A conditional shift-aligned transfer learning approach for prediction based on non-nested digital and physical data

Data-driven prognostics and health management techniques are increasingly applied by performance prediction, but the performance of deep learning models is limited by physical data scarcity. Although digital data can enrich the training samples, the differences of conditional distributions between physical and digital data are also intractable. To tackle the above challenges, in this paper, we propose a new transfer learning method for solving the conditional shift problem in regression prediction. Conditional embedding discrepancy (CED) is designed and used to construct the hybrid loss. Through a branching CNN network, conditional distributions between digital and physical data are aligned and regression prediction is achieved. Experimental results based on the test function demonstrate that the method can improve the prediction accuracy and robustness of the model.