Deep Learning Based Indoor Positioning System Using WiFi Trajectory CSI

Abstract

This dissertation presents a deep learning solution to a WiFi-based indoor positioning system (IPS). The state-of-the-art IPS depends on the radio map constructed by the fingerprints, which are represented by the characteristics of radio signals. The WiFi signal's channel state information (CSI) is prevalently utilized as the fingerprint of positions of interest in the indoor environment due to its fine-grained information. The performance of the IPS is often limited due to the multipath fading effects, especially in indoor environments involving multiple non-line-of-sight propagation paths. When the multipath fading effects are severe, the fingerprints of corresponding positions become unreliable; hence, the performance of the IPS is heavily degraded. Furthermore, the mapping function between the input of the IPS and the output of the predicted positions can usually be represented by various conventional classification algorithms. However, the conventional algorithms exhibit complicated calculation costs and low accuracy, especially in the case of enormous prediction space. In this dissertation, to solve the issues mentioned above, a novel representation of fingerprints that utilizes trajectory CSI observed from predetermined trajectories is introduced. Additionally, a novel IPS utilizing the trajectory CSI is proposed. The trajectory CSI continuously collected along the predetermined trajectories is fundamentally different from the conventional CSI collected from the predetermined stationary positions. The CSI along a trajectory can be captured as a whole in the observation. The trajectory CSI can then be utilized to boost the accuracy and robustness of the IPS. To solve the laborious and time-consuming process of collecting trajectory CSI, a generative adversarial network (GAN) is employed. The GAN can enlarge the training dataset by just a limited amount of available trajectory CSI. Therefore, an augmented training dataset can be effectively acquired, significantly reducing the cost of preparing the datasets for the model training. To fully exploit the trajectory of CSI's spatial and temporal information, a deep learning network of a one-dimensional convolutional neural network–long short-term memory (1DCNN-LSTM) is adopted. Besides, this dissertation presents a transformer-based IPS, an endeavor to first apply the popular neural network used in natural language processing (NLP) to indoor positioning problems. The transformer neural network is capable of parallel computation with its attention mechanism reduces the training time tremendously compared to the conventional recurrent neural network system. The transformer can capture the dependencies between the CSI of the positions in a trajectory in an indoor environment. The dependencies then can be further utilized to boost the positioning accuracy significantly. To fully exploit the relations among positions, the direction of arrival is estimated from the WiFi signal and used with the collected CSI as inputs to the transformer-based IPS. The performance of the IPS based on 1DCNN-LSTM and the IPS based on the transformer, in this dissertation, is evaluated by a radio frequency testbed, which is hardware-implemented using digital signal processors and a universal software radio peripheral for both access point and mobile device of WiFi. Based on the trajectory CSI, the proposed IPS is verified to far outperform the state-of-the-art IPS through extensive experimental tests and computer simulations.