Dynamic Estimation of Track Irregularities via Deep Learning

Monitoring track irregularities is critical for highspeed railway safety. Currently, track irregularities are mainly detected by Comprehensive Inspection Trains during operational gaps, which cannot meet the demand for high-frequency, low-cost monitoring. The current research trend involves the use of in-service vehicles' vibration response to estimate track irregularities. However, accurately deriving track irregularities from vibration response remains a key challenge due to their complex nonlinear relationships. We developed a deep-learning model based on inertial measurement unit data from onboard devices and track irregularity measurements from inspection trains. Our proposed FMCformer model uses vehicle-body vibration accelerations and angular velocities as inputs. It uses a Frequency Enhanced Block to strengthen frequency-domain features, multiscale convolutional kernels to extract features across various wavelengths, and a linear attention mechanism to capture global dependencies. The model outputs vertical track irregularities at wavelengths of 2-42 meters. Experiments on high-speed railway data demonstrate that the FMCformer outperforms state-of-the-art models.