Task failure prediction for wafer-handling robotic arms by using various machine learning algorithms

Abstract

Industries are increasingly adopting automatic and intelligent manufacturing in production lines, such as those of semiconductor wafers, optoelectronic devices, and light-emitting diodes. For example, automatic robot arms have been used for pick-and-place workpiece applications. However, repairing automatic robot arms is time-consuming and increases the downtime of equipment and the cycle time of manufacturing. In this study, various machine learning (ML) models, such as the general linear model (GLM), random forest, extreme gradient boosting, gradient boosting machine, and stacked ensemble, were used to predict the maximum Cartesian positioning shift (i.e. the maximum eccentric distance) in the next handling time period (e.g. 1 min). A charge-coupled-device-based fault diagnostic system was developed to measure the critical positions of the robotic arm when transferring wafers. A novel data augmentation method was used to determine the correlation parameters in the dataset for the ML models. The prediction error for each algorithm was determined using the root mean square error (RMSE). The results revealed that the GLM exhibited the lowest prediction errors. The RMSEs of the GLM were 0.024, 0.032, and 0.046 mm for 3421 pickups, the last 1000 pickups, and 100 pickups, respectively, for the prediction target. Thus, the GLM is a promising model for predicting the task failure of wafer-handling robotic arms.

Precise 6DOF Localization of Robot End Effectors Using 3D Vision and Registration without Referencing Targets

“…Recent studies have investigated the problems related to position uncertainties of robotic arms end effectors and analyzed the life cycle of the robotic arm to improve the reliability of the wafer handling robot [2,3]. However, in general, planar translation of end effectors has been measured using only 2D imaging; with depth information missing, 6DOF localization of end effectors cannot be realized.…”

A Morphing Point-to-Point Displacement Control Based on Long Short-Term Memory for a Coplanar XXY Stage

“…The results demonstrated that the LSTM network could approximate the relationship between the input and output of a GNSS-integrated navigation system (INS) with high precision. In [ 14 ], a visual recognition system was combined with an artificial-intelligence machine-learning neural network to predict the maximum pick-and-place offset of a robot arm in the next minute. The developed LSTM model made predictions with high accuracy and reliability and met handling robot needs.…”