Development of an adaptive monitoring system for servo-electric welding gun in resistance spot welding

This master’s thesis proposes an adaptive, data-driven monitoring system to control quality in resistance spot welding (RSW) process. The study focuses on analyzing and segmenting process signals, including dynamic resistance, electrode displacement, voltage, and force signals with the aim of predicting the weld nugget diameter. Two different approaches are investigated in this work: feature-based machine learning and physics-informed machine learning. The feature-based model integrates features segmented from the process signals and utilizes them to train ML models to correlate the features with the weld nugget diameter. Physics-informed models utilize the dynamic resistance signal and physics constraints of the process to predict the nugget diameter. Three different models were tested using the extracted features: Linear Regression, XGBoost, and Neural Networks. The prediction accuracies of the models were R² = 0.48, R² = 0.91, and R² = 0.89 respectively. Physics-informed models included a hybrid model and a GRU autoencoder, with the hybrid model achieving R² = 0.94 and the GRU autoencoder achieving R² = 0.46 for nugget diameter prediction. Models were analyzed and compared by evaluating metrics such as R-squared, RMSE, and MAE.