A time–frequency representation guided deep domain adaptation method for aircraft EMA fault diagnosis under class imbalance and variable working conditions

Abstract The complexity and variability of operating conditions in electromechanical actuators (EMAs) often cause significant domain shifts in monitoring data, which severely degrade the performance of conventional fault diagnosis models. Developing robust cross-domain diagnostic approaches is therefore essential for reliable industrial applications. To address this challenge, this study proposes a time–frequency representation guided deep domain adaptation (TFRGDDA) framework. The proposed method integrates time-domain and frequency-domain representations through a tailored encoder incorporating 1D spectral convolution within a ResNet-based architecture. An auxiliary decoder is introduced to enhance latent feature representation learning. To mitigate distribution discrepancies between domains, the energy-based sliced Wasserstein distance is employed for domain alignment. Furthermore, an adaptive generalized linear unit activation function is incorporated to improve robustness under class-imbalance conditions. Extensive experiments conducted across multiple domain adaptation tasks on both a public NASA EMA dataset and a self-developed bearing testbed demonstrate that the proposed method consistently achieves superior diagnostic performance compared with several state-of-the-art approaches. The results verify its strong generalization capability and robustness under diverse operating conditions. Overall, the proposed TFRGDDA provides an effective and reliable solution for cross-domain fault diagnosis in EMAs.