인공지능 기반 GMA 용접 시 아크 사운드 분류에 따른 실시간 용접 품질평가 기술에 관한 연구

Abstract

A Study on the Real-time Welding Quality Evaluation Technology according to Arc Sound Classification on GMA Welding Based on Artificial Intelligence Kim MuSung Advisor : Prof. Bang Hee Sun, Ph.D. Department of Welding and Joining Science Engineering, Graduate School of Chosun University Deep learning is used for problem solving and data analysis in a variety of fields, and when deep learning and programming skills are acquired and understood, new opportunities can be found and innovative solutions can be created. At the same time, GMAW welding is the most basic and essential assembly process in production, high welding speed and productivity, and is widely used in automatic and semi-automatic welding systems. It is done using an arc between the metal electrode and the workpiece, using gas to replace the air next to the arc. Although GMAW welding quality assessment methods rely on welding experts such as ultrasonic signals, image processing, and optical signals, the characteristics of GMAW's arc sound play an important role in assessing welding quality and stability during welding. Using JVM (Java Virtual Machine)-based programming language (Kotlin) developed by JetBrains, we created the ability to measure and record arc sound during GMAW welding in an Android studio program and developed a mobile phone application that stores arc sound measurements in the form of time-series data on the cell phone's internal memory. Using the developed application, we constructed a real GMAW welding environment and recorded 150 arc sounds while performing GMAW welding under 50 current-voltage conditions on a 20 mm thick carbon steel. To determine the arc stability of the time series data of the stored arc sound, the current voltage waveform was stored through the Labview program during GMAW welding. In the current voltage waveform, the mean and standard deviation of the current value (Is), maximum current value (Ip), arc time (Ta), and short circuit time (Ts) at the start of the short circuit were analyzed and compared with the bead image to label the arc stability in the arc sound. As for the arc stability labeling method, a table was prepared using a binary classification method divided into 0 (Good) and 1 (Bad) and a multiple classification method divided into 2 (Under), 3 (Normal), and 4 (Over). After that, data was loaded and constructed in the stages of data pre-processing, data segmentation, neural network model generation, and model compilation to create a deep learning model. As a result, the accuracy of the training data of the binary classification model was 97.48%, the performance of the validation data was 86.67%, the accuracy of the training data of the multi-classification model was 95.80%, and the performance of the validation data was 80.00%. Using the developed application, we wanted to lead the field of welding innovation that can easily check the quality of the welding process and respond quickly to problems by analyzing stable arc sound in real time, and provide a tool that can easily evaluate arc stability during welding.