Neural Network Fuzzy Control of TIG Welding Backside Weld Width

Literature Overview

This paper by Gao Jinqiang, Wu Chuansong, and Liu Xinfeng from Shandong University, published in the Journal of Welding in 2001, addresses the challenge of real-time control of backside weld width in TIG welding processes. The authors designed a single-layer neural network fuzzy controller capable of automatically learning fuzzy control rules and adjusting them in response to system changes. A CCD camera was used to capture the front-side weld pool image, from which geometric parameters were extracted and used to predict and control the backside weld width through an established correlation model. This work was supported by the National Natural Science Foundation of China (Grant No. 59875053).

Technical Approach and Control Architecture

The TIG welding process is characterized as a highly nonlinear, strongly coupled, and time-varying system, which presents significant challenges for conventional control strategies. The authors recognized that traditional PID controllers are inadequate for such complex systems and proposed a neural network fuzzy controller as a more suitable solution.

The control architecture consists of several interconnected components. The CCD camera captures the front-side weld pool image in real time, and image processing algorithms extract geometric parameters such as pool width, pool length, and pool area. These parameters are then fed into a correlation model that relates front-side pool geometry to backside weld width. The neural network fuzzy controller uses this predicted backside width as feedback to adjust welding parameters such as current, travel speed, and torch angle.

The single-layer neural network fuzzy controller is designed to automatically learn fuzzy control rules from training data and adapt these rules as the system behavior changes. This self-adaptive capability is essential for handling the time-varying nature of the welding process, where factors such as joint fit-up variation, thermal distortion, and material property fluctuations continuously alter the process dynamics.

Process Characteristics and Control Challenges

The backside weld width in TIG welding is a critical quality indicator, particularly for single-pass welding of thin-section components where backside bead profile directly affects structural integrity and leak tightness. In pipeline welding applications, where single-pass full-penetration welds are common for thin-walled pipes, controlling the backside weld width is essential for ensuring weld quality without the need for backside cleanup.

The correlation between front-side pool geometry and backside weld width is not straightforward. Factors such as plate thickness, joint gap, welding current, travel speed, and arc force all influence the relationship. The neural network fuzzy controller addresses this complexity by learning the nonlinear mapping from front-side parameters to backside width and continuously adapting its control rules to maintain the desired backside width within specification.

The learning algorithm employed by the controller enables it to improve its performance over successive welds. This adaptive learning capability is particularly valuable in production welding environments where component geometry and material properties may vary from one weld to the next.

Engineering Practice Implications

The concept of real-time weld pool monitoring and feedback control has significant implications for automated welding systems in pipeline and pressure vessel fabrication. In the production of small-diameter pipes and pipe fittings, where manual welding may not achieve consistent backside bead profiles, automated systems with real-time monitoring can significantly improve quality and reduce rework.

For pipe fabrication shops producing ERW, HFW, or welded fittings, the principles demonstrated in this study can be extended to other welding processes such as submerged arc welding and flux-cored arc welding. The integration of visual monitoring with adaptive control represents a paradigm shift from open-loop parameter setting to closed-loop quality assurance.

The use of CCD imaging for weld pool monitoring also opens possibilities for other quality indicators, such as burn-through detection, lack of fusion identification, and weld bead profile assessment. These capabilities can be integrated into a comprehensive welding monitoring system that provides real-time quality feedback and process optimization.

Reflections and Study Insights

The work presented by the authors represents an early but pioneering application of intelligent control to welding process optimization. The combination of neural networks and fuzzy logic addresses the limitations of both approaches individually, leveraging the learning capability of neural networks and the rule-based reasoning of fuzzy logic.

From a modern perspective, the concepts demonstrated in this study have been further developed through advances in data analysis, data analysis, and computer vision. However, the fundamental principles of real-time weld pool monitoring, parameter extraction, and adaptive control remain highly relevant. The challenge of establishing reliable correlations between observable front-side parameters and critical back-side quality indicators continues to be an active area of research.

This study serves as an important historical reference for engineers interested in intelligent welding control systems and demonstrates the potential of combining visual sensing with adaptive control to achieve consistent weld quality in complex manufacturing environments.