Development of an AI-based acoustic disturbance-detection method for robotic arc welding processes

Disturbances generated during arc welding processes can be detrimental to the quality of welded structures, and existing automated disturbance-detection methods lack the capability for real-time deployment. This paper proposes an AI-based disturbance-detection framework for gas metal arc welding using microphone-recorded acoustic signals. Representative disturbances – shielding-gas interruption, tip wear, and cutting-oil contamination – were experimentally reproduced and acoustically recorded under three welding conditions. Acoustic features were then extracted from the Mel-spectrograms of the recorded welding sounds and combined with welding parameters (current, voltage, and travel speed) to train a multilayer perceptron classifier capable of identifying both the occurrence and type of disturbance. The trained model achieved an overall accuracy of 81.5% and a macro-F1 score of 82.9%, demonstrating reliable generalization performance. Time-series evaluation indicated that the model could maintain a stable classification performance from the early stage of welding and immediately after the onset of disturbance. Furthermore, a SHAP (Shapley Additive exPlanations) analysis revealed that the decision criteria of the model were physically interpretable: high-frequency attenuation was dominant in shielding-gas interruption, while low-frequency vibration components were characteristic of tip wear. Both the spectral intensity and its variance were identified as key features for accurate disturbance classification. The proposed approach provides a low-cost, noncontact, and real-time monitoring solution that can be easily integrated into robotic welding systems and adapted to various industrial environments, thereby contributing to the realization of autonomous and explainable in-process quality assurance in smart manufacturing.