Machine learning-based model for detecting uneven wear and temperature deviation events in hot forging process

The aim of this study was to perform online modification of unobservable but crucial processing parameters for a hot forging tool in real time to maximize workpiece quality and to extend the useful life of the hot forging tool. The hot forging process involves with plastic deformation of workpiece and the inevitable tradeoff is rapid wear of the forging tool surface. However, some uncontrollable factors, e.g., placement of the workpiece in the cavity and varying temperature of the inner forging tool, may substantially affect the nonlinear degeneration of the forging tool. Inadequate workpiece quality or failure of the forging tool may then require unexpected changes in the production schedule. To address this issue, this study fitted accelerometers to a forging machine to acquire the full range of motion required for the machine to complete a forging process. The wear and thermal experiments were also conducted simultaneously to label events on the vibratory signals. A trend analysis was performed to analyze vibratory signals in both the time domain and the frequency domain. Since the proposed feature extraction method could only reveal trends in the specific frequency bands, however, extracted features and labeled events were used as input parameters and output targets, respectively, in a supervised learning model. Then, a systematic design approach for using the model architecture is proposed, including a strategy for model training between model underfitting and overfitting conditions. The methodologies proposed in this study are then evaluated in two case studies.