Predictive maintenance of railway suspension systems using multi-level time–frequency vibration analysis

Railway suspension systems are critical for ride quality, operational safety, and track maintenance. Degradation in primary suspension components, such as reduced stiffness or damping, can cause excessive vibrations, higher track wear, and passenger discomfort. Traditional maintenance strategies, including time-based or corrective approaches, often fail to detect early-stage deterioration and can lead to unnecessary replacements or service disruptions. This study proposes a predictive maintenance framework leveraging multi-level, multi-axis vibration data and machine learning models to classify suspension degradation on curved tracks. Acceleration signals are collected at wheelset, bogie, and car body levels along longitudinal, lateral, and vertical axes. Time-frequency features are extracted using Fast Fourier Transform (FFT), while zero-padding standardizes raw time-domain signals. Five deep learning architectures consisting of CNN, LSTM, GRU, CNN-LSTM, and CNN- GRU are trained and evaluated for classification accuracy, convergence speed, and computational efficiency. Results indicate that CNN with zero-padded time-domain input achieves the highest accuracy (0.98) and fastest convergence, outperforming recurrent and hybrid models. Sensitivity analysis highlights that Z-axis vibrations from bogie and car body provide the most informative data. The proposed approach enables early fault detection, reduces sensor requirements, and supports real-time condition monitoring.