Monitoring and prediction of the process energy in multi-stage cold forging using recurrent and self-attention based neural networks

Abstract The effects of greenhouse gases on the global warming and the natural limitations of manufacturing resources oblige the cold forging manufacturers to pay more attention to their energy consumption. Although the material production entails the higher energy demand for the part manufacturing, understanding the energy requirement of single processes in a press shop still remains crucial for fully determining the life cycle assessment of cold forging parts. The application of classical energy measurement tools to monitor cold forging processes is rather challenging due to the complexity of the manufacturing environment and the resulting expense for the measuring equipment. For this reason, this study explored the processing of the forming forces for the real time monitoring and prediction of the required forming energy. For this purpose, a two-stage cold forging process was considered and two tasks were designed. The first task was the prediction of the cumulated energy during the forming stroke based on the recent in-stroke energy history and the second task focused on understanding the effects of one forming stroke on the energy evolution of the future ones. For both tasks, the performance of recurrent and self-attention based neural networks was compared and the results showed high predictions accuracy for the different model architectures. Being able to predict the forming energy would allow to identify the process variations due to the fluctuation of the part geometry, material properties and surface quality, the tool wear state and the machine downtimes.