Quality Monitoring and Control Algorithm in Arc Welding Process Using Deep Learning

Abstract

Welding has been playing a critical role in industrial manufacturing worldwide. Automation lines using welding robots have been built in many manufacturing lines, and as a result, productivity has been improved. However, when unexpected disturbances such as welding deformation and weld shape change occur, it is impossible to adequately cope with them during welding, resulting in poor quality of the welds. Therefore, complex welding operations still depend on the experience of skilled welders. To meet the requirements for manufacturing efficiency, cost reduction, and quality, traditional welding manufacturing must be developed into smart welding manufacturing. Therefore, in this thesis, we developed a deep learning-based prediction model that monitors the back bead and weld gap related to welding quality in real time, and control the motion of the welding robot based on the quality prediction results. This thesis is divided into three main parts. First, The back-bead is often not uniformly generated owing to the change that occurs in the narrow gap between the base metals during butt joint GMAW. To solve this issue, we developed a convolutional neural network-based back-bead prediction model. Specifically, scalogram feature image data were acquired by performing Morlet wavelet transform on the welding current measured in the short-circuit transform mode of the GMAW process. The acquired scalogram feature image data were then analyzed and used to develop labeled weld quality training data for the convolutional neural network model. The model were compared with welding data acquired through additional experiments to validate the proposed prediction model. The prediction accuracy was approximately 93.5%, indicating that the findings of this study could serve as a foundation for the future development of automated welding systems. Second, for many years, the welding of hull blocks in narrow and restricted working environments has primarily relied on the experience of skilled welders. To automate this process of hull block welding, detecting weld gaps and controlling the weld deposition rate are the main issues that need to be addressed. Therefore, we propose an algorithm based on a convolution neural networks (CNNs) for weld gap detection and weld deposition rate control. First, through a welding experiment, the welding quality resulting from weld gap variations was analyzed. Subsequently, the welding current and arc voltage signals measured during the welding experiment were transformed into signal waveform images, and the proposed CNN model was trained on the image data labeled according to the welding quality. The developed CNN-based weld gap detection model exhibited an average prediction accuracy of 91%. Third, In the hull block welding process, it is difficult to produce a constant external weld bead owing to changes in the weld gap in a weld seam, thereby failing to meet the required theoretical throat thickness specification; therefore, for decades, the hull block welding process has relied on the experience of skilled welders. Herein, we propose an innovative automated welding system based on deep learning that can monitor the weld gap in real time and control the welding robot to ensure good welding quality. The welding signal generated in the welding process is synchronized with the robot motion by a trigger device, and feature variables are extracted and analyzed in the time and frequency domains. The extracted feature variables are used as input in the trained deep neural network model, and the robot is controlled by the classification result. In the off-line performance verification for new welding data, an accuracy of approximately 93% is achieved. A decision-making system is then developed to control the weld deposition rate to meet the weld quality (theoretical throat thickness) requirements for each weld gap, and is successfully verified in the real-time test with a linearly varying gap from 0 to 5 mm. Through the three studies proposed in this thesis, a real-time welding quality monitoring system based on deep learning was constructed and its performance was evaluated.