Komatsu builds highly customised products, typically in small series, and this increases the risk of mistakes: When operators frequently shift between models and variations, they may overlook some particularly of the current specification. Komatsuhas a strong focus on finding and correcting errors, and this makes it a perfect testbed for QA technology.
Optical methods have the advantage of being flexible, non-invasive, and generally offer high precision. Their disadvantages is that they can be too slow for real-time applications, and lack robustness to reflective surface materials and varying lighting conditions. We are therefore interested in the optical system developed by Per Bergström and his team at LTU, as it combines real-time performance with a reasonable degree of accuracy. The real-time performance is possible since a pre-processing of the CAD-model is used to find deviations. Methods from robust statistics are combined with the pre-processed CAD-model to make the system robust. Robustness to lighting conditions is still a problem, but one we hope to make progress on within the scope of the project.
The geometrical measuring techniques is complemented by machine learning (ML). In the long term, this may avoid explicit programming of deviation thresholds, and allow the system to dynamically adjust to the production environment.
It is likely that IBM’s Watson suite for video/image analysis will be useful, and here we can turn Mikael Haglund at IBM and his team for advice. Work consists in isolating concrete QA tasks, gathering requirements around these, prototyping a solution, and verifying it under realistic conditions. The expected outcome is a prototype system for automatic quality assessment in vehicle assembly lines, that uses optical measuring techniques to detect geometric deviations in components, compared to prescriptive CAD models defining the nominal shape of the measured components.