at low speed and because of it's transient dynamics it is prone to sensor noise and disconnections. Induction motors have speed-dependent dynamics and the resulting state-space model has terms which are multiplied by speed, this model is strongly bilinear. Speed-dependent feedback is needed to give desirable dynamics to the state estimates. Starting from a state-space model for the induction motor, a closed-loop observer can be designed to estimate the motor states. A range of feedback methods for the observer have been considered, from gain scheduling to sliding mode techniques. These are evaluated in simulation, using a simplified model of the traction system. The simulation neglects many second order effects which would effect the real application. Using data from an induction motor test-rig the observers are shown to be able to track the motor torque during a change in operating condition. Only a limited set of data is available. The influence of parameter mis-match, noise and speed sensor errors are considered by deriving frequency domain expressions for the estimation error in the presence of uncertainty or disturbances. The effect of the observer's gain on its sensitivity to these are considered under conditions which occur in the real application. Using observer feedback to decouple sensors from the estimation a range of sensor fault-detection schemes are developed. In this way a bank of observers is designed which are independent of a different subset of sensors, this enables sensor faults to be isolated. These method are compared in simulation. A motor, inverter and instrumentation are set up, with a DSP to run an observer based sensor fault-detection scheme in real-time. This enables implementation aspects to be explored, such as discretisation, model mis-match and motor loading. These effect the detection by increasing fault-free residual or reducing the fault residual. For a each type of sensor the area of the motor operating range, where a fault is detectable is defined.