Characterisation of Condition Monitoring Information for Diagnosis and Prognosis using Advanced Statistical Models

This research focuses on classification of categorical events using advanced statistical models. Primarily utilised to detect and identify individual component faults and deviations from normal healthy operation of reciprocating compressors. Effective monitoring of condition ensuring optimal efficiency and reliability whilst maintaining the highest possible safety standards and reducing costs and inconvenience due to impaired performance.

Variability of operating conditions being revealed through examination of vibration signals recorded at strategic points of the process. Analysis of these signals informing expectations with respect to tolerable degrees of imperfection in specific components.

Isolating inherent process variability from extraneous variability affords reliable means of ascertaining system health and functionality. Vibration envelope spectra offering highly responsive model parameters for diagnostic purposes.

This thesis examines novel approaches to alleviating the computational burdens of large data analysis through investigation of the potential input variables. Three methods are investigated as follows:

Method one employs multivariate variable clustering to ascertain homogeneity amongst input variables. A series of heterogeneous groups being formed from each of which explanatory input variables are selected.

Data reduction techniques, method two, offer an alternative means of constructing predictive classifiers. A reduced number of reconstructed explanatory variables provide enhanced modelling capabilities ensuring algorithmic convergence.

The final novel approach proposed combines both these methods alongside wavelet data compression techniques. Simplifying number of input parameters and individual signal volume whilst retaining crucial information for deterministic supremacy.