Hidden Markov Inference Framework For Error Propagation Mitigation In Modular Digital Twins

In this article, we propose a Hidden Markov inference framework for residual modeling and mitigation in modular digital twins. We first fit a Gaussian Hidden Markov Model (HMM) to the residual sequence to infer latent error regimes, estimate regime transition probabilities, and compute smoothed posterior probabilities of regime occupancy over time. We then apply the Hungarian algorithm to establish a one-to-one mapping between the learned HMM states and meaningful ground-truth regimes. To illustrate how the framework works, we model the physical asset using a higher-fidelity physics simulator and the digital twin using lightweight ARX surrogates across a six-module pipeline. The results on this illustrative example show that the HMM reliably detects regime changes, provides interpretable regime posteriors, and enables state-dependent bias corrections and targeted intervention that reduce surrogate prediction error relative to an uncorrected baseline. Although demonstrated on a toy example, the proposed framework provides insight into hidden error dynamics, regime transitions, and targeted error mitigation in modular, black-box digital twins.