Date of Award


Document Type


Degree Name

Doctor of Philosophy (PhD)

Legacy Department

Mechanical Engineering


Wagner, John R


The detection of system anomalies represents an important task in safety critical systems. Over the past two decades, advances in data manipulation techniques and sensor technology have contributed to the widespread application of signal processing concepts for fault diagnosis. However, the majority of these signal processing, or data driven, diagnostics do not possess the redundancy level exhibited in conventional model-based methods. Among signal processing methods, time series analysis is advantageous since it offers insight into the underlying dynamics and forecasts system behavior. Furthermore, the data driven method allows the depiction of system dynamics with an increased level of redundancy. This dissertation establishes the foundation for a new comparative diagnostic methodology that examines the dynamic similarity between two systems. The first system represents a reference health condition; the second contains the plant signals to be assessed. Linear and nonlinear time series principles may be applied for comparison purposes. The behavior of a dynamic system operating in a steady-state mode at select equilibrium points may be described by small perturbations, stochastic in nature, about these points. In a multivariate context, the auto covariance matrix function of linear time series, as an invariant property, may evaluate the dynamic similarity between two signal clusters. For systems exhibiting cyclic (i.e., start, steady-state, shutdown) operational behavior, two nonlinear time series approaches have been proposed. Trajectories in a re-constructed control state space offer the opportunity to investigate the system's governing dynamics, and develop pertinent diagnostic measures. Deviations in the trajectory geometric features between the reference and the test condition may be quantified using both recurrence plots and Poincaré Sections techniques. The recurrence plot method emphasizes steady-state periods while the Poincaré Sections assess the system behavior in transient (ramping-up) conditions. To validate these diagnostic concepts, heavily instrumented electric power generating natural gas turbines were studied. A Solar Turbines 4.5 MW Mercury 50 gas turbine (Clemson, SC) demonstrated the performance of the linear time series method with induced faults that were as low as 2% of the nominal signal level. The consistent declaration of the fault with 5% noise-to-signal power has been exhibited. When compared to a conventional red-line alarm method, the proposed technique offered a 2% shorter detection delay time. Furthermore, faults were detected that did not manifest themselves in the observed signals which would have been missed by the alarm method. A cluster of three 85 MW General Electric GE-7EA gas turbines (Iva, SC) served to validate the nonlinear time series methods. An assessment of the equipment's health condition was performed based on historical operating data for a one year time period. The recurrence plots offered a macro-level appraisal of the turbine behavior during the test period, while the recurrence quantitative analysis was able to detect steady-state deviations of 1%. Equally, the utilization of Poncaré Sections for the same equipment cluster enabled the detection of interesting events, during start-up periods such as load rejection.