Understanding the Microwire Casting Process Through Convolutional Neural Networks

Abstract

Even though the Taylor-Ulitovsky process for producing microwire has existed and been widely used in the past century, there are various challenges facing the microwire manufacturing process, such as inconsistent wire diameter, constant breaks of microwire during fabrication, and the difficulty of producing wires with a smaller diameter. These challenges can make the microwire fabrication process inefficient, and this research aims to understand how thermal images from the fabrication process under various parameter settings can be used to assess and classify the quality of the microwire. Thermal videos and other process variables were collected from a microwire manufacturing lab, and the thermal image datasets from the video were trained using a pretrained Convolutional Neural Networks (CNN) in order to better understand how changing certain parameters for the microwire manufacturing process can affect the microwire quality. The features extracted from the thermal images using the proposed CNNbased machine learning algorithm is capable of classifying the microwire fabrication process into four stages, i.e., initialization stage, stable stage 1, stable stage 2, and ending stage. The stage classification accuracy reveals high repeatability and performance from the proposed CNN model. The results are promising since manufacturing process parameter settings can be adjusted and optimized by referring to thermal image characteristics, and therefore the CNN model can improve microwire quality and predict failure of the microwire.