Neural networks-based accurate IoT data reconstruction

IoT gateways commonly aggregate sensor readings to reduce bandwidth and energy consumption, but this aggregation prevents cloud applications from accessing fine-grained historical data. This work investigates whether lightweight perceptron-based neural networks can reconstruct original sensor time series from temporally aggregated values with sufficient accuracy and efficiency. We propose a two-stage approach in which edge-trained perceptron models learn local sensor behavior, while a cloud-based model reconstructs full-resolution signals guided by aggregation averages. Extensive experiments were conducted using the Intel Berkeley Research Lab dataset, exploring multiple aggregation rates, window sizes, and training configurations, and comparing the proposed method against six classical interpolation techniques. Results show that while spline-based methods often achieve the lowest reconstruction error, perceptron models remain statistically competitive in terms of MSE and exhibit different runtime characteristics across deployment settings. These findings indicate that lightweight neural models represent a viable trade-off between accuracy and computational efficiency for IoT scenarios with constrained computational budgets.