Improving sensor data quality with predictive models

Internet of Things (IoT) applications relies on sensors to understand and control physical environments. Sensors are subject to numerous potential faults and sources of inaccuracy, including noise, drift, biases, outliers, and missing data. Errors in sensor reading potentially lead to poor quality of results in applications, diminished accuracy for inferences, and incorrect control decisions. The error consequences range from wasted resources and sub-optimal decisions in monitoring applications to potentially calamitous actions in critical control applications. This work aims to improve data quality in sensing applications by combining noisy and faulty sensor readings with predictive models based on Recurrent Neural Networks (RNN). While imprecise sensors and RNN models fail to measure environmental variables when taken in isolation accurately, we found that a simple linear combination of sensor readings and predictions leads to improved data quality. We developed a heuristic that dynamically attributes coefficients in the linear combination based on the observed and expected mean and variance in sensor data. We found that in scenarios where sensor noise was significant, the heuristic improves data quality compared to sensor data and RNN predictions used in isolation. For a climatic monitoring application with sensors subject to different types and magnitudes of noises and faults, Root Mean Square Error (RMSE) for temperature estimation was reduced by an average of 33% when compared to sensor readings and by an average of 44% when compared to RNN predictions. Similar results were observed for an application monitoring electrical consumption.