The production of aerospace components is characterized by extensive inspection processes in order to ensure the high-quality requirements (e.g. EN 9100). However, the component inspection itself does not add any value to the workpiece, instead it causes a high expenditure of time and costs. The VerticalE2E project is therefore researching an adaptive approach to reduce avoidable inspection costs. By integrating a digital twin into the inspection planning process, the virtual measurement result can be used as a filter to identify and thus reduce avoidable inspection efforts. This means that only workpieces that are not clearly categorized as "OK" or "not OK" by the digital twin are inspected.
Within the CAD-CAM chain, inspection planning typically takes place in parallel with or after CAM planning and usually takes place once for a specific product. All quality inspection steps (including inspection characteristics, means and scope) are thus defined. The necessary inspections are not individually adapted for each workpiece. The prerequisite for virtual inspecting is the existence of a digital footprint of the manufactured workpiece. This includes all relevant production data (machine, sensor, simulation and quality data), which is stored synchronized for each workpiece. Due to varying data formats and readout frequencies, this process step poses a particular challenge for the implementation of adaptive inspection planning.
Based on this, the workpiece quality is modeled in the digital workpiece twin within the material removal simulation IFW CutS. By integrating virtual inspection equipment (here: regression models), the quality achieved is determined on the basis of data. The result of the quality modeling is ultimately new knowledge about the existing workpiece quality. Based on the result, a virtual decision is made on the further quality inspection of the workpiece. In the case of a clear decision, no further physical inspection is necessary and the inspection plan is reduced. A physical inspection is only necessary if the quality value determined is close to the tolerance limit. In this case, without the physical inspection, it cannot be assumed with sufficient certainty that the virtual quality inspection is unambiguous.
The latest results from the research project were recently presented on an international stage at the CIRP Conference on Manufacturing Systems in Cape Town. The model quality for predicting the surface roughness Ra and the form deviation ds was examined for the 3-axis pocket milling of aluminum workpieces. It was found that adding simulation data (material removal rate Qw and tool wrap angle ϕ) to the machine and sensor data (feed rate vf and active force Fa) reduces the expected model error by up to 34%. When using the Gaussian process regression, the average model error is between 7.1 % and 11.1 %. When forecasting the shape deviation ds, the addition of simulation data did not improve the accuracy of the forecast. However, a significant model improvement (+40 %) can be achieved by using machine learning methods compared to conventional regression methods. An average percentage error of 8.5 % - 11.5 % was determined in this context.
The results show that data-based approaches can already be used to achieve a high level of accuracy for the low-cost determination of workpiece quality. The integration of data-based decisions into planning processes therefore has great potential to increase efficiency and reduce the necessary empirical knowledge. "The presentation and discussion of the current research results in front of the international audience was a great experience. I was also able to make other valuable contacts during the conference," says project manager Leon Reuter.
For further information, please contact Leon Reuter, Institute of Production Engineering and Machine Tools at Leibniz Universität Hannover, on +49 511 762 18211 or by e-mail (reuter@ifw.uni-hanover.de).
