Deep transfer learning for conditional shift in regression

Deep transfer learning (DTL) has received increasing attention in smart manufacturing, whereas most current studies focus on the situation of marginal distribution shift in classification. We observe a new regression scenario in machine health monitoring systems (MHMS) with conditional distribution discrepancy across domains and try to propose a general theoretical approach for broader applications. In this paper, we propose a DTL framework CDAR, namely conditional distribution deep adaptation in regression. As only few labeled target data is available, in addition to only considering the prediction accuracy of individual samples, CDAR aims to preserve the global properties of the conditional distribution dominated by the target data. Thus, a hybrid loss function is constructed by combining the mean square error (MSE) and conditional embedding operator discrepancy (CEOD) in CDAR, and the target model is able to be finetuned by minimizing the designed loss function through back-propagation. The performance of the proposed CDAR is compared with two classical marginal distribution adaptation algorithms, TCA and DAN, and a specific method of DTL, FA. Experiments are carried out on two real-world datasets and the results verify the effectiveness of our method.