Background <p>The harmonic reducer, as a key transmission component of welding robots, directly affects the accuracy and stability of the robot’s motion trajectory, which in turn influences the quality and efficiency of hull welding. Due to the long-term operation of the harmonic reducer in harsh and complex welding environments, accurate in-situ fault diagnosis becomes particularly important.</p> Methods <p>In this paper, a small sample in-situ fault diagnosis method for a harmonic reducer of a welding robot utilizing Adaptive Laplace Wavelet Decomposition (ALWD) and Multi-Scale Attention Mechanism (MSA) is proposed. The method first utilizes ALWD to decompose the vibration signal from the harmonic reducer. Then, the decomposed signal components are fed into a dual-path convolution, and the multi-scale attention mechanism extracts critical information from a small number of samples of complex in-situ signals with multiscale features. The weighted and fused multiscale signals are then input into global average pooling (GAP) for dimensionality reduction and fault detection. Comparative ensemble experiments conducted on real industrial robot harmonic reducer datasets and publicly available datasets demonstrate the excellent performance of the proposed method.</p> Results and Conclusion <p>The experimental results show that ALWD-MSA achieves a fault diagnosis accuracy of 99.32% even with a small number of samples, which is significantly higher than other comparison methods. This suggests that the method has great potential for application in real-world shop floor production.</p>