一种基于卷积神经网络的CSI指纹室内定位方法

针对提高Wi-Fi指纹室内定位技术性能，提出了一种基于卷积神经网络（Convolutional neural networks，CNN）的信道状态信息（Channel state information，CSI）指纹室内定位方法。在离线阶段联合定位环境参考点的幅度差和相位差信息，利用CNN进行训练，保存训练后的CNN网络模型作为指纹；在线阶段，针对不同实验场景，对测试数据的幅度差信息和相位差信息进行加权处理，引入改进的基于概率的指纹匹配算法，利用待定位点的CSI信息并通过CNN网络模型预测待定位点的坐标。此外，为增强算法普适性，针对复杂室内场景，提出了双节点定位方案来提高定位精度。在廊厅和实验室室内两种不同定位场景进行了实验，信息联合定位算法分别获得了24.7 cm和48.1 cm的平均定位误差，验证了基于CNN的CSI幅度差和相位差联合定位算法的有效性。 

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.