A CSI fingerprint map-based indoor positioning algorithm using stacking fusion model for enhanced localization accuracy

The traditional indoor location sensing based on wireless signals is deeply affected by two factors, which are the impurity of source data and the singularity of location sensing model. To tackle this problem, we propose a Channel State Information (CSI) fingerprint sensing position algorithm based on Stacking Fusion Model (SFM) in this paper. Taking advantage of different data prediction and training models, we build a location sensing prediction model that combines Stacking ensemble learning with the corresponding machine learning algorithm and it is structured into two layers of learners. The SFM method is initially the first layer of base learners, four indoor location sensing algorithms: Support vector Regression (SVR), K-nearest Neighbor Regression (KNR), Random Forest (RF) and Weighted Mixed Regression based on SVR&KNR (WMR-SKR) are trained, and the prediction results are generated and output. Subsequently, the prediction results from the first layer are used as inputs by the second-layer learner, which trains a linear regression model. After the same processing, the CSI data of the sensing target is matched with the data output from the second layer of processing in the offline stage, thus completing the final location sensing result. We implement verification sensing experiments in two real indoor environments, and the results show that compared with only using the above four algorithms, the position sensing accuracy of SFM is improved 35.56%, 40.49%, 22.01%, 17.18% respectively in the laboratory environment, and by 38.61%, 59.34%, 34.74%, 23.46% respectively in the empty classroom environment. Considering the universality of experimental verification, in addition to the comparison with the four primary learners, the comparison with the external classical algorithm WKNN is introduced. The experimental results show that the positioning accuracy has also been further improved in the two environments. It should be mentioned that the above results can be achieved not only attributed to SFM, but also to the CSI data preprocessing algorithm proposed in this paper.