XDrift-RVFL: Drift-Adaptive Interpretable Learning for Resource-Constrained Edge Devices with Attribution Fidelity Guarantees

Tiny Machine Learning deployments on microcontroller-class hardware face three simultaneous demands: accurate prediction, adaptation to concept drift, and trustworthy explanations of model decisions. While prior work has addressed the first two independently, no existing method provides explanations that remain valid during online drift adaptation on devices with kilobyte-scale RAM. This paper introduces XDrift-RVFL, a windowed Random Vector Functional Link (RVFL) network co-designed with an Explanation Fidelity Certificate that provably bounds the change in feature attribution after each model update. The certificate exploits the closed-form Woodbury update structure to derive the Attribution Shift Bound (ASB), a per-sample upper bound on attribution change computable with zero asymptotic overhead relative to the adaptation step itself. When the bound falls below a calibrated threshold, the previous explanation is guaranteed valid and no recomputation is required. Experiments on the CWRU Bearing Fault, UCI Human Activity Recognition, and FEMTO-ST run-to-failure bearing datasets demonstrate that XDrift-RVFL matches or exceeds the predictive accuracy of the state-of-the-art Lite-RVFL baseline while providing the only method in the comparison that offers per-sample feature attributions, eliminating up to 97.6% of explanation recomputations and fitting within 157.5 KB RAM on the STM32F407VG microcontroller.