Sensor-based diagnosis of dynamical systems

In highly automated engineering systems, sensors provide the data needed for control, monitoring and diagnosis. The reliability and cost of these sensors are important issues. The thesis describes a model-based approach to handle sensor failures and to provide optimum sensor allocation in diagnostic systems. The first part of this work deals with providing reliable diagnosis of failures in a system, given the possibility that sensors might fail. The correctness of diagnostic results depends upon the reliability of observations. The observations can be erroneous because (1) faulty data reading by sensors, (2) modeling error, and (3) in case of a major fault, loss of model validity. A robust diagnostic system to handle the above possible causes of errors was developed. The diagnostic system uses the physical and temporal constraints imposed on observations in a dynamical system to identify the presence of erroneous observations. The second part of the work deals with minimizing the costs associated with sensors without sacrificing the diagnosability of the system. A Diagnosability Analysis Tool (DTOOL) was developed, which facilitates the analysis of diagnosability in terms of the sensors in the system. Three metrics to characterize diagnosability were defined: (1) Detectability, which gives the longest time that is needed to detect a failure, (2) Predictability, which gives the shortest time between the forewarning and the actual occurrence of a failure, and (3) Distinguishability, which describes the size of the ambiguity sets given a time limit for the observation. Using these metrics, DTOOL provides three kinds of analyses. In evaluation mode, the diagnosability characteristics of a design with a predefined sensor allocation are calculated. In advice mode, an arbitrary set of requirements can be defined for the diagnosability characteristics, and the tool generates a satisfactory sensor placement. The tool also provides test tree generation analysis.