عليرضا يزداني جو

An operator oversees the assembly of connecting conrod bearings in an EF7 automobile engine block, and a slight error in this process will cause serious damages to the engine.Engine bearing is one of the important components in all types of internal combustion (IC) engines, that any surface defects can reduce its performance and cause to other failures.Therefore, it is necessary to establish an automated inspection system in engine production line, instead of traditional inspections that mainly rely on human sensesTo this end, we first collect a dataset consisting of 2000 images from a 4-cylinder engine block. This dataset comprises 16 categories of images, which are used for defect detection in the assembly of engine bearings, and is publicly released to the community.A new complete dataset consisting of 500 images obtained directly from engine bearings in the form of images with different surface defects was also introduced in this study.We then employ three machine vision techniques, including classification, object detection, and segmentation to automatically monitor the assembly process. We leverage techniques such as transfer learning and data augmentation. Moreover, to boost the performance, we provide our neural networks not only with original training images, but also with explicit shape/structure signals created using neural-network-based filters such as DexiNed and the phase stretch transform filters. Performance eva‎luations show that the proposed approach achieves a very high accuracy, which makes it practical for adoption in production assembly lines.The semantic segmentation technique was employed for processing all bearing images using convolutional neural network with MobileNet architecture which was more appropriate for real-time applications than other ones, because of its fewer parameters. The model performances were reported under different training datasets. Without augmentation on the training images, an Intersection over unio‎n Score of 84.5% was achieved by MobileNet. Using Sobel filter and Data Augmentation technique, we obtained an IoU-Score of 93.8%, which was 9.3% higher than when the model trained only with the original images. Moreover, we showed how to implement an automatic inspection system based on the proposed method in the bearing assembly station of engine production line.

پايگاه داده , يادگيري ماشين , تشخيص عيوب سطح , ياتاقان‌هاي متحرك , موتور , شبكه‌هاي عصبي پيچشي

dataset , Machine Learning , Surface Defect Detection , Conrod Bearing , Engine , Convolutional Neural Networks

Alireza Yazdanijoo

Dr. Abdollah Amirkhani