Ultrasonic welding of carbon fiber reinforced thermoplastics (CFRTP) is a key technology in advanced manufacturing, owing to its potential for rapid, energy efficient assembly of lightweight, high-performance components. In order to reveal the ultrasonic welding behavior of CF/PPS and CF/PEEK laminates, our research group are conducting study using a servo press unit and CNT added energy director[1]. However, the intricate interplay of material properties, welding parameters, and process variations introduces nonlinearity in joint quality, making conventional predictive methods inadequate.

To deal with these challenges, machine learning have increasingly been employed to improve weld quality prediction. Several studies have demonstrated the potential of data-driven approaches in modeling welding processes. However, most have focused exclusively on either static experimental conditions or in-process monitoring data. Only a few efforts have attempted to merge these two distinct data streams, leaving a gap in understanding of the combined effects of experimental parameters and real-time process dynamics.

To address this gap, our study developed a novel two-stage approach. In the first stage, regression models were developed using experimental setup conditions. Tree-based ensemble methods such as Random Forest and Gradient Boosting Decision Trees (GBDT) were employed due to their interpretability. The GBDT model outperformed the Random Forest, achieving a coefficient of determination (R²) of 0.80 and a mean absolute percentage error (MAPE) of 5.05 %. Model interpretation identified CNT content as a dominant predictor, while parameters such as holding load showed a lesser effect. In the second stage, the integration of process monitoring data was achieved via a Support Vector Regressor with an RBF kernel. This integration improved the prediction accuracy further, increasing R² to 0.86 and reducing MAPE to 4.00 %.

By segregating and subsequently integrating two data streams, two-stage approach have the benefits of both methods: the interpretability for process optimization, and the high accurate prediction considering variations from welding process to process. This approach offers a potent tool for understanding and optimizing the ultrasonic welding process in CFRTP assemblies, data-driven decisions in manufacturing.

[1] S. Nishimura et al., in Proc. 7th International Conference & Exhibition on Thermoplastic Composites, 2024, pp.127-130.