Few-shot fault diagnosis for industrial robot transmission systems via a prototypical time-frequency mixer

Fault diagnosis in transmission systems is essential for ensuring the safe and stable operation of industrial robots. However, real-world applications are often constrained by the scarcity of labeled fault samples. To address this challenge, a novel prototypical time-frequency mixer (PTFM) is proposed to diagnose few-shot faults in industrial robot transmission systems. Vibration signals collected from the industrial robot are first processed by a temporal modeling branch to extract time-domain features, and a Fourier transform branch to generate frequency-domain representations. The two branches are subsequently integrated into a latent space to capture both local and global patterns. The resulting representations are employed to compute class prototypes via a prototypical network. This facilitates the classification of query samples based on their distances to the support set. In this way, accurate diagnosis of both seen and unseen faults under few-shot conditions is achieved by effectively leveraging spectral and temporal information. The proposed PTFM was evaluated on an industrial robot experimental platform, with generalization tests further performed on two public machine fault datasets. The results demonstrate strong diagnostic performance across both seen and unseen classes. Compared to the state-of-the-art methods, the proposed PTFM achieves improved adaptability and robustness in few-shot fault diagnosis for industrial robot transmission systems.