Feature extraction using wavelet analysis with application to machine fault diagnosis

• Main Author: Tafreshi, Reza
• Language: English
• Published: 2009
• Online Access: http://hdl.handle.net/2429/16965

Two different approaches have been used to diagnose faults in machinery such as internal combustion engines. In the first approach, a mathematical model of the specific engine or component under investigation is developed and a search for causes of change in engine performance is conducted based on the observations made in the system output. In the second approach, the specific engine or component is considered a black box. Then, by observing some sensory data, such as cylinder pressure, cylinder block vibrations, exhaust gas temperatures, and acoustic emissions, and analyzing them, fault(s) can be traced and detected. In this research the latter approach is employed in which vibration data is used for the detection of malfunctions in reciprocating internal combustion engines. The objective of this thesis is to develop effective data-driven methodologies for fault detection and diagnosis. The main application is the detection and characterization of combustion related faults in reciprocating engines; faults such as knock, improper ignition timing, loose intake and exhaust valves, and improper valve clearances. To perform fault diagnosis in internal combustion engines, cylinder head vibration data are used for characterizing the underlying mechanical and combustion processes. Fault diagnosis includes two main stages: feature extraction and classification. In the feature extraction stage, we have utilized wavelets for the analysis of acceleration data acquired at the cylinder head to capture meaningful features that include necessary information about the state of the engine. Wavelets have shown to provide suitable signal processing means for analysis of transient data and noise reduction. Wavelet packets, as a generalization of wavelets, offer even a more powerful data analysis structure to extract features that are capable of identifying combustion malfunctions. Various concepts of wavelets, wavelet packets, related algorithms and assessment techniques have been reviewed, analyzed and discussed. As a result of this research, a novel methodology for fault diagnosis has been developed. This has been achieved through critically investigating available methodologies employed in fault diagnosis and classification, and by understanding their shortcomings. The developed method not only avoids the demerits of the previous techniques, but also demonstrates superior performance. To compare the performance of the proposed approach with major existing methods, various sets of real-world machine data acquired by mounting accelerometer sensors on the cylinder head, as well as a set of synthetic data, have been extensively tested. === Applied Science, Faculty of === Mechanical Engineering, Department of === Graduate