Domain Adaptation for Localization Using Combined Autoencoder and Gradient Reversal Layer in Dynamic IoT Environment

The popularity of Received Signal Strength (RSS) fingerprint-based indoor localization is mainly due to ubiquitous nature of Wi-Fi signals. However, environment changes, device heterogeneity and change in Access Points (APs) results in domain shift between offline and online RSS fingerprints. This article proposes a novel Domain Adversarial Neural Network for Regression (DANN-R) over a compressed RSS representation derived from Autoencoders used as a dimension reduction technique to alleviate the challenges of a dynamic IoT environment. In addition, adversarially learn domain-invariant representation in DANN-R using gradient reversal layer (GRL) mitigates these RSS fluctuations by learning a common representation, where source domain (offline RSS data) and target domain (online RSS data) cannot be distinguished. The proposed method outperforms both state-of-art machine learning algorithms and deep domain adaptation frameworks on two public localization testbeds.