Visual Inspection System for Piston-Ring Parts via Machine Learning

Abstract

Development of a system that can automatically detect appearance defects of piston-ring components of the engine cylinder caused during the coating process. In the proposed system, piston-rings are sent from the feeder to conveyor belt, and an image captured by camera. Subsequently, the image is cut along with the shape of ring into small images. A convolution neural network (CNN) model to classify which piston-ring is a normal and anomaly. Finally, a robot arm is utilized to remove the anomaly piston-ring from the conveyor belt.

In our previous experiment, when a GPU-based computer was used to process images, the system could achieve approximately 90-100% accuracy based on the type of defects. To reduce the costs of system, we study single-board computer (SBC) with Google Edge TPU USB Accelerator to classify images, which exhibits good potential to replace GPU-based processing. Furthermore, this paper also proposes some approaches to improve processing speed when using proposes low-cost SBC platform.