Transfer Learning for Integrated Nested Digital-Physical Data in Product Performance Prediction

In the field of product performance prediction, the prediction accuracy of deep learning methods is limited by the scarcity of physical data. The integration of digital and physical data has been proposed to enrich the sample, but there is a problem of conditional shift between digital and physical data. Facing the above challenges, a transfer learning method for regression prediction based on conditional distribution embedding is proposed in this paper. Through the pre-training-fine-tuning of the Convolutional Neural Network (CNN) model, a hybrid loss function is constructed to align the conditional distribution discrepancies across the digital and physical data and to preserve the global consistency of the conditional distributions in the target domain. Experimental results based on a test function confirm that this method can significantly improve the prediction accuracy of the model.