Machine Learning-based MIMO Indoor Positioning

Abstract

The most widely used positioning system is Global Navigation Satellite System (GNSS), which uses traditional positioning techniques and cannot achieve satisfactory positioning performance in indoor scenarios due to Non-Line-of-Sight (NLoS) transmission. Fingerprinting is a non-traditional positioning technique that is robust to NLoS transmission in indoor scenarios. Moreover, Applying Machine Learning (ML) to fingerprinting positioning can significantly improve positioning performance. Therefore the main objective of this project is to investigate the effect of different Multi-Input Multi-Output (MIMO) antenna topologies, the number of MIMO antennas, ML algorithms, and Channel State Information (CSI) fingerprints on the performance of ML-based fingerprinting positioning. The four open-source datasets used for investigation were measured on the Massive MIMO testbed of ESAT-TELEMIC at KU Leuven. Three datasets were collected when Uniform Rectangular Array (URA), Uniform Linear Array (ULA), and Distributed ULAs as Base Station (BS) under Line-of-sight (LoS) transmission, and one dataset was collected on URA BS under NLoS transmission.

The antenna topologies studied in this project are three 64-antenna topologies and five 8-antenna topologies. The ML algorithms studied are Support Vector Regression (SVR), Fully Connected Neural Network (FCNN), and Convolutional Neural Network (CNN). The fingerprints studied are Channel Impulse Response (CIR) and Channel Frequency Response (CFR). The number of antennas studied is 8-antenna ULA, 16-antenna ULA, and 32-antenna ULA. The positioning error measures the fingerprinting performance, which is the Euclidean distance between the predicted and ground truth coordinates. All comparisons are presented using the empirical Cumulative Distribution Function (CDF) curves of the positioning error.

The investigation results show that increasing the number of antennas of ULA improves positioning performance. CIR fingerprints and CFR fingerprints have comparable positioning performance, 64-antenna URA has the best positioning performance, and the 8-antenna random array has the best positioning performance. The two Deep Neural Networks (DNNs), FCNN and CNN, have much better positioning performance than the traditional ML algorithm, SVR. However, the difference between the positioning performance of the two DNNs is negligible.

Popular Abstract

Over the past decades, Machine Learning (ML) has become increasingly popular as hardware computing power has increased. ML is widely used to implement Artificial Intelligence (AI), and unlike traditional computer programs, ML algorithms enable computer programs to achieve performance improvements as program inputs are updated automatically. This capability makes ML promising for a wide range of applications in other fields, including wireless fingerprinting positioning. The main objective of this project is to investigate the effect of different Multi-Input Multi-Output (MIMO) antenna topologies, the number of MIMO antennas, ML algorithms, and Channel State Information (CSI) fingerprints on the performance of ML-based fingerprinting positioning.

I will use an example to explain what is the machine learning concept. Suppose a teacher has two stacks of cards. The first pile has pictures of different animals on the front and the animal's name on the back of each card. The second set of cards also has pictures of the same group of animals, but they are taken from different angles and do not have the names of the animals on the back of the cards. The teacher shows the students the front and back of the first pile of cards and teaches them to match the pictures of the different animals with their names by identifying their features. For example, a panda has black and white fur and a big round head, and a giraffe has a long, thin neck and a huge orange-brown body. With the teacher's guidance, the students successfully mastered the ability to match the appearance of different animals with their names. The teacher then showed the students a second set of cards, and they tried to answer the names of the animals after looking at the pictures. In this type of test, students can make mistakes. After all, there is no such thing as a perfect student or teacher. For example, the teacher may not have found the best description of the animal, and the student may not be a good learner. Different students may be good at learning different subjects; some are good at learning from pictures, while others are better at learning from words. The above process of student learning is the same as the process of learning ML algorithms, where the teacher and the cards represent the data set in ML. The first pile of cards with name labels represents the training set, and the second pile of cards without name labels represents the data samples awaiting to be studied. Students represent ML algorithms, and students who are good at learning different subjects represent different ML algorithms.

In this project, the example becomes that of a teacher showing students a series of CSI and their corresponding location coordinates. Students learn the correspondence between the two, and then students observe CSI they have never seen before and can predict the location coordinates corresponding to this new CSI. The following example can help us understand the relationship between CSI and location coordinates. Imagine a room with a mobile wireless transmitter and a fixed receiver. When a signal is transmitted, the transmitter is at location A, and the receiver receives version A of this signal. Then, the transmitter moves to another location B, in the room and sends the same signal again, and the receiver receives another version B of the same signal. Are the two versions of the received signal the same? The answer is no. During signal propagation, the same signal passes through the wireless channel differently due to the relative positions of the transceivers and the environment. These two wireless channels "distort" the signal differently to different degrees. So when the signal reaches the receiver, it already bears the imprint of the unique wireless channel it has experienced. This different imprint is the CSI, which can be associated with different transmitter locations since different transceiver relative locations constitute different wireless channels. Some ML algorithms are better at learning this correspondence than others, so this project aims to explore and compare which ML algorithms are better at learning the relationship between CSI and its corresponding location. In addition, the measurement system in this project consists of many antennas rather than a single antenna. A MIMO antenna is a receiver that can view the wireless channel from different angles. Topologies and the number of antennas also determine the richness of the CSI from the angle perspective. The CSI has different expressions, which are different fingerprints, and the choice of fingerprints also affects the positioning performance. Therefore, this project also investigates the effect of different topologies of MIMO antenna and the number of antennas and fingerprints on the positioning performance of different ML algorithms.

• author: Chen, Qiyi
• supervisor: Xuesong Cai; Guoda Tian
• organization: Department of Electrical and Information Technology
• course: EITM02 20222
• year: 2023
• type: H2 - Master's Degree (Two Years)
• subject: Technology and Engineering
• report number: LU/LTH-EIT 2023-909
• language: English
• id: 9110826
• date added to LUP: 2023-02-16 13:31:42
• date last changed: 2023-02-16 13:31:42