On the Importance of Pre-Training and Spatial Context for Corrugation Detection from In-Service Metro Trains

We adapt a pre-training framework for corrugation detection using axle-box acceleration signals from commercial trains on the Milan metro network, where the speed, load, and wheel state vary continuously across different runs. Three novel aspects of our approach are the use of: (i) spectrogram pre-processing to account for speed variations, (ii) data augmentation techniques that are designed specifically for spectrograms, and, (iii) a Transformer-based architecture for aggregating contextual features of the track. We validate the approach using a previously developed pipeline achieving a recall of 97.3% on severe corrugation, confirming applicability to the difficult setting of the Milan metro line. Furthermore, when we move to an operationally realistic setting, raising the minimum corrugation detection threshold to 40 micrometers, and extending the feature representation to include a sequence of consecutive spectrograms (covering 6 meters of track), the pipeline achieves an area-under-the-curve of 0.978 and a recall of 94.3% with a false-positive rate of only 1.4% on healthy conditions. Zero-shot evaluation on a held-out section yields a mean Spearman correlation of 0.62, confirming the model generalises well to unseen track geometry.