TY - JOUR
AU - Nenadic, Nenad
AB - [
                                                       
                    
                    Summary
                                                       
                    
                    The remanufacturing and reusing of printed circuit boards (PCBs) is an important component of the circular economy. Current practices on the remanufacturing floor employ several manual processing steps, including identification of the PCB type, localization and degradation assessment of various components, and keying data entries into the system. The repetitive nature of the manual processing steps places a heavy burden on technicians, who tire, make mistakes, and introduce subjectivity into the assessment process. Furthermore, these tedious tasks make employees unhappy, which leads to high turnover. Machine learning has become state‐of‐the‐art for automating inspection tasks but typically requires a large amount of labeled data. We describe the process of introducing machine learning and computer vision for two tasks associated with the remanufacturing process: 1) part number identification and 2) localization of components with an assessment of their degradation. The components selected for the visual inspection were light‐emitting diodes (LEDs) because the traditional assessment was based on manual visual inspection. The solution incorporated commercially‐available solutions, newly‐trained models, and novel approaches for relaxing requirements for machine learning development, all integrated into one development environment. Specifically, the part‐number identification solution leveraged the Google Cloud Vision API for extracting character strings from images. The solution for the degradation assessment involved two steps, localizing components and classifying their health. The localization of components used a novel approach that employed classical, deterministic image processing and machine learning. The localized LED sub‐images were classified using a custom‐trained deep‐learning model. Because labeling can also be time‐consuming and expensive, we propose a localization scheme that leverages the efficacy of deep learning and significantly reduces the time required to label a dataset. LED localization and assessment performance showed a better than 97% detection rate on the validation data for a specific PCB when the false detection rate was held below 5%. In addition to software development, we explored and discussed trade‐offs related to different options for image captures, industrial cameras, and smart devices relative to the use of the captured images.
                                                    
                
                ]
TI - Technology Innovation for the Circular Economy : Recycling, Remanufacturing, Design, Systems Analysis and Logistics: Image‐Based Methods for Inspection of Printed Circuit Boards
DA - 2024-01-15
UR - https://www.deepdyve.com/lp/wiley/technology-innovation-for-the-circular-economy-recycling-RE0fH6Y39W
DP - DeepDyve
ER -