Generative Adversarial Network for Prediction of Workpiece Surface Topography in Machining Stage

Surface topography plays a key role in the service performance of parts, and is often used as a metric for detecting machining states as well. Hence, it is of great importance to online predict the surface topography to avoid surface deterioration and machining faults. However, alternative solutions for this purpose remain to be developed. This article presents a generative adversarial network to predict the statistical distribution of surface topography conditioned on cutting parameters, measured cutting forces, as well as system vibrations. The local dynamic characteristics and global quasi-static trend of measured process quantities are extracted by short-time spectrum technique, then mapped into the local surface features patch by patch. To improve the network performance, we embed several useful tricks, such as recursive residual block, skip connection, and global residual architecture into the surface generator so as to enhance its capacity of learning the complicated nonlinear mapping against limited training data. The model performance is evaluated and validated using the structure similarity index, the distance of extracted features by the visual geometry group network, as well as the perceptual index. The results indicate that the recursive architecture performs better than nonrecursive one on learning the nonlinear correlation between surface topography and measured process quantities. Besides, by applying adversarial learning, the texture characteristics can be well-captured to avoid the blur of high-frequency details.

Automated summary

This article introduces a generative adversarial network to predict the statistical distribution of surface topography based on cutting parameters, measured cutting forces, and system vibrations. By utilizing recursive architecture to learn the nonlinear correlation between surface topography and measured process quantities, and embedding useful tricks to enhance network performance, the model is able to capture texture characteristics effectively and improve learning. The results show that the recursive architecture performs better in learning the nonlinear correlation and avoiding the blur of high-frequency details through adversarial learning.