Integration of Artificial Intelligence with Computer Vision on Surface-Mount-Device's Fault Detection System

Fong, Hao Nan (2022) Integration of Artificial Intelligence with Computer Vision on Surface-Mount-Device's Fault Detection System. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text FongHaoNan_Thesis.pdf Restricted to Registered users only Download (7MB)

Abstract

With better technology over time, electronic components like resistors and capacitors had smaller form factors and were known as surface mount devices (SMD). SMDs were known to have retained or even better performances (reduced energy losses) with a significant size decrease as compared to conventional through-hole electronic components. However, the small form factors of SMDs were also a reason for increased probabilities in production faults. Common industrial machines that detect these faults uses image comparison methods which utilizes good samples for every tested circuit board with SMDs. This leads to the industrial problem of flexible operations, hindering the quality assurance for personalized boards not meant for mass production or not of that stage yet. The feasibility of an AI-based solution for SMD fault detection systems over traditional image comparison methodswere questioned as AI solutions provide flexibility in detecting variants of SMD faults, thus allowing personalized circuit boards to be checked without any pre-produced good samples needed. Hence, the project serves as a case study to study the feasibility of said topic. Individual AI models were to be trained with the selected datasets to achieve a minimum prediction accuracy of 80% from training alone. Then, different solution designs were established where test images were inputs to different combination of AI models that were executed procedurally. The overall prediction accuracy was the next to be obtained through specific calculations before the best theoretical and actual prediction accuracy was compared with the traditional image comparison method. The project kicks off with the training of individual YOLO and KERAS CNN models with training image sets until the prediction accuracy shows validation to be used for testing image sets. With all individual AI models trained, the testing images were then utilized with the AI models correspondingly and the written program codes generates two conclusions: the best prediction accuracy for the proposed multi-model solution design stands at 74.90% whereas the best prediction accuracy for the single-model solution design were low at 59.26%. The limitations of the project was that several forms of SMD faults were not covered due to the provided dataset’s true intention and the project’s outcome could be improved as multi-model solutions could be enhanced by having the individual AI models to share information among each other for scope narrowing in SMD fault identification. While the achieved accuracy for AI-based multi-model solution does not stand equal with traditional image comparison methods of near-100% quality assurance, the project was concluded with the improvements of accuracy as future works and that the flexibility of solution were the uniqueness to modern SMD fault detection systems.