Deep neural network-based cost function for metal cutting data assimilation

Metal cutting is a complex process in machining that is typically not easily modeled via the finite element method (FEM) due to insufficient models or uncertain conditions. To improve FEM analysis of such a complex process, it is desirable to estimate the model parameters and uncertain conditions based on measurements of actual machining processes, which could be realized by data assimilation. However, the application of data assimilation to complex processes such as metal cutting is not straightforward because a comparison of the actual process and the corresponding FEM images is not trivial. To overcome this issue, we consider an extension of the cost function based on the object detection and classification capabilities of deep learning by evaluating the similarity between the FEM results and the images acquired during the actual machining process. The overall procedure is demonstrated by investigating a cutting chip in a turning process, whose shape depends on the workpiece material, cutting conditions, and the cutting tool. We first trained a deep neural network using chip images acquired from a turning experiment. This resulted in 85.5% detection and classification accuracy for testing data obtained from the same experiment. The trained network is then used to detect the images generated based on FEM. It was confirmed that the confidence score calculated using the trained deep neural network can be used to quantify the difference of the cutting chip shape generated by FEM. This preliminary study revealed that a realistic cutting chip shape can be generated by estimating the work-hardening exponent and the static friction coefficient in FEM based on images obtained during the turning process. This confirms that a deep learning-based cost function can be used to achieve image-based data assimilation.