»A stream is made up of many small rivulets - it is similar in companies, where a stream of costs is formed from numerous individual amounts flowing in unnoticed. As part of the straightening process, we were able to identify a potentially unnecessary source of costs. By collaborating with Fraunhofer on the QuickCheck project, we received confirmation that there is considerable potential for cost optimization with the help of machine learning (ML).«
Thomas H. Schmid (Operational Controlling)
Contact at the AI Innovation Center
Xinyang Wu
ML-supported prediction of screw straightening quality
Initial situation
In the production of screws, slight initial bending regularly occurs, which is corrected manually before further processing in a hydraulic pressing machine. The screws are supported on two prisms with a fixed distance between them, while the press acts from above and changes the deflection - particularly at the central measuring point MP0. This manual straightening process is time-consuming and labor-intensive, requires specific experience and leads to quality-dependent fluctuations. In addition, individual screws remain outside the permissible tolerances despite pressing, which causes rejects and additional process costs.
Around 1000 data sets with four deflection measuring points (MP1, MP2, MP0, MP3) and relevant process parameters such as press position, press force, press stroke and prism distance are available for the analysis. The aim is to develop an ML model that reliably predicts the change in deflection at MP0 based on the initial deflection and the set process parameters. The aim is to automate the straightening process, stabilize the quality of results and reduce the use of resources in the production process.
Solution idea
Two ML methods are used for model-based prediction of the change in deflection at MP0: LightGBM and neural networks. Both methods are suitable for precisely mapping the non-linear dependencies between initial deflection and process-relevant parameters. In addition, feature engineering is used to derive technical characteristics such as local gradients, curvatures and process-related interaction variables. These features reflect central physical effects of the straightening process and improve the model quality. As a result, a robust prediction model can be developed that supports automated and quality-assured control of the straightening process.
Benefit
The experiments show the practical added value that the investigated models offer for process evaluation. Despite moderate R² values (MP0 approx. 53 percent), the regressors capture the deflection trend and thus enable a basic assessment of the behavior after pressing, even if MP1 only has a small explanatory share. However, the benefit of the classification models is particularly high: with over 86 percent accuracy, F1 scores >0.92 and a recall of 95.7 percent, they provide reliable information on whether MP0 is within tolerance. LightGBM offers additional added value through better probability calibration (ROC AUC). Overall, it can be seen that the classification is significantly more stable and reliable in this application and therefore provides a robust basis for quality assurance decisions.
Implementation of the AI application
The implementation of the AI application shows that the regression task can only be solved to a limited extent due to the limited data basis of around 1000 samples. The models cannot reliably map the physical laws of the screw position change in a purely data-driven manner. Physics-informed approaches are much better suited for such data-poor scenarios, as they directly integrate physical relationships and thus enable stable and consistent predictions even with little data. The classification task, on the other hand, can be successfully implemented with both ML models, as the binary decision »within vs. outside the tolerance« requires significantly less data. Based on this, pressing parameters - especially the pressing position - can be specifically optimized in the future in order to automate the process and reliably ensure quality.