Author ORCID Identifier
Year of Publication
Doctor of Philosophy (PhD)
Dr. Peng Wang
Welding is a vital manufacturing process across a diverse range of industries to permanently join metallic components into needed assemblies. Though widely used and studied, the process is not fully understood due to the myriad of interconnected physics that present themselves during the welding process. Electrical, thermal, structural, and fluidic mechanics are simultaneously at work influencing one another to achieve the bonding that manufacturers desire. In addition to general manufacturing issues, such as tool wear or misalignment, micro-level changes to the material structure and composition can have cascading impacts on the final weld. This complexity makes it difficult, if not impossible, for certain welding operations to be performed autonomously and inconsistencies in extant welding operations. Fortunately, Machine Learning (ML) techniques are suitably positioned to handle complex and high-dimensional correlations that can be used to develop novel autonomous operations and improve production consistency. This dissertation contains two big modules, highlighting the development and applications of ML for 1) adaptive process control and robotic automation of arc welding; and 2) in-production quality prediction and defect detection in Resistance Spot Welding (RSW).
To highlight how ML can be leveraged to produce a robotic solution to automate classically manual tasks, a vision-based arc-welder adaptive control scheme was developed. Currently, this operation is done by skilled laborers who visually observe the weld pool and adjust their movements based on their experience to achieve a consistent welding bead. This process can be automated by robots through the integration of vision-based weld process perception with data-driven process modeling and adaptive control. Specifically, upon a camera that captures the weld pool evolution, which is fed into a pixel-level image segmentation network is developed to outline the weld pool, allowing an estimate of the pool width to determine the welding quality, and detect split pool occurrences. The estimated pool width was then correlated with process parameters for the welding process modeling through a straightforward neural network. Last, upon the process modeling, an efficient adaptive online control method was developed to dynamically adjust process parameters to reach and maintain a desired welding state.
Additionally, the use of ML for quality control was shown with the realization of an interpretable data-driven model to predict the outcomes of spot welding on coated materials. This was accomplished by first developing an enhanced RSW process model that accounts for the melting of the InterDiffusion Layer (IDL) of the material coating in Press Hardening Steels (PHS). The time of this occurrence, along with other key phases of the welding process, were determined through the analysis of real-time multi-sensing data from a variety of weld scenarios. A bespoke signal processing technique was then developed to automatically identify when these phases occurred. This information was used to develop meaningful features that were used in conjunction with a MLP to predict IDL thickness, which was combined with the known process variables to predict expulsions, the sudden ejection of molten metal.
Digital Object Identifier (DOI)
This study is supported by National Science Foundation under Grant No. 2024614 in 2020.
Kershaw, Joseph, "MACHINE LEARNING FOR ADVANCING AUTOMATION AND QUALITY CONTROL IN ROBOTIC WELDING" (2023). Theses and Dissertations--Mechanical Engineering. 215.
Available for download on Sunday, December 14, 2025