A Lightweight Network Framework for Fault Diagnosis of Quadruped Robot Joint Modules

With the widespread deployment of quadruped robots across various application scenarios, they are required to adapt to diverse motion patterns when performing different tasks. This necessity imposes complex stresses and loads on the components within the robot's joint modules during movement, leading to a higher risk of wear and failure. To address these challenges, this study proposes a novel lightweight multi-scale neural network based on a dual attention mechanism and Depthwise Separable Convolution (DSconv), named Atten-MSDsN, for diagnosing faults in joint modules. This approach captures critical fine-grained features across the global scope of the signal through the attention mechanism, extracts multi-scale features from fault signals using convolution kernels of varying sizes, and reduces the number of parameters in the feature extractor via DSconv. Experimental analysis demonstrates that the Atten-MSDsN framework offers the advantages of being lightweight and robust.