In order to control a robot, high speed optical cameras are required, to help generate a feedback signal for a closed-loop control system. However, one of main challenges to build a highly accurate process is working with the vast amount of image data.
Cross domain transfer learning significantly improved the performance of the robot. It could also be further developed to predict and locate defects that occur during the robot assisted AM process.
- Run experiments to generate the image data
- Develop traditional deep learning model (which demonstrated the infeasibility of working with a very large dataset)
- Take advantage of the Google VGG model and streamline it for this particular purpose
- This model works with very little data, provides high accuracy and fast response speeds
- Connect this model to the robot control system an validated the results