CSTF-VM-FL: federated learning for turnout switch machine fault diagnosis via cross-scale time-frequency feature extraction and FedSpecSens aggregation

The turnout switch machine is a critical actuator in railway signalling systems, and its state directly affects train dispatching. Non-destructive testing methods are often used to assess its operational status, but data are distributed across different lines, making centralised sharing difficult. Additionally, the vibration signals of turnout switch machines are non-stationary, which complicates high diagnostic accuracy and generalisation in traditional models. To address these challenges, we propose a federated fault diagnosis framework that integrates short-time Fourier transform (STFT) with a cross-scale time-frequency refined vision mixer (CSTF-VM). The framework applies STFT to convert vibration signals into time-frequency spectrograms and uses CSTF-VM to enhance multi-scale feature extraction. A federated spectral sensitivity aggregation strategy dynamically adjusts weights based on client sample sizes and spectral sensitivity, improving global model performance under Non-IID conditions. Experimental results on real-world data demonstrate that the proposed method achieves 99.0% classification accuracy with ±0.5% standard deviation and superior inter-class separability in t-SNE visualisation, proving its effectiveness and robustness.