Digital-Twin-Based Modeling and Fault Prediction Method for Industrial Robots

摘要

Fault prediction technology anticipates potential failures by examining operational data of equipment and developing deep learning models. These techniques are extensively applied in the realm of electromechanical equipment within the manufacturing industry to facilitate proactive maintenance, minimize downtime, and enhance equipment dependability. Nevertheless, within intricate engineering domains like six-axis industrial robots, traditional prediction techniques may encounter an overwhelming computational burden and exhibit subpar performance in terms of predictive accuracy. In this study, the authors introduce the Lagrangian Convolutional Long Short-Term Memory Neural Network (LC-LSTM) as a novel approach for predicting faults in industrial robots. By combining a convolutional neural network (CNN) with a long short-term memory network (LSTM), the LC-LSTM model can effectively analyze the temporal data collected from individual axes. This integration allows for independent prediction of the rotation angle and torque for each axis. The enhanced Lagrangian Neural Network (LNN) is not only applicable for elucidating the dynamics and kinematics of robotic arms but also for directly modeling the correlation between force and motion. Consequently, it can more accurately forecast parameters like force, velocity, and acceleration for individual axes of industrial robots. After training 2660 sets of data containing various faults, the LC-LSTM model demonstrates the capability to predict distinct faults for each axis of industrial robots with an average accuracy of 95.45%, an average recall ratio of 95.58%, and an average precision ratio of 94.8%. Additionally, this study introduces a digital twin model (DTM)for industrial robots, which combines predictive modeling with digital twin technology to facilitate real-time monitoring and accurate tracking of equipment operational status. This methodology enables more precise failure forecasts, consequently improving the general dependability and effectiveness of the machinery. Subsequently, a reinforcement learning model is suggested to adjust the twin's parameters, guaranteeing a significant level of coherence between the DTM and the actual system as time progresses. To evaluate the dependability of the DTM, we confirmed its ability to autonomously update parameters. The state of the twin model, trained on 200 datasets, is observed during different fault incidents, achieving an average accuracy of 91.4%.