Adaptive Physics-Guided Transfer Learning Model for Predictive Maintenance
Prayas Lohalekar, Twinkle Joshi, Kavitha Thiyagarajan, A C Ramachandra, Zahrah Sataar
Abstract
Hybrid Digital Twins (HDT), which combine physics-based Reduced-Order Models (ROM), have shown strong promise for the predictive maintenance of rotating machinery by enabling realistic synthetic data generation and real-time fault and Remaining Useful Life (RUL) inference. However, existing HDT implementations often require extensive recalibration for each new asset, which limits the domain generalization and practical deployment across various motor types. Hence, this study proposes Adaptive Physics-Guided Transfer Learning (APG-TL) to enable physically consistent adaptation of HDT-based predictive maintenance. Initially, a compact physics ROM through Finite Element Analysis (FEA) and model-order reduction was developed to capture electromagnetic and mechanical dynamics. Subsequently, a One-Dimensional Convolutional Neural Network (1D-CNN) encoder was trained on combined ROM-generated synthetic runs and source hardware data. Furthermore, the ROM for a new target asset is calibrated using Bayesian parameter estimation to fit the physical parameters. Next, unsupervised Domain-Adversarial Adaptation (DANN) is employed to align the encoder features to target operations while enforcing a physics-consistency regularizer. Finally, online ROM parameter updates were deployed with occasional few-shot finetuning for the continuous real-time prediction of faults and loweffort adaptation. The proposed APG-TL achieved better results in terms of accuracy (96.54%) than the existing DT with RUL estimation (DT-RUL).