Robotic MAG welding defects and quality assessment with a defect threshold decision model-driven method

Research question

How can multimodal process signals reveal defects during robotic welding of excavator booms?

Method

The system acquires pool images, arc sound, current and voltage, with Res2-MobileNetV3 extracting lightweight visual features. A defect-threshold decision strategy combines these features for monitoring and is evaluated in excavator-boom welding.

Main findings

Recognition accuracy was 96.59% on the paper's test set. Separate excavator-boom trials reported approximately 90% average recognition accuracy for four tested defect types, with alarms within 2 s.