Vision system for quality inspection of automotive parts based on non-defective samples

Nowadays, companies in the automotive industry focus on delivering high-quality products to their customers, however, this task tends to be more complex as new car models emerge because new quality requirements must be learned. Currently in some companies, vision systems are used for the part quality inspection process, however, their learning process requires many correct and defective data to generate better predictions. Although it is possible to learn from correct samples, it is difficult to learn from defective parts because they are difficult to find in a company with strict quality standards. In this work, the implementation of machine learning classifier algorithms is proposed to detect correct and defective samples of different part types from the learning of only samples that meet quality standards. The feature extraction from images corresponding to suspension control arms and engine front covers was carried out, then a data augmentation process was applied to be analyzed by classifying algorithms in two stages: Part Identification and Geometric Quality Inspection. As a result, it was obtained that the Support Vector Machine classifier was the best algorithm in both stages, resulting in 100.0% accuracy in identifying the parts, 96.0% accuracy in detecting defective suspension control arms and 100.0% accuracy in finding defective front cover arms.