No abstract

The aim of this work is to present a new methodology for the automated analysis of the cross-sections of experimental chip shapes. It enables, based on image processing methods, the determination of average chip thicknesses, chip curling radii and for segmented chips the extraction of chip segmentation lengths, as well as minimum and maximum chip thicknesses. To automatically decide whether a chip at hand should be evaluated using the proposed methods for continuous or segmented chips, a convolutional neural network is proposed, which is trained using supervised learning with available images from embedded chip cross-sections. Data from manual measurements are used for comparison and validation purposes.

Thermal errors are among the most significant contributors to deviations of products manufactured on modern machine tools (MTs). Reducing them is typically achieved through either design adaptation, active cooling of the MT and its environment, or compensation using measurements or model-based predictions. Model-based compensation strategies promise to have the lowest environmental footprint by far. In general, a compensation model needs to be accurate, robust to changing boundary conditions and must require only minimal experimental efforts as this reduces the productivity of the MT. Model inputs such as temperature measurements or the power consumption of various components, can be used to predict the thermal errors. The temperature inputs require additional sensors, effort and cost for the MT manufacturer to install and ensure up-time while the power consumption could be logged and are typically provided from the control system anyway. Adaptive compensation models are created using four different sets of inputs consisting of 13 temperature sensors and 7 power measurements. While the best results were obtained with all 20 inputs, the 7 energy recordings give similar results as the 13 temperature sensors if the environmental temperature is considered. The volumetric RMSE was reduced by 72% and the maximal error from 32.75 µm to 9.5 µm. ARX models proved to be suitable and even outperform more complex model structures such as LSTM and especially those without time dependency such as feed forward neural networks.