Otomatik Optik İnceleme Cihazlarının Tasarımı ve Baskı Devre Kart Kusurlarının Sınıflandırılması
Özet
Detection of card errors that may occur in Printed Circuit Board (OCB) production lines becomes more difficult day by day due to the increasing complexity of the board, the increase in the number of circuit elements used and the decrease or change in size, and it cannot occur at the desired speed. At this point, fast and reliable error detection and a successful detection rate with the correct use of signal processing methods are the requirements expected from such devices and software. In this thesis, it is primarily planned to develop a device that can be used in the field and is as reliable and accurate as portable and fixed Automatic Optical Inspection (AOI) devices. In commercial and scientific terms, such a portable device is not available in the sources and in the market. However, in artificial intelligence-based systems, this success depends on the breadth and quality of the database. Within the scope of the thesis study, image sets were collected and a data bank was established for four PCB errors (missing solder, slipped circuit element, short circuit and lifted circuit element). Obtaining qualified data containing new errors is far from always achievable. For this reason, even if the initial success rate is low, a system has been designed that increases the success rate in the total data process and continues to update and learn its information in case of missing or incorrect detections. In the system defined as a user in the loop, user information is included in the system as feedback and a continuous training infrastructure is designed to increase the error detection success rate of the learning process. In this thesis, an interface for analyzing and tracking images was also designed, and a model training and error detection infrastructure was developed with deep learning-based algorithms. The success of error detection is increased by combining the information known by the system and the information gained through user feedback in a loop. Training cycles were used in compound and discrete networks for the models, and with a constant comparison between them, the most successful network was instantly selected for error detection. With an example cycle, the composite model showed superior success compared to the discrete model in short circuit error and achieved a success rate of 97.64 percent. In case of missing solder error, the discrete model showed a more successful result with 77.77 percent and was preferred in the decision-making unit. In the designed system, the trained network is more successful in the errors it has learned and adapts to new errors by using the user requests in the new errors it has learned over time and maintains the speed element that is important in the production line with the YOLO algorithm that provides fast error detection.
