Indoor positioning using weighted magnetic field signal distance similarity measure and fuzzy based algorithms

• Main Author: Bundak, Caceja Elyca
• Format: Thesis
• Language: English
• Published: 2021
• Subjects: Algorithms; Fuzzy algorithms; Electromagnetic field
• Online Access: http://psasir.upm.edu.my/id/eprint/92752/1/FS%202021%2026%20-%20IR.pdf; http://psasir.upm.edu.my/id/eprint/92752/
• Institution: Universiti Putra Malaysia

Indoor localisation based on the magnetic field has drawn much research attention since they have a range of applications in science and industry. Magnetic-based positioning systems are infrastructure-free and can be sensed by magnetometers embedded on smartphones. Unfortunately, magnetic field intensity data only consists of three components magnetic field signals compared to Wi-Fi, using multiple access points. There is a high chance that a similar reading of those three components obtained at multiple locations. A magnetic-based positioning algorithm should fully utilise the three components of the magnetic field intensity data. This thesis analyses the positioning accuracy changes by using different similarity measures on a specific magnetic field vector and proposed an algorithm using different weighted magnetic field signal distance similarity measures. For the first method, various metric distances used for the MF signal components are studied and the results showed that Euclidean distance and square distance give low distance mean error compared to square root and Manhattan distance. Then, three proposed different signal weighting functions, namely actual weight, square weight, and square root weight are applied in each MF signal similarity measure and compared with the state-of-the-art of Euclidean distance to estimate location. Additionally, the effect of signal weighting function is investigated further using multiple K values of K nearest neighbour (KNN) algorithm. According to the results, the square root weighting function has a lower position error of 8.156 m than Euclidean distance with an improvement of 5.5%. Also, the use of (K=5) of KNN for the square weight of dmy1 distance measure gives the lowest mean estimation error of 7.188 m. Another problem in MF IPS is there are few studies focused on using the Euclidean distance and the area between the reference points to improve the accuracy in the position estimation. Therefore, for the second objective, another algorithm named the fuzzy algorithm is designed which combines the clustering algorithm, matching algorithm, triangle area algorithm and average Euclidean algorithm used to estimate location. Firstly, the MF RPs database is reconstructed into a cluster database using the clustering algorithm. Each trained RP and other nearby RPs are clustered together at a certain distance. A matching algorithm is used to match between the top 10 ranked RPs with the nearest Euclidean distance to the TP with the RPs clustered. For the triangle area algorithm, the smallest triangle area is selected from the triangle formed from the matching RPs cluster to estimate location. In contrast, the average Euclidean algorithm is based on the average Euclidean of the RPs from the RP cluster set. The lowest average Euclidean distance is chosen, and the average estimated location of the RPs is calculated. Lastly, for the fuzzy algorithm, a rule-based decision is applied to select whether the triangle area or average Euclidean algorithm is used to find the final estimated position. The fuzzy algorithm shows a localisation accuracy of 5.889 m, which is better than the KNN and Weight MF signal algorithm with an improvement of 31% and 27% respectively. Both algorithms have achieved the target accuracy below 8 m and better than KNN. Although the fuzzy algorithm achieved better accuracy than the weighted MF signal distance similarity measure, the weighted MF signal distance similarity measure was less time-consuming than the fuzzy algorithm. Therefore, both algorithms need more improvement in future works to achieve a better estimation location.