FFT-Informer: A Hybrid Frequency–Time Domain Model for Track Irregularity Identification

Abstract

Assessing track condition quality is essential for ensuring operational safety and efficiency in high-speed railways. Traditional track irregularity detection relies on dedicated inspection vehicles, which are costly and limited by low inspection frequency. To address these limitations, this study proposes FFT-Informer, a novel deep learning model integrating the Informer architecture with fast Fourier transform (FFT) for hybrid time–frequency domain analysis of spatial series data. Through comparative evaluation with Informer, Transformer, and Autoformer, FFT-Informer demonstrates superior performance in three key dimensions: its sparse attention mechanism achieves focused feature extraction while reducing redundant parameters; at the optimal sequence length (480, matching the 120 m wavelength), it attains the lowest prediction errors (mean absolute error/root mean square error) and near-perfect correlation (0.99) with true amplitude fluctuations; and power spectral density analysis confirms its exceptional accuracy in replicating spectral energy distribution, particularly for mid-to-high wavelengths. With minimal output variability ( σ = 0.03 – 0.04 ), FFT-Informer outperforms existing models in consistency and interpretability, offering a robust solution for real-time track condition monitoring. This work bridges the gap between time domain sequence modeling and frequency domain analysis, advancing data-driven approaches for infrastructure health management.