Fault-Tolerant Sensor Fusion for Low-Cost Indoor Temperature Monitoring: A Monte Carlo Study of DHT22/AM2302, BME280, and SHT31 Sensors

Low-cost temperature sensors are widely used in Arduino- and IoT-based environmental monitoring, but a single sensor can produce misleading readings when it drifts or experiences an intermittent fault. This study evaluates whether a computationally inexpensive sensor-fusion rule can improve the fault tolerance of a three-sensor temperature node. DHT22/AM2302, BME280, and SHT31 sensors were modeled using manufacturer-reported temperature-accuracy values. For each of five independent fault probabilities (0%, 1%, 5%, 10%, and 20%), 200,000 Monte Carlo trials were generated over a 15–35 °C indoor range, for one million trials in total. Baseline sensor error was modeled as bounded uniform error within the reported accuracy magnitude. Gross faults were added as a transparent stress-test assumption: an additional random signed offset of 1.5–5.0 °C. Four estimators were compared: arithmetic mean, inverse-variance weighted mean, median, and a proposed gated weighted fusion rule that rejects readings more than 1.0 °C from the three-sensor median before re-weighting the retained readings. With no faults, weighted and gated fusion achieved an RMSE of 0.101 °C, compared with 0.141 °C for a simple mean. At a 5% fault probability, gated fusion reduced RMSE to 0.205 °C, 55.4% below the simple mean (0.460 °C). At a 10% fault probability, gated fusion achieved 0.371 °C RMSE versus 0.638 °C for the simple mean. At the most severe 20% condition, the median produced slightly lower RMSE than gated fusion (0.689 versus 0.700 °C), showing the limit of accuracy-weighted fusion when simultaneous faults become common. The results support a simple design principle: accuracy weighting is useful during normal operation, but a robust consensus gate is needed before trusting those weights in fault-prone low-cost sensing systems. This is a computational preprint. No physical DHT22/AM2302, BME280, or SHT31 measurements are claimed. The gross-fault distribution is a simulation stress-test assumption. The manuscript has not been peer reviewed.