The application of time series predictive maintenance model on rolling stock critical systems

The reliability analysis of railway rolling stock systems determines how efficiently passenger services operate and how sustainable operations remain over a specified period. The proposed framework presented in this paper implements supervised learning to predict system downtime while using time-series forecasting to evaluate incident risk levels. Three regression models, Bayesian Ridge Regression, Support Vector Regression (SVR), and K-Nearest Neighbors (KNN), were trained on historical failure records to estimate Mean Time to Repair (MTTR). The naive baseline MTTR data is used as input for modelling and prediction downtimes, and further analysis using the Bayesian Ridge model provided the most accurate results. Seasonal Autoregressive Integrated Moving Average with Exogenous variables (SARIMAX), Long Short-Term Memory networks, Convolutional Neural Networks, and Exponential Smoothing approaches were used to analyse weekly and monthly CCTV camera failure incidents. The risk of system failure incidents was classified as low, medium, and high-risk levels to support decisions on maintenance schedules, spare parts, and resource allocation in advance. The dynamic modelling using machine learning and time series modelling approach presented in this paper can support maintenance managers make decision on optimal preventive maintenance activities and hence reduce maintenance downtimes.