Online Sound Based Arc-Welding Defect Detection Using Artificial Neural Networks

Heavy steel industries are constantly distressed with improving quality, reliability, and robustness of their product. Nevertheless, they are also concerned with reducing welding processes cost, men-power, and material waste. This leads to an interest in control techniques, automation, and robotization. Monitoring the welding process of continuously fed melting wire electrode is still an open problem, due to the instability caused by the electrical variables. We identify expressive gaps in the understanding of stability and quality of welding deposition. In this work, we introduce a low-cost system that monitors the stability of the weld bead and the transfer mode of the MIG/MAG welding process. We propose a non-intrusive solution that uses the sound signal produced by the electric arc to monitor and analyze the process in real-time. In this sense, we propose an Artificial Neural Network (ANN) based methodology to identify some types of discontinuities in the welding process. To support and foster the development of sound signal based neural network, we produce a sound signal dataset based on the experiments which recreate real situations in the welding industry. Processes with adequate conditions and welds with two types of discontinuities were executed to validate the methodology. The model performance is evaluated using both classification accuracy and confusion matrix. Results show to be possible to identify discontinuities in the welding process by looking only at the sound generated by the electric arc.