Uncertainty-Awareness in Explainable Artificial Intelligence (xAI) for Sensor Network Metrology

Machine learning models are increasingly used in sensor network applications, but the influence of measurement uncertainty and data quality on model explainability remains insufficiently understood. In this work, we investigate how degradation of input data quality in the form of increasing uncertainty propagates into SHapley Additive exPlanations (SHAP) for machine-learning-based soft sensors in an urban air-quality monitoring network in Helsinki. Two tree-based ensemble models, namely bagging and XGBoost regressors, were trained to estimate ozone concentration using meteorological and air-quality measurements from distributed sensor nodes. A Monte Carlo framework was developed to introduce physically plausible perturbations into the training data using nominal sensor uncertainty levels derived from representative manufacturer specifications and scaled using uncertainty multipliers to simulate progressively degraded sensing conditions. The resulting influence on SHAP feature attributions was analyzed across different geographical sensing clusters. The results show that the global hierarchy of feature importance remains relatively stable under increasing uncertainty levels, while local SHAP attributions become increasingly variable. Temperature and wind-related variables consistently dominate the explanatory structure, but also exhibit the strongest uncertainty-induced broadening in attribution distributions. Furthermore, denser sensing clusters produced more stable explanations, while XGBoost exhibited greater attribution sensitivity to degraded input data quality than bagging models. The study demonstrates that explainability in sensor-network machine learning should be treated as uncertainty-sensitive and distributional rather than deterministic. The proposed framework contributes toward uncertainty-aware explainable artificial intelligence for metrologically informed assessment of sensor-network data quality and interpretability.