S-Space College of Engineering/Engineering Practice School (공과대학/대학원) Dept. of Mechanical Aerospace Engineering (기계항공공학부) Theses (Ph.D. / Sc.D._기계항공공학부)
Unsupervised Learning and Diagnosis Method for Journal Bearing System in a Large-scale Power Plant
- 공과대학 기계항공공학부
- Issue Date
- 서울대학교 대학원
- 학위논문 (박사)-- 서울대학교 대학원 : 기계항공공학부, 2016. 2. 윤병동.
- Rotor systems are frequently used in machines and facilities for various industrial applications. Often the rotor systems fail to deliver their designed performance, thus resulting in substantial financial loss. These issues are very critical to some industrial sectors (such as power plants) where journal bearing rotor systems are regularly used. As a result, it is common to implement a diagnosis tool to such rotor systems. Automated diagnosis using data-driven techniques can enable detection of anomalies during early stages and can thus contribute to improved safety and increased cost savings. In the process of developing diagnosis algorithm, the robustness is one of the best important issues. Furthermore, for the diagnosis of a variety of fault conditions that may occur in a real system, the application of unsupervised learning techniques is needed.
In order to facilitate the development of robust diagnosis methodologies for the rotors in journal bearing systems, this research aims at advancing two essential research areas: Research Thrust 1 – datum unit optimization and Research Thrust 2 – omnidirectional regeneration (ODR) of gap sensor signals. In Research Thrust 1, the optimal datum unit will be defined by the comparison of separability and classification performance among feasible datum units. In Research Thrust 2, highly accurate and robust diagnosis approach using ODR signals will be introduced. The virtually generated ODR signals for circumferential direction can fully represent the vibration behavior of rotor system. Following the development of Research Thrust 1 and 2, Research Thrust 3 – unsupervised learning framework for power plant will give the basis of extension to the actual power plant diagnosis. Deep learning for unsupervised technique with high-level feature gives the reliable clustering results for power plant diagnosis.