AI enhanced multimodal sensor fusion framework for explainable drift-aware decision making in large-scale IoT and cyber physical environments

Abstract A multimodal fusion system for explainable and drift-adaptive decision making is essential in the era of Internet of Things and cyber-physical systems. In this paper, we propose an AI-powered fusion system to alleviate performance degradations due to distributional drift and improve explainability of deep decision models. Specifically, using modality-specific quality-aware encoders along with temporal fusion backbone, we produce calibrated decision scores based on multimodal sensory inputs. Drift monitoring is performed in feature space and prediction space to detect drift with less latency and adaptation is utilized to recover decision quality with minimal drift overshoot. Under the 10-subject leave-one-subject-out evaluation on WESAD, the proposed framework achieved mean macro-F1 and AUROC values of 0.923 and 0.964, respectively, compared with 0.859 and 0.960 for the early-fusion baseline. In the case of SWaT, the framework resulted in a mean macro-F1 of 0.966 for attack category classification, which is 0.055 higher than the LSTM-AE baseline. For the SWaT framework, the AUROC score reached 0.999 while its AUPRC was 0.911, lower than the best baseline AUPRC of 0.975. At the same time, ECE dropped from 6.8 to 2.1%, and the perturbation-based explanation fidelity was above 92%. Meanwhile, the stage-wise SWaT detection rate was above 0.985 with a false alarm rate below 0.02. For deployments of up to 400 simulated nodes, the inference latency was still 43 ms per decision window, peak memory footprint was 472 MB, and communication load was reduced by 46.15% compared to the cloud-only processing option.