Time Series Prediction of Weld Seam Coordinates for 5 DOF Robotic Manipulator Using NARX Neural Network

In general, welding is a process in which two workpieces are joined together. The edge interface of the two halves are called weld seam. The main scope of this paper is to perform prediction analysis of 3D weld seam coordinates based on Non-Linear Auto Regressive with Exogeneous Input (NARX) Neural Network using various training functions and training ratios. Because developing a model for such complex processes using analytical techniques is time-consuming and prerequisite knowledge of the process is needed. Training NARXNN with the appropriate combination of learning rate, training-testing ratios, momentum coefficient and training function for the prediction of robot coordinates is a challenging task in Neural Networks. This work investigates Gradient Descent based Back Propagation, Scaled Conjugate Gradient method, Resilient Back Propagation, Levenberg–Marquardt algorithms in determining the 3D coordinates of weld seam. The proposed work compares the training algorithms based on Mean Absolute Error (MAE), Mean Square Error (MSE), and Root Mean Square Error (RMSE) for the real-time experimental data of weld shape. Experimental analysis is performed using the data obtained from real-time weld seam detection using 5 DOF robotic manipulators.