Real-Time Welding Tip Condition Classification in Arc Welding Using Acoustic and Electrical Signals with Machine and Deep Learning

Abstract

During the process of arc welding, the welding tip wears out due to the high temperature brought by high current. The degradation and contamination of welding tips will cause lower welding quality, higher energy consumption, and uncertain downtime. Thus, reliable real-time monitoring is important for productivity and quality in arc welding. This thesis proposes an artificial intelligence model, based on acoustic and electrical signals during welding, to classify welding tip conditions into three states: normal, cleaning needed, and replacement needed. In collaboration with Heungkuk Metaltech Co. Ltd. (HK), the audio files were collected during the welding of compact track loader rollers (CTL roller). All audio recordings are fixed in 20 seconds with 44.1 kHz sampling rate. These files are processed in time and frequency domains and undergo multi-scale feature extraction. The band energy ratio in certain frequency ranges in three states is calculated and compared. The difference shows the feasibility of prediction based on acoustic signal. At the same time, electrical features are extracted from current and voltage signals recorded during welding. The introduction of electrical signal is based on the change of resistance as the welding tip wears out. After extraction, two sets of features are given to two lightweight classifiers, multilayer perceptron (MLP) and histogram-based gradient boosting (HGB), to get trained and fused using late fusion. To align with real maintenance logic, a cascaded decision strategy is adopted. A first-stage gate distinguishes normal and abnormal, followed by a second-stage classifier separating cleaning and replacement, which has generally better result than direct 3-class prediction.