Application of Generative Adversarial Network Tabular Data Synthesis for Federal Learning-based Thermal Process Performance Prediction

Process performance prediction now has a fresh and efficient method thanks to machine learning. Existing techniques do not provide good data protection capabilities. The novelty of this work is proposed and validated the use of virtual synthetic thermal processing process performance data as input to machine learning models, where the ‘train on synthetic data - test on real data’ approach is used to pioneer a novel framework for predicting thermal processing process performance. First, the data generated by the table generation adversarial network is applied to the federal learning model for performance prediction. Based on the input-output relationship curve, an evaluation index is proposed for the generation of data for thermal processing performance prediction. Finally, the effect of the generated sample size on the prediction of the machine learning model is investigated. The model is trained using 10,00 synthetic design data and tested using 915 real experimental data. The results show that the synthetic data contribute to the good performance prediction capability of the machine learning model. The use of this method will help to extend the application of federal learning based thermal processing process performance.