The quality inspection of finished workpieces generally represents a non-value-added process. In order to increase profitability, a reduced inspection effort is therefore of particularly high interest for manufacturing companies. Frequently, random sample-based inspection plans are used to reduce the effort in the final quality control. Although this reduces the general inspection effort, there are still avoidable inspection efforts and at the same time a slip-through of quality-critical components. The work package "Self-adapting quality inspection" of the VerticalE2E project therefore deals with the integration of data-based quality inspections into inspection planning. With the resulting quality knowledge, a targeted dynamization of the inspection scope of quality-critical components is possible. For quality prediction, machine data (feed, speed, axis positions), simulation data (cutting depth, width, tool wrap angle, metal removal rate) and sensor data (cutting force) were synchronized with the tactile surface roughness Ra (average 0.449 µm). The predictive ability of different machine learning methods was then investigated: support vector regression, gradient boosted trees and Gaussian process regression. For all algorithms, nested cross-validation was used as an evaluation method of prediction quality. For the formation of the model evaluation parameters, the cross-validation was performed 100 times in each case and then the mean and standard deviation were formed from the runs. Thus, the risk of random-based over- or underestimation and overfitting is minimized. A correlation analysis provided additional insight into which of the available data significantly affect the predictive performance of the regression models. The presented evaluation thus determines which variables are necessary for an efficient data tap. In the present case, these are the feed rate, the metal removal rate and the resulting active force. The depth of cut showed the least influence and is therefore not considered in detail for the present application. For the prediction of the surface roughness, the results of the Gaussian process regression were the best. Here, an average coefficient of determination R² of 0.83 was achieved with a scatter of about 0.08. The mean absolute error was 0.038 μm. The second best model quality was achieved with the support vector regression. Here, an R² of approximately 0.82 with a scatter of 0.099 is present. The mean absolute error is 0.037 μm. The Gradient Boosted Trees showed the lowest reliability. They achieved an R² of 0.696 with a high dispersion of 0.234 and a mean absolute error of 0.043 μm.
Based on the evaluations, it can be stated that data-based prediction models are basically suitable for determining the surface roughness. It has already been shown that methods such as support vector regression or Gaussian process regression are to be preferred. In the further course of the project, the data-based quality predictions will now be integrated into the test planning process. In this way, time-consuming and costly quality inspections can be planned in a targeted manner in the future as part of adaptive inspection planning, reducing the inspection of components that are safely within the required tolerances.
Contact:
For further information, please contact Leon Reuter, Institute of Production Engineering and Machine Tools at Leibniz Universität Hannover, by phone +49 (0) 511 - 18211 or by e-mail (reuter@ifw.uni-hannover.de).
