Adaptive Visual Quality Inspection Based on Defect Prediction From Production Parameters

Abstract

At the end of a production process, the manufactured products must usually be visually inspected to ensure their quality. Often, it is necessary to inspect the final product from several viewpoints. However, the inspection of all possible aspects might take too long and thus create a bottleneck in the production process. In this paper we propose and evaluate a methodology for adaptive, robot-aided visual quality inspection. With the proposed method, the most probable defects are first predicted based on the production process parameters. A suitable classifier for defect prediction is learnt in an unsupervised manner from a database that includes the produced parts and the associated parameters. A robot then steers the camera only towards viewpoints associated with predicted defects, which implies that the trajectories of robot motion for the inspection might be different for every product. To enable dynamic planning of camera trajectories, we describe a methodology for evaluation and selection of the most appropriate autonomous motion planner. The proposed defect prediction approach was compared to other methods and evaluated on the products from a real-world production line for injection moulding, which was implemented for a producer of parts in the automotive industry.

Keywords

• Industrial informatics
• injection moulding
• machine learning
• production parameters
• quality inspection
• robot motion planning