Date Available


Year of Publication


Degree Name

Master of Science in Mechanical Engineering (MSME)

Document Type

Master's Thesis




Mechanical Engineering

First Advisor

Dr. Peng Wang


Resistance spot welding is a crucial manufacturing process used across a wide range of industries for permanently joining metal components. Characterized by its applications in the automotive industry, resistance spot welding is valued for its speed, efficiency, and relatively low cost to set up and maintain. The process involves running a pulse of electrical current through two metal sheets to liquify the material and create a permanent bond. The process complexity necessitates precise control over various parameters to ensure acceptable results, emphasizing the importance of quality control. Because there are no low-cost and non-invasive techniques to inspect welds, strategies utilizing sensor data are required to determine weld quality. The spot welding process involves complicated physical interactions and a multitude of unknown factors, making it challenging to model using physics based methods. Because of this, machine learning strategies are the best way to capture the complex relationships between sensor data and weld quality. These strategies often rely on real-time data from high-cost sensing methods or large amounts of high-cost lab data for training. These limitations can make quality control cost-prohibitive, highlighting a need for innovative virtual sensing solutions. By using virtual sensing, the need for high-cost sensors and excess lab data can be eliminated, while simultaneously improving the accuracy and versatility of quality predictions.

This thesis introduces a novel approach for virtual signal generation utilizing the power of transformer models, renowned for their effectiveness in capturing complex sequential relationships, combined with transfer learning techniques to produce a model that can quickly be adapted to novel welding scenarios using a small amount of new training data. By fine-tuning a model pre-trained on a diverse set of welding scenarios, the limitations of traditional data-driven approaches can be overcome. The complex relationships between known and desired data are captured by the traditional transformer during pre-training, while the effects of various welding and material properties are captured by separate embedding layers. This results in a model that can be up to 43.4% more effective than a standard transformer at making predictions under conditions that differ significantly from the training data. By generating this virtual sensing data, traditional machine learning models designed for quality prediction can become more generalizable, allowing them to be deployed without the need for investment in additional sensors.

Digital Object Identifier (DOI)

Funding Information

This research lab was funded by the National Science Foundation's Standard Grant (no.: 2237242) in 2023

Available for download on Saturday, April 25, 2026