Using Machine Learning to Improve the Modelling of Machining and Cutting Processes

Using Machine Learning to Improve the Modelling of Machining and Cutting Processes

Leitung:  C. Weißenfels, P. Wriggers
Team:  M.Sc. Dengpeng Huang
Jahr:  2018
Förderung:  China Scholarship Council (CSC)

Metal cutting is a fundamental process in industrial production. The fast and accurate on-line prediction of metal cutting processes is crucial for the Intelligent Manufacturing (IM). With the advent of high-speed computing, robust numerical algorithms and machine learning technology, computational modelling serves as a tool for not only accurate but also fast predicting the complex machining processes and understanding the complex physics.

Physical Mechanisms: The coupled physical phenomena in metal cutting can be summarized as large plastic deformations, adiabatic shear band formation, chip separation as well as frictional contact.

Material Model: To capture the shear band formation, the Johnson-Cook flow stress model within the hyperelasto-plasticity framework is used to account for the strain hardening, strain rate hardening and thermal softening. The chip separation is described by the Johnson-Cook fracture model with a supplementary condition for the stress triaxiality. This condition allows a more accurate measurement of the chip size and chip spacing.

Solution Method: To deal with the large topology change during metal cutting, a recently developed Galerkin type meshfree approximation scheme, the Optimal Transportation Meshfree (OTM) method is applied. This method has advantages in algorithmic robustness and convenience in fracture modelling.

Machine Learning: Machine learning technology can play great roles in computational modelling of metal cutting process, such as speed up the simulation and data driven computation. The Artificial Neural Network (ANN) has been used as a tool for numerical modelling of non-linear material behaviours. The ANN based material model can be trained off-line with the experimental data and thus has the possibility of improving the model by re-training when the additional experimental data sets become available.