Machine learning for indoor positioning based on received signal strength

Indoor positioning is a key technology enabler for various smart systems that require location-based optimization and automation. Despite its potential, the field of indoor positioning faces numerous challenges in terms of accuracy, precision, computational complexity, power consumption, robustness, scalability, and cost. This thesis proposes machine learning approach to improve indoor positioning system that makes use of Received Signal Strength (RSS) positioning metric, in terms of accuracy, processing time, and robustness. We consider machine learning approach to be a potential alternative solution due to its ability to model nonlinear functions without a priori assumptions on the distribution of the data and the noise. We present several improvements in location fingerprinting approach with the objective of reducing the number of training data, improving robustness, and decreasing the length of processing time. We combine two existing location fingerprinting methods, which are Location Estimation using Model Tree (LEMT) and LANDMARC, to improve the accuracy and robustness of the location fingerprinting when training points are sparsely distributed. Performance evaluation shows that the proposed continuous-output modification of LEMT achieves 0.5 m to 2 m reduction in the 90th and 95th error percentile. LEMT method and its derivatives, however, have long training time and testing time because these methods rely on M5 model tree which is computationally intensive. The M5 model tree is used in LEMT to model the functional relationship between reference devices and tracking device. To address this problem, we propose the use of Extreme Learning Machine (ELM) regression model in place of M5 model tree to shorten the processing time. We found that incomplete data in the RSS fingerprint can either be imputed with a low RSS or filtered out, without negatively affecting the accuracy of ELM regression. With the use of ELM, the location fingerprinting method experiences 15 to 40 times improvement in training time and 6 to 10 times improvement in testing time, with accuracy degradation around 0.1 m to 0.5 m. Depending on the application involved, this may be considered reasonable or acceptable. We also evaluated the performance of the proposed methods using an open-access benchmark dataset containing no reference device, measured across 15 months. When reference devices are not present in the environment, we can use a subset of training fingerprints as reference data in place of the fingerprints of reference devices. Since we only use a subset of fingerprints, we reduced the amount of work in collecting repeated training data. Our proposed method produces superior accuracy across all months, indicating its ability to adapt to temporal variation despite having minimal amount of repeated training data, especially for the 11th month and beyond when the benchmark methods start to experience significant accuracy degradation.Doctor of Philosoph