Exploring the performance of real-time data imputation to enhance fault tolerance on the edge: A study on environmental data

Real-time data streams from edge-based IoT sensors are frequently affected by transmission errors, sensor faults, and network disruptions, leading to missing or incomplete data. This paper investigates the application of lightweight, real-time imputation methods to enhance fault tolerance in edge computing systems. To this end, we propose to integrate a modular imputation engine on edge system supporting lightweight forecasting models selected for their computational efficiency and suitability to operate on real-time data streams. To assess the performance of different popular lightweight forecasting models for real-time applications, a simulation framework is introduced that simulates the operation of the imputation engine, replicates sensor failure scenarios and allows controlled testing on real-world systems. Imputation accuracy is evaluated using Mean Absolute Error (MAE), 95th percentile error, and maximum error, with results benchmarked against sensor tolerance thresholds. The simulation framework is used to explore imputation on environmental data based on observations collected from a weather station. The findings show that Holt–Winters Exponential Smoothing delivers the highest accuracy for real-time imputation across environmental variables, outperforming simpler models suited only for short-term gaps. Errors grow with longer forecasts, confirming imputation as a temporary solution. Evaluations against sensor-specific thresholds offer practical insights, and execution profiling proves these models are lightweight enough for deployment on low-power edge devices, enabling real-time, fault-tolerant monitoring without cloud dependence.