Machine learning classification of surface fracture in ultra-precision diamond turning using CSI intensity map images
Monica L. Nogueira, Noel P. Greis, Rachit D. Shah, Matthew A. Davies, Nicholas E. Sizemore
Abstract
With increasing focus on automation of machining processes in a smart factory, attention has turned to implementation of machine learning (ML) not only for tasks such as process parameter optimization, tool wear monitoring, and surface roughness prediction, but also for near-real-time determination of part quality using image data for inline inspection. As a first step, this work compares the performance of ML approaches using post-process image data, as opposed to numerical machining process data, for the discrete prediction of surface fracture in single-point diamond turned germanium for IR optics. Typical classifiers such as Random Forest and Support Vector Machine are employed for processing numerical machining parameters, while feedforward neural networks (FNNs), convolutional neural networks (CNNs), and deep CNNs (DCNNs) are utilized for processing intensity map images of the turned surfaces. The feasibility of using FNNs, CNNs, and DCNNs for identifying fracture in intensity maps is initially explored on raw images, and then by employing edge segmentation and edge detection techniques such as the Sobel, Prewitt, and Laplacian operators, as well as denoising methods as the Adaptive Gaussian Threshold filter. Finally, hybrid CNN and hybrid FNN models that combine numerical process data and image data are explored. Keywords