Correct-and-then-detect: Adaptive fault detection through data-driven digital twins and transfer learning

In a digital twin-based fault detection framework, it is essential to have a reliable and robust digital twin to generate reference values representing nominal behaviors. Traditional digital twin modeling approaches require deep understanding of functioning principles and failure mechanisms of the system, which are not always available in practice. Furthermore, as systems operate, maintaining an accurate digital twin becomes challenging as system's behavior tends to deviate from its nominal cases due to various factors like aging, uncertainty and sim-to-real gaps. To address these issues and better support predictive maintenance, this paper proposes an integrated framework for data-driven digital twin modeling and fault detection. First, we develop a design-of-experiments-based approach to generate training data that can efficiently cover the potential task space. The generated training data are then used to train a data-driven digital twin, i.e., a deep learning model that predicts system performance at a given time instant given observed covariates from a time window prior to it. Next, we develop a new adaptive fault detection method based on the developed data-driven digital twin. The algorithm works in an adaptive manner: the reference values generated by the digital twin are first compared with raw observations for residual-based fault detection. Once an abnormal observation is detected, its predicted nominal value is used to replace the corresponding observation in the internal feedback buffer before constructing subsequent context windows, thereby reducing fault-induced contamination in later predictions. Finally, a transfer learning mechanism is applied to adapt the data-driven digital twin to system's evolution and cross-domain distribution shifts using verified healthy target-domain data. Real experiments are conducted on two real robots. The results confirm that the data-driven digital twin accurately captures nominal system behavior and that the proposed framework maintains robust fault-detection performance under the considered cross-domain conditions.