Clustering of Learning Sub-models for Quality Prediction in a Resource-Efficient Tool Grinding Process

The prediction of workpiece quality in process planning, using machine learning models, is a common-researched topic. Until now, trained models were static and could not update themselves with new data. However, this aspect is crucial when considering the continuously changing manufacturing circumstances in regards to new process parameters, materials, and workpiece geometries. In addition, repeatedly training process models with an extended mixed dataset decreases the prediction quality due to the increased data divergence. This paper presents an approach to automatically generate sub-models, which maintain the prediction quality even if novel data is considered. The challenge is to define the amount and content of these sub-models through clustering. Tool grinding experiments will be conducted with different process parameters, materials, and workpiece geometries in order to obtain a divergent dataset. Subsequently, cluster approaches are compared to obtain dynamic growing models, which enable optimized planning for a more resource efficient process. Finally, the method will be generalized in order to ensure a process-independent usage.