Simulation-Based Data Augmentation for an Inline Wear State Detection During Blanking
摘要
Manufacturing systems are often characterized by a complex dynamic behavior. In this context, machine learning (ML) offers great potential for a better understanding of such systems to reduce operating costs while at the same time ensure a reliable process control. However, the training procedure including the training data generation to set up the ML model is time-consuming and cost intensive. Especially, the consideration of faulty process states to avoid unbalanced data sets is associated with a high technical effort due to their low likelihood of occurrence in a running production. Therefore, this study proposes a simulation-based data augmentation method which integrates synthetic data into the training procedure of ML models to estimate wear in high-speed forming processes inline. The synthetic data is generated by a rudimentary simulation of the punch phase during a blanking process, which allows a parameterized variation of the abrasive wear states. To quantify the potential of the approach, the performance of the ML model based on synthetic data is compared with a reference ML model trained with experimentally acquired force signals. The proposed method significantly reduces the effort required to create a training data set, and only requires an initial set of time series representing a single wear state to validate the simulation model. At the same time, the augmented ML model estimates the current wear state during the blanking process with a mean absolute deviation of 15 µm from the target wear state described by the cutting edge radii of the tool, considering sensorial acquired force signals as a test input.