A domain generalization approach with multisource fusion for antenna drive system fault diagnosis under unknown time-varying speeds

Driven by advances in artificial intelligence, deep learning has been extensively applied to fault diagnosis in rotating machinery. However, collected fault data often fail to cover the entire range of operational speeds. This limitation presents significant challenges for intelligent diagnostic algorithms when identifying faults under unknown and time-varying operating conditions. To address this, a domain generalization method based on multisource information fusion (MSIF) is proposed. The method characterizes faults by integrating high-frequency acoustic emission signals with low-frequency vibration signals to capture complementary information across frequency bands, while incorporating real-time rotational speed information. Specifically, inspired by Vision Transformer architectures, a fine-grained MSIF model is developed. The model’s robustness to speed variations is enhanced through a multirotational-speed-scale feature extraction module and a multirotational-speed-scale feature aggregation module. Additionally, deep fusion of heterogeneous information is accomplished via a fusion attention mechanism. Experimental validation utilized five classes of fault data collected from the azimuth drive system of a fully movable 2.4 m aperture reflector antenna, under both step-varying and sinusoidal-varying speed conditions. Eight challenging diagnostic tasks are constructed. Results demonstrate that the proposed method achieves optimal diagnostic performance, with an average accuracy of 90.27%, representing a significant improvement over comparative models and confirming its strong generalization capability under unknown speed conditions.