Use this URL to cite or link to this record in EThOS:
Title: Predictive condition monitoring of rotating machinery with application to gas turbines and compressors
Author: Li, X.
ISNI:       0000 0004 7659 7472
Awarding Body: London South Bank University
Current Institution: London South Bank University
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Rotating machines, such as gas turbines and compressors, are widely used due to their high performance and robustness. These machines typically operate under adverse conditions, such as high loads and high temperatures and are thus subject to performance degradation and mechanical failure. In an effort to solve this problem, condition-based maintenance (CBM) was introduced to minimize safety risks and operational downtime hazards as well as to reduce maintenance and operation costs. One of the most critical aspects of CBM is the provision of incipient fault diagnosis and prognosis regarding the system's performance under faulty conditions. Traditional monitoring and alarm systems are currently widely used in the oil and gas industry to evaluate whether values of individual sensors exceed a threshold. Predictive maintenance requires techniques that are far more elaborate. Over the past decades, multivariate data-driven methods have attracted interest for condition monitoring in modern industrial plants due to the rapid growth and advancement in data acquisition technology. However, applications of these methods in industry are not widely reported. In view of the lack of research using real industrial data, this investigation focuses on the development of multivariate diagnostic and prognostic models that are applicable to operational industrial gas compressors and turbines, the early detection of faults, the identification of fault-associated variables and the estimation of performance deterioration after the appearance of faults. Although an increasing number of case studies of multivariate statistical monitoring has being reported in the past few years, the data employed in those studies are usually simulated data that are collected from simulation programs. The condition monitoring data of real industrial rotating equipment are generally not accessible by the public due to commercial confidentiality. Using condition monitoring data collected from operational industrial gas compressors and turbines, this work aims to provide case studies to demonstrate the capabilities of novel multivariate statistical monitoring approaches to detect faults and estimate the impacts of those faults on plant operations. Traditional statistical monitoring approaches are based on the assumption that the underlying processes are linear and static. However, this assumption might not hold true for real industrial processes because sensory signals affected by noise and disturbances often show strong nonlinearities, and the operating conditions often vary with time. As a result, static and linear methods may not be suitable for real-world applications because they provide incomplete representations of such systems. To address the limitations of standard multivariate statistical monitoring approaches for systems with both nonlinear and dynamic properties, canonical variate analysis (CVA) together with kernel density estimation (KDE) are employed in this work to detect diverse types of faults in rotating machines. The control limits associated with the proposed model were calculated based on the Hotelling's T^2 and Q metrics. The results obtained showed that the proposed method is effective for providing incipient fault diagnoses in the early stages of performance deterioration. For the purpose of fault diagnosis, 2-D contribution charts are utilized to identify the most fault related variables. The developed contribution plots can provide greater insights into the root causes of the faults and how the faults propagate to the remaining parts of the system. Predictive condition monitoring and preventive maintenance are seen as the means both to achieve high reliability and availability of complex rotating machines and to reduce unplanned production shutdowns. To achieve these goals, it is necessary not only to implement effective fault detection and diagnosis but also to react to the detected faults by continuously assessing and predicting the health status of the system. To test the capabilities of CVA for performance estimation, this method is first used to build a time-invariant state-space model of the dynamic system using purely historical condition monitoring data. The proposed method is applied to rotating equipment operating under both healthy conditions and slowly evolving faulty conditions to demonstrate its applicability and effectiveness. The use of a time-invariant model for system identification limits its application to linear and stationary processes. The use of a time-varying model can overcome this limitation by allowing model adaptation to rapid changes in system operating conditions of time-varying processes. To address the challenge of implementing prognostics in real-world applications with both dynamic and nonlinear properties, the time-invariant CVA model is extended using recursive least squares (RLS), resulting in the improved adaptive CVA prognostic model. The extended CVA method proposed in this work is evaluated using data captured from rotating machines operating under rapidly varying healthy conditions as well as faulty conditions. Furthermore, to account for the impact of environmental factors on a system's performance, in this work, CVA combined with long short-term memory (LSTM) is used to estimate the behaviour of a centrifugal compressor after the occurrence of a fault using data captured during the early stages of deterioration. The results of this study indicate that CVA can effectively capture the system dynamics for large-scale complex rotating machines, thereby enabling the early detection of faults, the diagnosis of the root cause of the detected faults and the prediction of system behaviour after the appearance of faults. A systematic fault detection, isolation and estimation scheme can be developed based on the proposed techniques, based on which the whole plant-wide process can be monitored at both the plant-wide and the unit levels, and the monitoring information can be used to improve maintenance decisions and to reduce unscheduled downtime.
Supervisor: Dudley, S. ; Duan, F. ; Mba, D. Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral