MULTI-SENSOR 3D INSPECTION SYSTEM FOR ENHANCED MANUFACTURING QUALITY

The current research trends in the manufacturing industry focus on improving the manufacturing processes to enhance the quality of the finished part. These parts must be subjected to quality control tests based on the manufacturing requirements, as they are manufactured. Vision-based inspection is one of the most popular inspection systems in large-scale manufacturing industries. However, this inspection technique loses valuable information due to the lack of depth data and its heavy reliance on a fixed light source. Therefore, an online light-independent scanning system that can benefit from the depth data is needed. In this study, we employed a modular multi-camera quality inspection system for real-time quality monitoring in manufacturing environments. We designed the system to identify and classify surface-level defects and verify the dimensional qualities of manufactured specimens. By applying 3D point cloud data processing techniques, we identified surface-level defects, overcoming the limitations of 2D image-based methods. In our research, we introduced a novel method combining Local Surface Variation (LSV) with Density-Based Spatial Clustering of Applications with Noise (DBSCAN), to achieve improved automatic isolation of defective points on part surfaces compared to conventional clustering algorithms. We demonstrated these methods and experiments using a sample assembly, designed and printed in-house. We classified defects based on predetermined criteria identified from the geometric features of defect clusters. Additionally, we conducted a geometric analysis of the part assembly to assess dimensional accuracy. We apply these methods to a real-world specimen to validate our developed methods. The novelty of this research lies in the utilization of a combination of advanced methods to identify defects and dynamically adjust process parameters of the combined algorithm, enabling autonomous defect detection and classification for any type of sample without requiring manual intervention. The classification algorithm we developed does not require training since it's based on predetermined criteria, making it easier to set up the system for different manufacturing scenarios. The dimensional analysis yielded less than 5% error when compared to the ground truth value, which is an acceptable error range and is one of the most promising aspects of the algorithm. Future work will focus on adapting these techniques for broader additive manufacturing applications and integrating an automated sorting system to separate defective parts.