Search

Abstract: In recent decades, condition-based maintenance has been acknowledged as a cost-effective maintenance program and widely used in many engineering systems. Diagnostics and prognostics are critical components of condition-based maintenance, responsible for providing information about present and future system health conditions. These two components are respectively an integrated process covering several essential aspects that should be cared about to ensure successful implementation of diagnostics and prognostics. This thesis focuses on certain aspects for diagnostics and prognostics, respectively. When doing diagnostics, data processing is an important stage responsible for providing reliable data for use. Data processing usually includes feature extraction, data cleaning and feature selection. Feature extraction is responsible for extracting from the raw data characteristic features representing system conditions. Data cleaning is necessary for removing outliers caused by the noise during data collection. Feature selection is responsible for capturing and removing useless features generated during feature extraction. This thesis focuses on data cleaning and feature selection. When doing prognostics, noise may appear in condition indicator values. The condition indicator is extracted from condition monitoring data and is able to reflect the health conditions of monitored assets. Generally, using the noisy condition indicator values may result in unreliable predictions for prognostics, so there is a demand of the method that can provide predictions without the effects of noise. Support vector machine (SVM) is recognized an effective tool for classification and prediction that are needed by diagnostics and prognostics. SVM is a supervised-learning method and is reported to have better generalization ability and superior performance for small sample cases over other supervised learning methods such as artificial neural network. This thesis develops SVM-based methods to solve some problems existing in diagnostics and prognostics. The contributions of this thesis are summarized as follows: 1. An SVM-based data cleaning algorithm that removes outliers for effective diagnostics. 2. An SVM-based feature selection algorithm that removes useless features for effective diagnostics. 3. An analytical method that selects SVM model parameters for effective on-line system condition prognostics. 4. An intelligent optimization-based method that selects SVM model parameters for effective on-line system condition prognostics.