Deep conditional shift alignment in transfer learning for regression under complex data conditions

doi:10.1016/j.eswa.2026.133288

Deep conditional shift alignment in transfer learning for regression under complex data conditions

2026-09-29

0

PRE

AI

DOI：10.1016/j.eswa.2026.133288

Original

Original request for help

Save

Abstract

En 中文

Deep learning is increasingly adopted in engineering applications. Transfer learning is a practical remedy to the scarcity of physical data and discrepancies between digital and physical data. However, the problem of conditional shift in regression prediction remains a challenging yet underexplored area. In this study, a new transfer learning method is proposed for conditional shift alignment in regression problems, considering its applicability in dealing with qualitative-quantitative mixed variables and non-nested data cases. A new conditional embedding discrepancy, grounded in the theory of conditional distribution kernel embedding, is designed to quantify inter-domain conditional distribution differences. This metric is then combined with the conditional embedding operator discrepancy to form a dual conditional alignment mechanism. Based on deep convolutional neural networks, inter-domain conditional distributions are adaptively aligned, and regression prediction is achieved through model pre-training and fine-tuning. The method is validated using three numerical examples and two application cases, including the reliability analysis of truss bridges and battery state-of-health estimation. Comparisons with prior studies and ablation variants demonstrate that the proposed method performs better in prediction accuracy and model robustness under complex data conditions.

Keywords:

Transfer learning

Conditional shift

Conditional kernel embedding

Regression

Deep conditional shift alignment in transfer learning for regression under complex data conditions

Abstract

Journal

Organization

Cited Papers