Federated Domain Generalization for Time-Series Classification via Dynamics-to-Domain Generation

Time-series classification supports applications such as activity recognition and health monitoring, but non-stationary and heterogeneous sensor streams often lead to out-of-distribution shifts that are amplified in federated settings. When each client holds only a small and biased subset of data, federated domain generalization becomes difficult because the client provides limited diversity to learn robust invariances. We revisit this challenge through latent segment dynamics and interpret domain shift as variations in fine-grained local behaviors and their temporal transitions. We propose FedD2G, a dynamics-aware generative framework for federated domain generalization. FedD2G decomposes each sequence into segments and learns dual-view dynamic descriptors using a Dual-view Segment Encoder, where time-domain features capture local morphology and frequency-domain features capture periodic cues that often distinguish domains. It then introduces Sequence-Fused Transition sampling with a class-conditioned Markov transition model to assemble descriptors into temporally coherent sequences, enabling controlled synthesis that expands beyond observed source distributions without sharing raw signals. Experiments on cross-domain federated time-series benchmarks show consistent improvements over strong FL and FDG baselines, reaching 78.5% average accuracy under small-scale data availability. We further provide empirical evidence that the learned dynamics reveal latent domain variants and offer effective guidance that helps bridge the source-target gap through distributional expansion.