Harmonic drives are a key component in many high-end systems and a potential weak point affecting operational stability. Thus, studying fault diagnosis methods for harmonic drives is of significant importance. In recent studies, the use of multi-sensor data for joint diagnosis has gained increasing attention, though most existing models still rely heavily on fault data. To address this challenge, this paper proposes a diagnostic model named Hierarchical Synergistic Fusion Network (HSFNet), which achieves accurate identification of harmonic drive faults under small-sample conditions. First, the model maps multi-sensor data into the time-frequency domain and dynamically recalibrates input weights to emphasize different sensor signals. Then, a multi-source features fusion module is designed to capture fault characteristics from receptive fields of different scales, enhancing interactions and fusion of both same-scale and cross-scale features to strengthen the model’s representational capacity. Finally, a classification projection head compresses features and outputs diagnostic results. Validation and comparative analysis on the collected dataset demonstrate the superiority of the proposed method.

fault diagnosis, multi-sensor, small samples, harmonic drive, hierarchical synergistic fusion network