A cognitive approach for a robotic welding system that can learn how to weld from acoustic data

Abstract

Laser beam welding is the method of choice for the high-quality joining of materials. However, for industrial production these systems have to be set up and calibrated manually with much effort. Our objective is to apply intelligent data processing that results in a cognitive technical system that can learn how to weld, speed up the configuring process, and reduce costs. While monitoring laser welding with cameras and optical sensors has already been demonstrated elsewhere, this paper emphasizes the benefits of monitoring with acoustic sensors and feature extraction. Using acoustic sensors, the cognitive system is more sensitive to strong optical radiation. Several combined methods such as wavelet analysis, fast Fourier transformation, and linear dimensionality reduction are evaluated with sensor data from real experiments. Finally, as machine learning, the results are classified with learned reference data to obtain reliable information for monitoring and possibly using closed-loop control.