Towards Condition-Based Maintenance of Railway Locomotives: A Data-Driven Analysis of Fault Behaviour and Predictive Modelling

This study investigates data-driven prognostics and health management (PHM) approaches for Indian Railways locomotives using historical operational and sensor datasets acquired from the REMMLOT monitoring platform. The investigation aimed to identify fault-prone subsystems, evaluate the predictive potential of machine learning algorithms, and characterise the stochastic nature of locomotive fault behaviour. Exploratory analyses revealed pronounced right-skewed fault distributions, with ventilation, auxiliary converter, and traction related faults emerging as dominant categories, particularly under high-traction, power-intensive conditions. In comparison to linear regression, ensemble models were found to be more effective in nonlinear dependencies between energy consumption and traction parameters across operational modes of powering, propulsion, and regenerative braking. Classification experiments employing resampling strategies such as random oversampling and undersampling achieved moderate discrimination, with gradient boosting models yielding the highest performance. Time-series forecasting employed a Seasonal Autoregressive Integrated Moving Average (SARIMA) configuration and a Long Short-Term Memory (LSTM) recurrent neural network provided complementary insights into the temporal dependencies in fault occurrences. Coupled with subsequent statistical randomness and entropy-based analyses revealed that the overall fault occurrence sequence was dominated by stochastic fluctuations, exhibiting high Shannon and permutation entropy values and only weak transient dependencies. The findings collectively indicate that locomotive fault dynamics are governed by semi-random processes with intermittent patterns rather than persistent deterministic trends. These insights have direct implications for railway condition-based maintenance (CBM): while predictive modelling can inform targeted interventions for high-risk subsystems, hybrid approaches combining data-driven inference with rule-based diagnostics are recommended to mitigate inherent stochastic variability.