Diagnosis of machine tools based on inverse problem approaches and time-frequency techniques

This thesis presents the development of a generalized framework for condition monitoring and diagnosis of machine tool systems. Diagnosis is an "effect-cause" problem wherein the goal is to determine the unknown cause of some observed effects. Inverse problem approaches are well suited to solving such problems, and are used as the basis for the framework developed in this thesis. Usually, the diagnosis of machine tools is undertaken using very process-specific methods, which severely limits the applications of the resulting diagnostic system. The generality of the developed framework is ensured by decomposing diagnosis into fundamental "building blocks", and then defining the relationships between these blocks so that they can be used for a variety of diagnostic tasks. Using inverse problem approaches allows for the posedness of the diagnostic system to be considered, allows the limits of the diagnostic system to be determined and works to prevent false alarms or missed detections when performing monitoring operations.; For diagnosis, specific features relating directly to the health of the system need to be identified. Using time-frequency transforms as operators within the diagnostic framework, it is possible to isolate features that uniquely represent an operational state of the system being diagnosed. The use of the recently developed technique of selective regional correlation (SRC) localizes these health-specific features in the time-frequency domain. This improves the results of correlation analysis by ensuring that the features used are intrinsically related to the machine tool being diagnosed.; The generalized framework is successfully applied to two different diagnostic tasks: a machine condition monitoring (MCM) application and a more complex tool condition monitoring (TCM) application in which four different states of wear on machine tool cutters are identified. The framework developed is shown to be flexible enough to be adapted both of these tasks and yields results superior to conventional time domain correlation-based techniques.