An Inspection and Classification System for Automotive Component Remanufacturing Industry Based on Ensemble Learning

This paper presents an automated inspection and classification system for automotive component remanufacturing industry, based on ensemble learning. The system is based on different stages allowing to classify the components as good, rectifiable or rejection according to the manufacturer criteria. A study of two deep learning-based models’ performance when used individually and when using an ensemble of them is carried out, obtaining an improvement of 7% in accuracy in the ensemble. The results of the test set demonstrate the successful performance of the system in terms of component classification.

Viewpoint placement for inspection planning

Inspection planning approaches so far have focused on automatically obtaining an optimal set of viewpoints required to cover a given object. While research has provided interesting results, the automatic inspection planning has still not been made a part of the everyday inspection system development process. This is mostly because the plans are difficult to verify and it is impossible to compare them to laboratory-developed plans. In this work, we give an overview of available generate-and-test approaches, evaluate their results for various objects and finally compare them to plans created by inspection system development experts. The comparison emphasizes both benefits and downsides of automated approaches and highlights problems which need to be tackled in the future in order to make the automated inspection planning more applicable.

IMPROVED TOW-END DETECTION FOR FIBER PLACEMENT INSPECTION USING MACHINE LEARNING

A machine learning approach to improve the detection of tow ends for automated inspection of fiber-placed composites is presented. Automated inspection systems for automated fiber placement processes have been introduced to reduce the time it takes to inspect plies after they are laid down. The existing system uses image data from ply boundaries and a contrast-based algorithm to locate the tow ends in these images. This system fails to recognize approximately 10% of the tow ends, which are then presented to the operator for manual review, taking up precious time in the production process. An improved tow end detection algorithm based on machine learning is developed through a research project with the Boeing Advanced Research Center (BARC) at the University of Washington. This presentation shows the preprocessing, neural network and post‐processing steps implemented in the algorithm, and the results achieved with the machine learning algorithm. The machine learning algorithm resulted in a 90% reduction in the number of undetected tows compared to the existing system.