Learning from machines – AI in cutting technology

 

Can manufacturing be learned? Can machines become better the more they produce? How do you process high-frequency signals from the production process? Within the framework of various initiatives, research groups and the cluster of excellence "Internet of Production", the research group is Product and Process Monitoring tackles these topics. We analyze data from the ongoing production of our partners in industry. For example, how can tool breakages in drilling be monitored automatically even in older machines? Current monitoring solutions expect the operator to manually set threshold values for each operation. Methods from the field of machine learning allow to determine at once the type of operation (supervised learning) and to detect abnormalities in the signal (unsupervised learning). This enables for monitoring even in a very heterogeneous manufacturing field (in our case 114 different operations).

We are currently working on wear monitoring during milling. Both directly, with the help of an integrated camera and convolutional neural networks (cNNs); and indirectly, with the help of process signals. Here, machine learning can process several sensor signals simultaneously (sensor fusion). A special challenge in milling is the rapidly changing engagement conditions, which can be considered in machine learning by means of an engagement simulation in order to train a model that is as generic as possible.

The direct, optical wear measurement relies on instantaneous segmentation through cNNs (deep learning). Thus, a wear classification is carried out pixel by pixel, which makes it possible to determine more than just VBmax: new, more representative form values of the wear surface can be examined.