Overcoming cold start and sensor bias: A deep learning-based framework for IoT-enabled monitoring applications

Abstract

Target localization is an essential aspect of smart environmental monitoring, which has gained prominence thanks to the Internet of Things (IoT) paradigm. Localization is the process of determining the location of an unknown target within an area of interest (AoI) based on data gathered by IoT sensors. Existing target localization works assume full knowledge about the sensors without considering the cold start problem, which may arise when new sensors join the network. This is exacerbated by the problem of bootstrapping in IoT sensors, where the system needs to identify and register unknown sensors. Additionally, the variability in sensors characteristics is inevitable due to the numerous factors that may alter the sensors behavior, raising the problem of sensor bias. Existing works dealing with the aforementioned problems are inadequate, since they involve a pre-processing step and assume the bias to be static over time, resulting in inaccurate localization systems. Therefore, this work proposes a robust and dynamic, holistic localization system that can find the location of an unknown emitting target regardless of the sensors characteristics. The proposed system comprises of two phases: 1) characterization and localization phase; where a deep learning model is utilized to characterize the sensors based on their locations and readings only, whereas Bayesian algorithm is employed to determine the unknown target location by leveraging the predicted sensors characteristics, and 2) correction phase, where an update metric is devised to refine the predicted sensors characteristics based on the estimated target location. The validation of the proposed approach was conducted using real-life and synthetic datasets and compared against an existing benchmark, where it shows an improvement of ∼79%, in terms of both localization and prediction errors. Thus, showcasing the effectiveness and potential advantages of the proposed approach in localizing unknown target(s).