Deep Varitional Negative Correlation System For Indoor Localization

WiFi sensors are widely used in the field of indoor positioning.At present,a large ma-jority of Wi Fi fingerprint-based indoor positioning methods use a two-stage approach.In the first stage,RSSI(Received Signal Strength Indication)signal data is collected and recorded offline,and the corresponding algorithm is used to train the model to fit its coordinates.The second stage is online testing,which is used to evaluate positioning accuracy.Although the two-stage fingerprint positioning has good performance,its positioning accuracy will decrease significantly as time passes and environmental changes.The main reason is that the model can-not adapt to the drifting of the RSSI signal since it would change over time,temperature and humidity which leads to the degradation of the positioning model.How to develop the long-term stability of the Wi Fi fingerprint indoor positioning model is the main research focus of this thesis.To tackle this problem,the paper proposes a new deep variational negative correlation positioning model by enhancing the representation ability of multiple ensemble learners.The model has the following two contributions.It is the first model to apply the negative correlation learning method to the field of indoor positioning.By imposing negative correlation constraints on the reconstruction loss and the coordinate regression loss,each learner can exploit different hidden layer features,reducing mutual covariance,thus effectively overcome the occurrence of overfitting and greatly improve the generalization of the model.Moreover,we combine the variational inference with negative correlation learning for the first time in indoor positioning.Through the variational autoencoder,the hidden layer features subject to the continuous Gaus-sian normal distribution,which further improves the robustness and adaptability of the model.In the experimental environment,the average localization error at the initial time is reduced from 1.57 m to 0.77 m and the average localization error is only 0.89 m in 60 day's interval by our deep negative correlation learning model.While adopting the deep variation negative cor-relation model,the average localization error at the initial time is reduced from 1.57 m to 0.74 m and the average localization error is only 0.86 m in 60 day's interval.In conclusion,our model has significant improvement comparing with other comparative models.Experiments show that the combination of the negative correlation method with variational inference can not only sig-nificantly improve the positioning accuracy in a short time interval but also make the model work stably in a long time interval,which can effectively improve the positioning accuracy of the model.In addition to the deep variational negative correlation positioning algorithm,another core contribution of this paper is that we design and implement a convenient positioning system for the users.The system includes an embedded collector,mobile App and Linux cloud platform.The embedded module adopts ESP32 to collect RSSI signal,then the data is sent to mobile App and cloud platform through Bluetooth and NBIOT devices in the format of JSON message.The mobile App can display real-time RSSI data,providing coordinated information and facilitating users to track the Wi Fi-based anchor station.Linux cloud platform can store data of different periods and different collection devices,record historical motion coordinates,query historical data intersection and other operations,which is friendly for users to analyze.Finally,the deep variational negative correlation system platform achieves high accuracy and high sensitivity,which makes users to use the location service more intuitively and conveniently.