Toward Foundation Models for Domain Transfer under Limited Target Data in Prognostics and Health Management

Abstract

Domain shift remains a major challenge in prognostics and health management (PHM), where machine learning models trained on one operating condition or machine often experience significant performance degradation when transferred to a different target domain. Although domain adaptation methods have been widely studied for intelligent fault diagnosis and related PHM tasks, many existing approaches still rely on sufficient target-domain data. However, target-domain data are often extremely limited when new machines, sensors, or working conditions are encountered in real industrial applications. This research aims to achieve reliable domain transfer with limited target-domain data and to explore the development of pre-trained and foundation-model frameworks for PHM applications. The proposed research follows a progressive path from conventional deep domain adaptation methods to pre-trained transformer-based models and ultimately toward PHM-oriented foundation models capable of learning transferable representations across heterogeneous industrial datasets. The proposed framework integrates time-series tokenization, transformer-based representation learning, large-scale pre-training, and parameter-efficient downstream adaptation for PHM tasks. Overall, foundation-model-based approaches can provide a promising direction for developing more transferable and data-efficient PHM systems.