Machine Learning Driven Quality Inspection for Cell Phone Assembly Process of a Manufacturing Company in Ethiopia
No Thumbnail Available
Date
2023-06
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
In the process of cell phone component production, assembly and transportation, various types of defects will inevitably occur. Quality inspection in assembly process in time can avoid waste of manpower and material resources. In addition, it can monitor the production process, ensure high-quality products and reduce production costs. In this thesis work, the quality inspection problems in the cell phone assembly line of a manufacturing company in Ethiopia are studied. Most of the inspection techniques used in the company are manual based with inspectors check for defects using their eyes. The main challenges in the inspection process are detection accuracy and speed. In a one year inspection report, from the total defects found at the final inspection, more than 31 % are missed from receiving inspection. The manual visual inspection is also slow. For example, it takes at least three seconds to inspect a phone screen surface defects for a well-trained inspector. The inspection problems are categorized into five as image based, sound based, QR and barcode inspection, other functional tests and weight test. Based on the nature of the inspection problems and literature review, the image based and sound based problems are found to be best solved by machine learning (ML) driven solutions. Convolutional Neural Network (CNN) class ML models with their development procedure are proposed for image based problems while Recurrent Neural Network (RNN) and CNN class models are proposed for the sound based problems with the development procedures. To demonstrate this, five CNN class ML models are developed for phone screen inspection and their performances are compared. The dataset used for training and testing of these models is prepared by combining defect free images taken from the company with mobile phone screen surface defect segmentation (detection) dataset (MSD). The YOLOv8 model that is trained from scratch results in best performance. The validation accuracy is found to be 97.5 % and it takes 25 milliseconds to inspect surface defect of a phone screen. This result suggests the potential of ML driven approaches for solving the inspection problems of the case company. By taking defective and defect free images of phone screen from the company, a dataset is prepared to train a pre-trained YOLOv8 model. The validation accuracy is found to be 100%, but this model is prone to error threat. This is because the size of the dataset used is very small.