Development of a Cloud- and Edge-Architecture for adaptive model weight optimization of a CNN exemplified by optical detection of hairpin welding

The beginning of a global reorientation towards an increasingly conscientious approach to nature and the human habitat has been accompanied by changes in industry and society. The automotive industry, where a transition from combustion to electrically powered vehicles is currently underway, is also concerned with this change. In addition to increasing the capacity of the battery, improving the efficiency of the electric motor is essential. To achieve these goals, however, new technologies such as hairpins for the stator are needed. An important process step involves the welding of two pairs of hairpins, which often leads to welding defects. Nevertheless, expert knowledge in this field is limited. Optical monitoring of the welding process with the help of a convolutional neural network (CNN) is a good approach. This approach can compensate for the low level of expert knowledge and detects and classifies welding defects directly in the production line. However, the disadvantage of optical monitoring is that production conditions and the surrounding environment change over time. This has an impact on optical detection and can negatively affect the accuracy of a CNN. For example, the camera perspective can change, which has a negative effect on optical quality monitoring. Therefore, this paper presents an approach for monitoring and evaluating the quality of a CNN in a cloud instance online. If a deteriorating quality is detected, the CNN in the cloud is re-trained by continuously collected data and then automatically deployed to the production line. This allows the CNN to adapt to the changing environmental conditions. The present approach is demonstrated and validated with real data of the stator production process. Compared with the current state-of-the-art, this control loop is highly automated and requires a minimum of human intervention.