HC-TSKNet: a high-channel attention fusion framework for inter-shaft bearing fault diagnosis under variable-speed aero-engine conditions with high-accuracy and high-robustness

Abstract Addressing the scarcity of data reflecting realistic variable-speed operating conditions in current research on aero-engine inter-shaft bearing fault diagnosis, and overcoming limitations in existing methods concerning time–frequency resolution, global cross-channel interaction capability, and robustness under variable speeds, we proposed a High-Channel Triple Attention and large selective kernel (LSK) Module Fusion Network (HC-TSKNet) characterized by high precision and strong robustness. The method first employs synchrosqueezing wavelet packet transform to transform the raw vibration signals into high-resolution time–frequency spectrograms. This effectively mitigates the spectral smearing issue prevalent in conventional time–frequency analysis such as continuous wavelet transform, significantly enhancing the clarity of discriminative fault signatures. The core innovation of HC-TSKNet lies in its synergistic integration of three key techniques: through an increased channel count, the model size is compressed and the per-sample inference time is reduced, while preserving a satisfactory level of diagnostic accuracy; utilizing a Triplet Attention mechanism to capture global cross-channel and spatial interaction dependencies, overcoming the limitations in cross-dimensional interaction inherent in conventional attention mechanisms like convolutional block attention module and squeeze-and-excitation network; and introducing a LSK module to adaptively weight local features extracted by deep convolutional kernels of different scales spatially, enabling dynamic focus on localized fault patterns under variable-speed conditions. Validation experiments conducted on an inter-shaft bearing fault dataset from aero-engine dual-rotor simulator test rig demonstrate that HC-TSKNet achieves exceptional diagnostic accuracy of 99.75% under variable-speed conditions, significantly outperforming mainstream methods including Swin-Transformer and recent state-of-the-art models in related domains. It is particularly noteworthy that the model successfully overcomes significant variations in signal characteristics induced by differences in sensor mounting positions, demonstrating robust adaptability to diverse sensor configurations. Ablation studies, hyperparameter experiment, and robustness validation further substantiate the effectiveness of its key modules, the superiority of its structural design, and its substantial potential for engineering applications.