Fault diagnosis in time series data with application to railway assets

Abstract

This thesis focuses on diagnosing incipient faults in railway train assets through large data sets. It investigates data mining methods to detect changes in big time series data to enable service by employing three case studies that separately investigate (1) the doors, (2) the engines, and (3) the wheels. Each case study comprises data with different characteristics: the first case study examining engines presents datasets with variable speed and load; the second case study examining door data highlights start-stop characteristics, with discontinuities in the data; and the third case study examining wheel data contains fast transitions from normal to slip behaviours in acceleration and deceleration. A time series architecture is proposed. It describes key steps for multivariate time series analysis and enables iterative improvement reusing results for the next iteration. A fault diagnosis is made for each case scenario, and procedures for synchronisation and alignment; pre-processing; and methods for feature extraction, classification or clustering are thus presented. For engine fault diagnosis, the proposed graphical method has the best performance. In the case of the door fault diagnosis, the K-nearest neighbours method has the best performance. In the wheel slip diagnosis case study, the combined wavelet and LSTM methods present the best accuracy. Limitations include data quality issues, key input data uncertainties, and applied classification deficiencies. The main challenges include big datasets that are desynchronised, wrongly time-indexed, noisy, redundant, and unlabelled; infrequent faults; the lack of monitoring in several subsystems; and incomplete or missing maintenance records.