Vision-Based Measurement System for Automated Quality Inspection of CNC Turned Components

Abstract

Manual inspection in the high-precision manufacturing industry often struggles to meet the demands of high-volume production and strict dimensional tolerances. This paper presents a vision-based measurement system for the automated inspection of CNC-turned parts. The methodology integrates a custom-built, isolated backlit inspection station and sub-pixel edge detection algorithms developed in Python/OpenCV. Optical calibration was configured through a 22 full factorial design of experiments, yielding a ground sample distance of 0.0323 mm/px and a mean reprojection error of 0.093 px. Validation against a coordinate measuring machine on a 108-part batch showed a root mean square error below 0.0260 mm for all dimensions. Error sources were quantified through stability tests, a gauge R&R study, and an uncertainty budget. Overall classification accuracy was 99.1%, 95.4%, and 92.6% at tolerances of ±0.0762, ±0.0508, and ±0.0254 mm, respectively, with sensitivity of at least 97.4% and positive predictive value between 76.2% and 92.6%. The system inspects 100% of the parts, evaluating five dimensions in 4.72 s per piece, over 81% faster than manual inspection, with full digital traceability.