LiTFSE_MLNet: A lightweight frequency-domain multi-label network for transient-robust compound fault diagnosis of three-phase motors

To address the performance degradation of CNN-based fault diagnosis methods under transient operating conditions, this paper proposes LiTFSE_MLNet, a lightweight frequency-domain multi-label network for transient-robust compound fault diagnosis of three-phase motors. Under transient conditions, fault characteristic frequencies drift with rotational speed, causing steady-state trained models to fail despite achieving over 99% accuracy on steady-state data. The model employs a Short-Time Fourier Transform (STFT) frontend to capture time-frequency dynamics, followed by a CNN backbone equipped with a frequency-axis Squeeze-and-Excitation (FreqSE) module that enhances discriminative spectral components while compressing the temporal dimension and reducing computational cost. The overall architecture provides wide-band coverage for frequency drift under transient conditions. The FreqSE module works collaboratively with a multi-label classification head to ensure robustness under speed-varying conditions. Experiments on the Tsinghua Three-Phase Motor Dataset under a steady-to-transient generalization setting show that LiTFSE_MLNet achieves 91.74% overall transient test accuracy, outperforming the second-best method by 6.27% points. The training-to-testing gap is only 7.50%, the smallest among eight compared methods. The average exact match rate for compound fault combinations is 91.96%. The collaborative architecture effectively mitigates cross-condition performance degradation, without relying on individual module optimization. With its lightweight design and transient robustness, LiTFSE_MLNet offers a deployable solution for vibration-based condition monitoring in industrial motor systems operating under transient conditions.