Project: Intelligent manufacturing

Problem definition

Manufacturing tolerances pose sensitive cost consequences in the field of materials handling. Especially for system elements of assemblies consisting of several individual parts, tolerance chains arise that require an enormously complex mathematical interpretation. Pump systems, for example, consist of various elements of housing parts, rotor elements, bearing shells and bearing seats as well as drive shafts and others.
The benchmark for product quality is a defined required minimum flow rate, which is massively impaired by inaccurate fits, relative play or even grinding. Knowledge of these properties is only of limited help, since they result from a large number of individual dimensions and combinations of components.
In order to reduce the scatter of the actually achieved flow rate, it is usually only possible to manufacture more accurately across components. In this way, the individual part deviations are reduced and the error influence on the assembly is minimized. However, this approach is accompanied by exponentially increasing manufacturing costs and thus offers no added value in terms of the economic efficiency of the end product.

Objective

The possibility of examining the assemblies for their actual performance in practical tests is to be used in order to make better use of the individual parts produced and measured in series.
Here, the goal should not be a reduction of manufacturing deviations as in the classical sense. Measurement data should be used in the step of combining the individual parts to form the end product, so that an optimum achievable flow rate can be realized by pairing the individual parts accordingly.
Manufacturing deviations can be determined concretely by measuring the actual dimensions of the parts. Due to the complex component geometries, several hundred individual dimensions result. By using machine learning, relevant information is to be generated here. The information on the actual relevant dimensions on the achievable output can have a significant influence on the manufacturing costs.
If a relation of the measured values to the final product properties can be determined, a prognosis of the production quantity can be generated by combining different deviating individual parts.

Result

The correlations of the individual part dimensions to the target variables of the pump delivery rate could be comprehensibly demonstrated. On this basis, the influence weights of the individual dimensions were taken from the machine learning algorithms. This data is used directly to realize a manufacturing optimization in the short term. In this way, decisive tolerances in manufacturing are reduced in accordance with the algorithms. The theoretical additional costs of the smaller tolerances can be compensated for in other component areas of lesser influence, since these have no relevant effect on the target quantity.
In addition, test rig trials are carried out on an ongoing basis to generate new information on the actual flow rate. This data basis is constantly fed into the dynamic training process.
The aim is to increase the prediction accuracy of the algorithms with regard to the achievable flow rate. An advantage is the possibility to use measured single parts directly to make a prediction and to verify it afterwards on the test bench.
Individual parts with high production deviations no longer have to be removed from the production line as a preventive measure. They can be combined by AI algorithms with suitable pairings of the system components to ensure the required quality of the assembly.