A CNN-based CSI fingerprint indoor localization method

To improve the performance of Wi-Fi fingerprint indoor positioning technology, a method based on convolutional neural networks (CNNs) for channel state information (CSI) fingerprint indoor positioning is proposed. This method fully exploits the feature extraction capabilities of CNNs, applies the combination of amplitude difference and phase difference information as training data in the offline phase, and uses the trained CNN network model for an online test. In the online phase, for different experimental scenarios, by analyzing the variance of the amplitude information and phase information, the amplitude difference and phase difference information of the test data are weighted to obtain a certain universal weight factor for a better positioning result. At the same time, considering the characteristics of terminal mobility during real-time positioning, the CSI information sampled twice in succession is adopted as test data to increase the diversity of test data. To address the disadvantage of poor positioning performance of traditional probability-based positioning algorithms, an improved probability-based fingerprint matching algorithm is introduced. By passing the CSI information of the point to be located through the CNN network model, it can output the probability average value corresponding to the reference position with the highest probability in all test data packets and weight it with the reference position coordinate to estimate the point to be located. In addition, to enhance the universality of the algorithm, a dual-node positioning scheme is proposed for complex indoor scenes to improve positioning accuracy. Experiments are conducted in two positioning scenarios in a corridor and laboratory, including the amplitude difference positioning performance, the average positioning error of each positioning method, and the performance comparison of positioning algorithms. The information joint positioning algorithm obtains an average positioning error of 24.7 and 48.1 cm, which verifies the effectiveness of the proposed algorithm.