Arc Signature Classification using Machine Learning Approach to Identify Weld Defect Conditions in a Robotic Pulsed Gmaw Process

Abstract

Welding is one of the major metals joining processes employed in fabrication newlineindustries especially in the manufacturing of boilers and pressure vessels. Quality of newlineweld is an important aspect for such industries, considering the severe operating newlineconditions. Industries are looking forward for some kind of a real-time process newlinemonitoring/control using sensors and signal processing systems that will ensure the newlinerequired weld quality. In this work, an attempt is made to establish a correlation newlinebetween the current and voltage signatures with undisturbed arc condition producing newlinegood weld and disturbed arc condition producing porous and burn-through welds in a newlinerobotic pulsed Gas Metal Arc Welding (GMAW) of carbon steel plates used in boiler newlineapplication. Experiential studies are carried out to establish undisturbed and disturbed arc newlineconditions producing good weld and defective weld conditions respectively. newlineUndisturbed arc is established by following good weld practices established as per newlineWelding Procedure Specification (WPS). Disturbed arc conditions are intentionally newlineestablished by creating porosity and burn-through weld defects using experimental newlinedesigns. An experimental set-up is established to acquire current and voltage newlinesignatures of the undisturbed arc producing good weld and disturbed arc conditions newlineproducing burn-through and porosity welds. The behaviour of undisturbed and newlinedisturbed arc conditions are studied by analyzing the pulse parameters of current and newlinevoltage. Weld arc signatures and their correlation with good weld and defective weld newlineconditions are studied using Probability Density Distribution (PDD) plots of current newlineand voltage. Machine learning algorithms are used to build statistical models to newlineclassify the good weld and defective weld conditions. Statistical features extracted newlinefrom time, frequency and wavelet domains of current and voltage transients are used newlineto train and test the statistical models proposed in this study. Machine learning newlineclassifiers viz. decision trees, support vector machines ...