Research on Fault Diagnosis of Robot Arm With Dynamic Simulation and Domain Adaptation

Abstract

The main challenges in the field of fault diagnosis of robot arms lie in the difficulties of acquiring fault data and ensuring model applicability. For a fault robot arm, the trained models typically only perform well on test data and cannot be effectively applied in practical scenarios. As a result, the time cost is very much to construct adequate fault datasets. The paper proposes a dynamic simulation method for obtaining fault data to address these issues. The motion feature of the arm with joint faults is replicated by the simulation software, thereby obtaining vibration signals in the fault mode as samples. Additionally, under the main framework of Deep Learning (DL) with an end-to-end feature extraction capability, a Stacked Continuous Wavelet Transform (SCWT) method is designed to visualize timing signals graphically based on the traditional wavelet transform. Furthermore, to enhance traditional DL performance, a dual-channel architecture for data fusion within DL is designed to enrich the feature space and improve fault-distinguishing ability. Lastly, a Domain Discriminator $G_{d}$ is designed to identify the upper bounds for differences between spatial distributions of simulated and actual fault data. By the domain discriminator, the feature distribution of target and source data is aligned, facilitating the transfer application of the simulation-trained diagnostic model on the actual fault. The proposed method is tested and evaluated using a self-constructed experimental data set. The results substantiate its effectiveness and superiority.

Keywords

• Robotic arm
• dynamic simulation
• data fusion
• transfer learning
• deep learning