A multi-scale CNN-transformer hybrid network with parallel attention mechanism and local linear unit for cross-condition fault diagnosis

Domain shift caused by variations in rotational speed and load under cross-condition scenarios may distort the statistical distribution and feature representation of vibration signals, posing a challenge to the reliable deployment of intelligent fault diagnosis systems for rotating machinery. Existing multi-scale CNN–Transformer hybrid architectures generally do not explicitly consider the influence of such condition-induced domain shifts. To address this issue, a multi-scale CNN-Transformer hybrid method integrating a Parallel Attention Mechanism (PAM) and a Local Linear Unit (LLU) is proposed to enhance the stability of feature representations under varying operating conditions. In the proposed method, raw vibration signals are directly used as inputs, and a multi-scale CNN module is employed to capture transient impact features across multiple temporal scales. The PAM then adjusts feature responses in both channel and temporal dimensions through parallel attention branches, enabling adaptive feature reweighting to alleviate condition-induced statistical variations. Furthermore, a Transformer encoder embedded with LLU is adopted as the backbone to incorporate local structural modeling through depthwise separable convolution while preserving the global dependency modeling capability of self-attention. Experiments on three benchmark datasets (PU, PHM09, and CWRU) show that the proposed method achieves average accuracies of 80.15 %, 93.96 %, and 98.17 %, respectively, under unseen operating conditions, consistently outperforming several representative comparative methods. Additional ablation studies further verify the contribution of PAM and LLU to robust feature representation learning. Moreover, noise robustness experiments under multiple signal-to-noise ratio (SNR) levels demonstrate that the proposed method maintains stable performance under moderate noise conditions, highlighting its practical reliability in realistic industrial environments. Our code is available at https://github.com/caie8201/Multi-Scale-CNN-Transformer-Hybrid-Network.