Optimizing machine learning models for uncertain manufacturing environments in sheet metal forming

The integration of sensor technology represents a promising approach for the reliable and cost-effective real-time monitoring of production processes in sheet metal forming. The combination of time series data with Machine Learning algorithms enables the approximation of complex process states. In particular, deep neural networks have demonstrated remarkable performance under laboratory conditions for the monitoring of blanking processes. However, in real-world production environments, the model performance deteriorates due to pervasive environmental factors that induce uncertainty to the underlying data distributions. Therefore, this work presents investigations that show the impact of different sensor modalities (force, acoustic emissions, acceleration) on model accuracy under uncertain manufacturing conditions. Subsequently, the Robust Bayesian Hyperparameter-Optimization, a robustness-focused, novel adaptation of the Bayesian Hyperparameter Optimization is proposed. The evaluation of the signal-specific model structures highlights acceleration signals as the most robust input. Furthermore, the proposed Robust Bayesian Hyperparameter-Optimization approach can increase robustness across all datasets, including a maximum improvement of over 20% classification accuracy, without degrading performance in the source domain.