Deep Learning-Based Automated Optical Inspection System for the Additive Manufacturing of Diamond Tools

Cutting tools with orderly arranged diamond grits using additive manufacturing show better sharpness and longer service life than traditional diamond tools. A retractable needle jig with vacuum negative pressure was used to absorb and place grits in an orderly arranged manner. However, needle hole wear after a long service time could not promise complete grit adsorption forever. This article proposed an improved YOLOv5s to detect the adsorption status of diamond grits on pinholes to maintain the planting rate of diamond grits in each matrix during the additive manufacturing process. First, the added detection head extracts higher level semantic information. Second, depthwise separable convolution + batch normalization + sigmoid linear unit modules containing depthwise separable convolutions (DSC) are used instead of convolution + batch normalization + sigmoid linear unit to reduce the number of parameters. Introducing DSC into the Bottleneck1 module results in faster computational speed than introducing bottleneck. Finally, coordinate attention is added at appropriate locations to improve detection accuracy. The improved YOLOv5s achieves an average 19.6% reduction in both parameters and floating point operations per second. The inspection system performance was validated by collecting data on a large number of vacancies and worn vacancy pinholes. Compared with the original YOLOv5s, the detection time for a layer of diamond grits with the system based on the improved YOLOv5s model decreased from 6.35 to 5.06 ms, and the detection accuracy was higher than 98%. When the absorption rate was detected below 95%, a redo command was given. The equipment has been in continuous operation for 1 year, and the vacancy rate of diamond grits in the orderly arranged diamond green segment produced by this additive manufacturing equipment is less than 5%.