Högskolan i Skövde

Due to the advance of the technologies, most car manufacturing companies are using robots for the manufacturing process. But, the defect detection is mostly done manually, which is not accurate and efficient. The purpose of this project is to find the best machine learning approach which can detect the unexpanded sealing of the battery lids of the cars. The data for the analysis consists of 56 video files of applying sealing on 56 different battery lids, which were all expanded, which were converted into 32,728 images. The battery lids are extracted from the images using the Template Matching method in the OpenCV. After removing the noisy images, the images are reduced to 19,535. Both unsupervised and supervised methods are used for the analysis. Different thresholding approaches are investigated on both the methods, to find the best thresholding approach to threshold the sealing from the battery lid. The Random Forest (RF) and Convolutional Neural Network (CNN) are the supervised models used for the analysis. Before building the models, 9,846 images were labelled as ‘Expanded’ and 9,689 images as ‘Unexpanded’, by comparing images using the Structural Similarity Index Measure (SSIM) metric. Even though, both the RF and CNN models give similar high accuracy results, the RF using the original images performed little better than the other models with an accuracy of 0.99693. It is concluded that both the methods give high accuracy results, without applying any thresholding methods. From the thresholded methods, Gabor filter performed little better than other methods, even though all the methods give high accuracy results. Also, the RF performed little better than CNN, even though both models produced high accuracy results.