Transfer learning in LSTM autoencoders for rail anomaly detection across diverse operational conditions

In this study, transfer learning was applied to a Long Short-Term Memory Autoencoder for the detection of rail anomalies across diverse operational conditions. Separate models were trained initially on Laser Doppler Vibrometer data collected from two distinct sites: the Transportation Technology Center in Pueblo, Colorado, and a BNSF railway in Cleburne, Texas. Each site was characterized by unique environmental setups and train speeds (e.g., 10 mph, 20 mph, 30 mph). A cross-dataset evaluation was performed, testing each model on data excluded from its training. Accuracy was observed to decrease in these evaluations, particularly when speed differences exceeded 10 mph or experimental setups differed. Transfer learning was then implemented, retraining portions of the pre-trained models with the new datasets. The results showed that performance has significantly improved after the implementation of transfer learning. It indicates that transfer learning is essential for ensuring accuracy and adaptability across varying conditions for practical applications.