Handling WSN Communication Faults at the Edge with Confidence Attribution for Data Imputation

Due to the nature of Wireless Sensor Networks (WSN), several factors can interfere with sampling and communication. These faults may compromise data quality and disrupt timing and power requirements due to re-sampling and re-transmission. Recent works in the literature propose data imputation and confidence attribution mechanisms as alternatives to overcome missing, or bad-quality, data. Nevertheless, these solutions often lack mechanisms to evaluate the quality of data imputation on-the-fly. This work combines a confidence attribution mechanism previously proposed by the authors with a Deep Autoencoder (DAE) to promote an effective data imputation mechanism able to handle transient faults in WSNs. We rely on the ability of Deep Autoencoders to learn data correlation in order to attribute confidence to data based on the loss of information in the encoding-decoding process. The fine-tuning of the confidence attribution parameters considers the discrepancy between the original confidence attribution method and the loss of information calculated for original and predicted data. Finally, through a case-study, we demonstrate that the DAE-based confidence attribution is capable of matching the confidence attribution that relies on the comparison between original and predicted data in more than 86% of the cases, without requiring the original data.