Fault Classification and Correction based on Convolutional Neural Networks exemplified by laser welding of hairpin windings

Abstract

The automotive industry is facing a change from combustion engine-powered to electrified vehicles. Besides the traction battery, the electric engine is one of the most important components of the electrified powertrain. In order to increase the energy efficiency of the electric motor, wound copper wires are replaced by enameled rectangular copper wires, known as hairpins. In order to produce a conductive connection between hairpins, it is necessary to weld them together. Currently, the automated laser welding of copper is a poorly understood process. Such new production processes are still unknown in comparison to classic engine production and there is only little expert knowledge available. The integration of Industry 4.0 techniques and advanced data analytics provides the opportunity to understand the process of copper welding more thoroughly. A common understanding of advanced data analytics differentiates between predictive and prescriptive analytics. One of the most promising developments in advanced analytics is Machine Learning (ML). There is a wide range of different types of algorithms, theories and methods. An example of these are Convolutional Neural Networks (CNN). They have been designed for learning multidimensional data, such as images or even videos. This paper presents such a CNN to detect welding defects of hairpins. Depending on the classified defect, a rework concept is given (prescriptive analytics). The input parameters are the visual information are derived from of a 3D camera. Using the welding process as an example, the paper illustrates a newly developed method based on the CRoss Industry Standard Process for Data Mining (CRISP-DM) for the development of the CNN. In this context, the paper deals in detail with data preprocessing, modeling and evaluation. The newly developed methodology and architecture of the CNN achieves an accuracy of over 99 percent to predict the defect class.