基于深度负相关学习的室内WiFi指纹定位技术

Deep Negative Correlation Learning for WiFi Fingerprint Positioning

本文提出了一种新的基于深度负相关学习的WiFi室内定位模型。通过负相关学习约束,多个学习器能学习到不同的表征特性,从而有效降低模型的过拟合,并极大地提升其泛化能力。同时该模型将负相关学习方法应用到去噪自编码器和回归预测器上,并利用深度学习方法使其很好地适应随环境和时间变化的RSSI信号,提高了在长时间间隔内的定位性能。利用负相关学习方法使定位模型在初始时候的平均定位误差从1.57 m下降为0.77 m,60 d平均定位误差也仅为0.89 m,误差仅仅只增加0.12 m,验证了负相关学习能够削弱环境变化对定位的影响。

A method based on deep negative correlation learning for WiFi indoor localization is proposed.By adopting the negative correlation learning strategy on autoencoder and regressor modules,this method leads to much diversified decision surface,thus cancelling out the errors from individual models from averaging over all models and resulting in much better generalization ability.The experiments show that the deep negative correlation learning substantially enhances the model performance on the positioning during the long time interval.After implementing the negative correlation learning,this method enables the average localization error at the initial state to drop from 1.57 m to 0.77 m.The average localization error only increases 0.12 m during the 60-day-experiment,which indicates that it can weaken the impact of environmental changes on positioning.