»SimVision« – Simulation-Based Visual Inspection for the Industry

Development of coated cutting tools is a highly specialized and expensive process. Coating enhances the tool’s cutting performance and extends its lifetime. When the coating is not applied correctly, the tool will malfunction and potentially cause further damage to the object it was cutting. However, the coating defects are measured in micrometers and are close to impossible to collect in a catalogue. How can automate the inspection process? This is where the rule-based synthetic data comes into play. In the »SimVision« project, researchers from the Fraunhofer Institutes ITWM, IGD and IST work together with TU Wien to develop scalable pipeline for training Artificial Intelligence with synthetic data.

»SimVision« stands for »Simulation-Based Visual Inspection«. A central component is the creation of a synthetic data set for Machine Learning with synthetic image data that is realistic enough to prepare AI models for practical use. But how realistic does this data need to be? How does the data influence inspection accuracy? And how can we model defects which we have never seen? These questions are at the heart of the research.

Application Example: Increasing Efficiency Through Automated Inspection in Production

The components inspected in the project include cutting tools with diamond coatings of different granularity, causing varying appearance. Conventional AI-based methods reach their limits when it comes to reflective, textured or geometrically complex surfaces and little training data is available. Such surfaces make it difficult to reliably detect defects such as delamination, scratches, cracks or wear and tear.

In »SimVision«, we are developing a flexible inspection system that automatically inspects surface quality at different stages of coating production. The inspection system is made possible by our rule-based synthetic data. It means that the synthetic data is based on mathematical models of surface texture, defects and tool geometry. By controlling all the parameters, we can generate photorealistic training data with a wide variety of possible surface conditions, including edge cases. This improves the accuracy, flexibility and cost-efficiency of the inspection system.

Overarching Goals of the Project: Flexible Quality Inspection Through AI

1. Optimization of Image Synthesis Processes: Physics-based modeling of surfaces, defects and light reflections creates realistic, synthetic training data for AI models.
2. Establishing Simulation-Based Visual Inspection as a New Research Discipline: With a focus on industrial knowledge transfer, we develop practical solutions that can be quickly transferred to real production processes.