Earticle

본 연구는 Three.js 기반의 3D 이미지를 통해 입체적인 부품에 대한 검사 및 상태 확인을 위한 목적으로 수행 하였으며, 실증을 위해 SMT 공정 중 PCB에 장착된 부품의 납땜 상태를 확인하는 품질 검사를 가정하여 연구를 진행하 였다. SMT 품질 검사 과정에서 PCB 기판에 장착된 부품과 납땜 상태를 3D이미지로 변환하여 360도 회전이 가능하도 록 시각화함으로써 2D 이미지에서 확인하기 어려운 납땜 상태를 확인 할 수 있도록 하고자 하였다. 여러장의 2D 이미 지를 통하여 3D 이미지를 생성하는 기존 방법들 중 사진측량 소프트웨어(3DF Zephyr), 논문 기반 기법(Instant NeRF, Gaussian Splatting), 깊이 맵 기반 단일 이미지 재구성 방법 등을 사용하여 보았으나 이 방법들은 성능 및 품질 측면에서 제안한 방법에 미치지 못하였다. 본 연구는 입체적인 부품에 대한 품질 검사에 보조 수단이 되는 방법을 제안하였으며, 이를 통해 입체적인 부품에 대한 품질검사에서 신뢰성을 향상시킬 수 있을 것으로 기대한다.

This study was conducted for the purpose of inspecting and confirming the status of three-dimensional components using 3D images based on Three.js. For verification, the study was conducted assuming a quality inspection to check the soldering status of components mounted on a PCB during the SMT process. During the SMT quality inspection process, the components mounted on the PCB and the soldering status were converted into 3D images and visualized so that they could be rotated 360 degrees, thereby enabling the confirmation of the soldering status that is difficult to confirm in 2D images. Among the existing methods for generating 3D images using multiple 2D images, photogrammetry software (3DF Zephyr), paper-based techniques (Instant NeRF, Gaussian Splatting), and depth map-based single image reconstruction methods were used, but these methods fell short of the proposed method in terms of performance and quality. This study proposed a method that can be an auxiliary means for quality inspection of three-dimensional components, and it is expected that this will improve the reliability of quality inspection of three-dimensional components.

본 논문은 딥러닝을 기반으로하여 SMT 공정 중 스크린 프린터를 통과한 PCB를 확인하여 솔더크림이 지정 된 위치에 일정한 높이로 도포되지 않은 경우 불량 PCB로 간주하고 해당 PCB의 불량 위치를 경계상자로 설정하 도록 하는 딥러닝 적용 불량검출 AI엔진 시스템 설계 및 개발에 관한 연구이다. 솔더크림 도포 불량 유형은 자주 발생되는 7가지로 구성하였으며 실증 업체로부터 불량 PCB를 수집하여 촬영하고 데이터셋을 준비하였다. 학습에 사용된 데이터셋은 Train, Validation, Test의 그룹으로 나누어 RT-DETR 모델을 기반으로 훈련시켰다. 연구 결 과 목표로 삼은 7개 유형의 결함이 성공적으로 탐지됨을 확인할 수 있었다. 이후 추가적인 현장 실증 및 고도화를 통하여 SMT공정 중 스크린 프린팅 결과물을 대상으로 실시간 불량을 식별해 내는 AI머신비전 시스템으로써 역할 을 크게 할 것으로 기대된다.

This paper is about the design and development of a defect detection AI engine system based on deep learning that checks a PCB that passed through a screen printer during the SMT process, and if the solder cream is not applied to a specified location and at a certain height, it is considered a defective PCB and sets the defective location of the PCB as a bounding box. The types of solder cream application defects consisted of seven frequently occurring types, and defective PCBs were collected from a verification company, photographed, and a dataset was prepared. The dataset used for learning was divided into groups of Train, Validation, and Test and trained based on the RT-DETR model. As a result of the study, it was confirmed that the seven types of defects targeted were successfully detected. In the future, through field verification and advancement, it is expected to play a significant role as an AI machine vision system that identifies defects in real time for screen printing results during the SMT process.

Face the process yield rate improvements of motherboard, although general enterprises finish deployment goal of each functions by overall quality managements, through quality improvement methods, industry engineering methods, plan-do-check-act (PDCA) methods and other improvement solutions, but it is only can be improved partially and unable to enhance the yield rate of product to the target. It only can takes one step ahead to enhance the process yield rate of motherboard with six sigma ($6{\sigma}$) overall DMAIC process and tactics. This research aimed to use six sigma quality improvement tactics by DMAIC systematic procedure and tactics, and find the key factors that effect to the process yield rate of surface mount technology. It also identified the keys input and process and output index to satisfy customer requirements and internal process index. The results showed that the major effective factors by fishbone and process failure modes and effects analysis (PFMEA). If the index of input and output that can be quantified, the optimum parameter can be found through design of experiment to ensure that the process is stable. If the factor of input and output that cannot be quantified, we found out the effective countermeasure by Mind_Mapping, make sure whole processes can be controlled stably, to reach the high product quality and enhance the customer satisfaction.

3-Dimensional System in Package (3-D SiP) structure (Amkor calls it mPossum-molded Possum) using double side Surface Mount Technology (SMT) and double side molding was evaluated in order to achieve small/thin form factor as well as good functionality by integration and double side layout. As the new platform on laminate substrate basis, molding process was challenge in mold flow balance at top and bottom side and package warpage control over the overall assembly process. There were two types of different molding process evaluated with 1) 1-step molding which was done at both side at the same time and 2) 2-step molding which was done at the conventional molding process twice. Mold simulation helped to narrow down the material selections and parameters available before actual sample build. There were many challenges for this first trial in design/ parameter and material types but optimized them to enable this structure.

The surface mounting technology of 0402 electric chip part is necessary to fabricate a high density and multi-functional module, but there is a limitation of the technology, like as a bridge and self-alignement. This work estimated SMT processing factors of 0402 chip. To obtain optimum SMT process, we evaluated effects of stencil thickness, shape of hole on printability and mountability. Printability shows best results under the thickness of $80{mu}m$ with circle hole shape and 90% square hole shape. In case of chip mounting process, chip mis-alignment and bridge was occurred rarely in same conditions. In more thin stencil thickness, $50{mu}m$, strength of 1005 chip parts was poor, because of amount of printed solder was insufficient.