Assembly torque data regression using sEMG and inertial signals

Abstract

Assembly torque monitoring is an important aspect of assembly process monitoring. Currently, in manual assembly lines, torque monitoring sensors are usually installed on assembly tools or assembly jigs. This approach is restricted by the working environment and assembly space. To this end, operators can use wearable devices containing sensors without being restricted by the tool and working spaces, with the added advantage of portability. To use wearable electromyography equipment to monitor the assembly torque, this paper proposes an assembly torque monitoring method using surface electromyography (sEMG) and inertial signals. Two torque regression models, one based on a heterogeneous-kernel-based temporal convolutional network (Het-TCN) and the other based on a two-stream convolutional neural network (CNN), are proposed and compared with CNN, long short-term memory (LSTM), and temporal convolutional network (TCN) models. The experimental results show that, compared with using sEMG signals alone, performing an assembly torque data regression using a combination of sEMG and inertial signals can significantly reduce the regression error. The proposed Het-TCN model has the best regression performance, with an average error of 3.31 N m an RMSE of 9.18 %, and a coefficient of determination R2 of 0.72. Thus, the proposed Het-TCN along with the application of both sEMG and inertial signals can help effectively perform assembly torque data regression, which can be used for assembly torque or operation monitoring.