Reliability analysis and data driven modelling of railway component failure

• Main Author: Wang, Jinlong
• Other Authors: Pang Hock Lye, John
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2022
• Subjects: Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/159237
• Institution: Nanyang Technological University

Reliability analysis and modelling of the railway component failure modes and mechanisms are essential for its proper functional performance, effective maintenance, and safe operation. However, it is challenging to apply the theoretical methods to the in-field failure record data which are heterogeneous in format. According to the data type, this research established a systematic historical maintenance data analysis framework in capturing the dynamic behavior of reliability indicators related conditioning variables. Failure modes and mechanisms from five components were investigated which including: (1) Welded rail break failure; (2) Rotating bending fatigue test for welded rail material; (3) Train door failure; (4) Rail wear degradation failure; (5) Wheel flat degradation failure. The one-dimensional time-to-failure type data sets in the first three cases were modeled by Weibull distribution models to characterize the reliability. Such statistical models are under some distribution assumptions which may not be fully satisfied in practice, but the models can be fitted with limited data samples and their interpretabilities are high. Data sets in cases (4) and (5) have larger data size and higher feature dimension which contain richer information. Therefore, data-driven machine learning based methods have been developed and modified to provide better degradation modelling capabilities. The novelty of this research mainly including: designing customized raw data processing, developing robust modelling framework, and providing implementable knowledge feedback for maintenance application. A novel method of combining Support Vector Regression with Archard wear law to predict the wear behavior of the rail steel was developed in case study (4). The proposed physical knowledge guided robust nonlinear regression analysis framework for multidimensional degradation data dealing with atypical hidden outliers demonstrated significant improvement in the modelling results. Pre-process of the raw wear data which involving feature importance analysis, physical model guided feature generation and outlier detection is developed within the framework for the SVR model’s robust learning. A Support Vector Classification model was developed in case study (5) to predict the alarm level of defect wheels three to five days before the Wheel Impact Load Detector (WILD) monitoring system actual report which could gain a longer time window for the maintenance resource arrangement. A novel wheel flat size prediction method based on Gaussian Process Regression model and Principle Component Analysis is another major contribution in this work. Human interpretable linear functions were derived for the in-field judgement of defect wheel flat size. Additionally, the established principles for predicting defect wheel flat sizes contribute to the optimization of alarm thresholds to improve maintenance efficiency. This research adopted a range of process techniques and algorithms to develop effective frameworks for railway component failure maintenance data clean. A WILD sensor big data set containing of 40 million samples was transformed through a proposed varied time window feature extraction method to integrate with another maintenance measurement dataset precisely. The developed feature extraction method guaranteed a larger amount of available training data to improve the machine learning accuracy. This research also paid attention to the interpretation and visualization of machine leaning based reliability analysis models and results. To explain the machine learning algorithm underlying mechanism, a game theory based SHAP(SHapley Additive exPlanations) analysis and a Principle Components Analysis based feature reduction methods were developed for machine learning explain and visualize accordingly.