Research On Indoor Positioning Algorithm Based On Transfer Learning

With the rapid development of the Internet of Things,wireless local area networks have been widely used and deployed.Therefore,Wi-Fi indoor positioning systems based on Received Signal Strength(RSS)fingerprints do not need to add additional equipment.It has attracted the attention of researchers in the field of indoor positioning.Due to the complex and changeable indoor environment,RSS fingerprint features change over time,which reduces the effectiveness of the offline fingerprint database constructed as a matching basis,resulting in poor performance of the traditional RSS fingerprint positioning algorithm in the online positioning stage.Therefore,how to use the existing offline fingerprints and a small number of labeled online fingerprints to optimize training is the key to improving the algorithm performance of positioning.Transfer learning can effectively deal with the problem of the difference in the distribution of the source domain and the target domain,when the small number of labeled data in the target domain.Therefore,this paper will use transfer learning to improve the RSS fingerprint-based WiFi indoor positioning algorithm.The main tasks are:1.This article first uses the public data set UJI?LIB?DB?v2.1 and self-collected data set for studying the characteristics of indoor RSS.The analysis of experimental results shows that RSS is relatively stable in a short period of time.Through the analysis of experimental results,it can be seen that RSS has the characteristics of short-term relative stability and long-term large volatility.the RSS fingerprints collected by different collection angles are compared,also the RSS distributions are found to be different.These influencing factors have caused the changes in RSS fingerprint characteristics;then using the traditional RSS fingerprint positioning algorithm to predict the positioning of UJI?LIB?DB?v2.1 data set it is verified that the positioning performance of the traditional indoor positioning algorithm is greatly reduced when there is a difference between the RSS fingerprint distribution in the online phase and the offline phase.2.This paper proposes an indoor positioning algorithm based on the maximum mean discrepancy(MMD)of quad-core fusion.The marginal distribution and conditional distribution difference of fingerprints in the offline phase and the online phase are quantified using quad-core fusion MMD.The quad-core fusion MMD is used to obtain the optimal transfer matrix,and the transferred offline fingerprint data is obtained based on the optimal transfer matrix for training the model,and then using model for online fingerprint positioning.The results show the indoor positioning algorithm proposed in this paper effectively reduces the negative effects of the feature difference between the two stages.Only a small amount of online fingerprint samples are required to obtain a reliable model without re-collecting a large number of fingerprints to reconstruct the fingerprint da tabase.3.This article proposes an indoor positioning algorithm based on transferred Convolutional Neural Networks(CNN)model.First,the CNN model structure is designed and the RSS fingerprint is converted to a grayscale image.After the RSS fingerprint s convert into a grayscale image,not only the strength of the RSS value can be reflected by the color depth of the pixel,but the positional relationship of its AP is straightforwardly displayed;Then use a large data set to pre-train the CNN model,save its model parameters.Then using a small number of labeled fingerprints in the self-collected data set to adjust the pre-training model parameters with a transfer model strategy combining freezing and fine-tuning,In order to prevent excessive fine-tuning from overfitting the model,the parameter fine-tuning range is limited by adding regularization,Finally,using transferred model to position the self-collected data set.The experimental results show that the indoor positioning algorithm proposed in this paper can transfer a pre-trained model based on a large amount of labeled fingerprint data of the source domain when the target domain with insufficient label data by adjusting its network parameters reasonably.it can obtain a positioning model with good performance.