Robot dynamics-based cable fault diagnosis using stacked transformer encoder layers

Abstract

Industrial robots play a vital role in manufacturing systems, engaging in tasks such as welding, painting, and assembling. To prevent catastrophic manufacturing stoppage, it is essential to diagnose faults in control cables of robots in time. This paper proposes a hybrid fault diagnosis method that integrates a robot dynamic model with deep learning-based fault diagnosis to classify the severity of cable faults. Specifically, the proposed method incorporates both the measured torques obtained from measured currents and nominal torques from the dynamic model, achieving robust cable fault diagnosis under varying operating conditions. The measured cable current signals that contain the fault information are used to calculate the joint torques, and a robot dynamic model is used to obtain the nominal joint torques using joint angles and angular velocities. Subsequently, a stacked transformer encoder-based classifier is constructed with the obtained torque disparities as inputs and fault severity probabilities as outputs. Experimental results validate that the proposed fault diagnosis method provides higher accuracy compared to existing methods, highlighting the efficacy of integrating a dynamic model with learning-based fault diagnosis. Furthermore, we conducted a quantitative and qualitative comparison between our proposed method and other recent fault diagnosis methods.