Radio Map Anomaly Detection For Fingerprint Indoor Positioning Based On GAN

Localization-based service is a dispensable part of modern life.At present,global navigation satellite systems such as GPS?GLONASS?Galileo and china's Beidou,can provide high-accuracy positioning services in outdoor scenarios,while the corresponding technology of indoor localization is still immature,hence,indoor localization has received continuous attention in recent years.Fingerprinting-based localization has attracted intense interests and gained popularity in practical uses due to its convenience,low-cost,energy efficiency,and ubiquitous availability in mobile devices.Its localization effect heavily depends on the real-time reflection of the signal environment by fingerprint map,as a result,we need to conduct anomaly detection on indoor signal environment.Existing anomaly detection methods rely on localization algorithms or traditional machine learning methods,resulting in low accuracy,therefore,we use a large number of easily available fingerprints without location tags to train GAN-based deep anomaly detection model for detecting anomalies of indoor signal environment,it provides better decision basis for updating fingerprint database to improve stability of indoor positioning accuracy.An anomaly detection model based on generative adversarial network called RADGAN is proposed.By learning the signal spatial distribution characteristics of normal location environment from a large number of normal fingerprints without location labels,RAD-GAN can judge whether the location environment has changed.At the same time,in order to deal with the training instability of the GAN-based model and improve the anomaly detection effect,we use the hidden vector of reconstructing fingerprint to generate new one,strengthening the training of the generative network.The experiment results show that the AUC(Area Under ROC Curve)of RAD-GAN is0.96,effectively dealing with the challenge of low recall rate of anomalies.The mechanism of fingerprint attention is proposed to further improve the effect of anomaly detection on indoor localization signal.RAD-GAN may fail to learn certain normal fingerprint distribution patterns,hence,it's difficult for the model to judge the position with weak abnormal changes.We introduce the internal fingerprint attention mechanism to make the generative network better simulate the feature distribution of normal fingerprints.Besides,in consideration of difference between strength of signal source within same fingerprint,we propose the external fingerprint attention mechanism to improve the model's ability to learn the normal fingerprint distribution.The experiment results show that our model improves the average recognition rate of abnormal data at the location where the change is not obvious by 10%,demonstrating that sensitivity to anomalous data of RAD-GAN is apparently promoted.