Ensemble machine learning model for prediction of inner welding defects in orbital GTAW (Gas Tungsten Arc Welding) process with monitored by multi-sensor system

Abstract

This study proposes a novel real-time defect detection system based on multi-sensor fusion, integrating an infrared (IR) camera, a laser line scanner, and an acoustic sensor to monitor welding conditions during the GTAW process for carbon steel and stainless steel. Sensor data are processed to extract thermal, geometrical, and acoustic features, which are used to train two sub–machine learning models based on artificial neural networks (ANN) for detecting porosity and lack of fusion defects. The sub-model outputs are combined into an ensemble model achieving an average accuracy of 94.3 % in K-fold cross validation, but this restricts applicability to carbon/stainless steel welds unless retrained. The system can predict defects within approximately 10 s after welding, reducing inspection time by about 30 times compared to conventional X-ray methods, enabling substantial improvements in industrial process efficiency. The proposed approach offers a distinct advantage over prior single-sensor systems by combining complementary thermal, and acoustic sub-machine learning model into a unified predictive model, enabling not only defect detection but also accurate classification of defect types.