Image-based Supervision System Design And Implementation Of Metal Atomization Process

In the 21st century, the rapid development of artificial intelligence and the internet of things technology constantly promote the innovation and progress of manufacture industry, Germany proposed "industry 4.0" project aims to promote the intelligent level of manufacturing. Meanwhile, "China manufacturing 2025" claims clearly for the development of intelligent industry in our country. As the representative of the intelligent algorithm, machine vision and machine learning have made continuous development for replacing manual operation as much as possible since their appearance, and they are more and more widely applied in the field of industrial. Metallurgical industry, with its poor production conditions, more needs machine vision to play an important role.Gas atomization always occupy the dominant position in the market because of its good sphericity, high uniformity and low cost in the production of metal powder. But the particle size of metal powder always fluctuated easily because of various factors in production process. Many researchers have made lots of experiments and tried to find the mechanism about these factors, the most famous one is proposed by Lubanska.H. The Lubanska formula propose that many factors include the gas velocity and the mass flux of gas and liquid metal mainly influence the particle size of powder, and multiphase coupling effect in the process of atomization also greatly increase the complexity of its mechanism model. It is not realistic to predict the particle size by the mechanism model with these factors which are hard to measure.As the most important part of gas atomization, the atomization process always contain some state parameters which can be the representation of current situation. This paper estimates the metal powder particle size from a new perspective. First of all, it proposes a GTjmedfilt algorithm based on median filter for image noise-reduction and restoration, this method eliminates noise distinctly and preserves the detail information as much as possible. By using value-based Otsu threshold segmentation in HSV color space, it separates target from background and extracts some significant geometrical information of the target, and uses colorimetric thermometry for the temperature field reconstruction. Finally, we select a number of appropriate production data according to the simplified Lubanska formula for model training and establish the LS-SVM-based prediction model for estimating the particle size and span of metal powder by analyzing atomization process image.This method makes effective use of gas atomization mechanism, while avoiding the complexity of it. By comparing with the actual value, it is confirmed that the prediction model could estimate the D50 particle size and powder span accurately under acceptable error range. Using machine vision and machine learning replace manual operation also enhances scientificity and stability of the system and provides quantitative data reference for practical production. At the end of the article gives the system design process of framework and function design process and the application in production.