Soft Computing Approaches for Wireless Localization
Name
Kristjan Solmann
Abstract
Localization algorithms in wireless sensor networks have been an interesting field of research for the past few years. The combination of different requirements such as storage space, computational capacities, communication capabilities, and power efficiency make it challenging to develop a localization algorithm. One subcategory of localization algorithms is neural network based algorithms where a model is trained on data from anchor nodes.
In this thesis, the fuzzy transform is introduced into the context of existing neural network-based localization algorithms. The fuzzy transform is proposed as a replacement for dimensionality reduction methods in the algorithms. Before feeding the data into the neural network for the purpose of training, its dimensionality needs to be reduced as it is the main factor of performance. The performance of the fuzzy transform compared to the full dataset and a common reduction method principal component analysis is evaluated by testing on simulated data. The obtained results show that the proposed fuzzy transform based approach outperforms principal component analysis in localization accuracy while being comparable in neural network training time. The fuzzy transform performed comparably or slightly worse with the full dataset in localization accuracy, but showed much shorter training times.
In this thesis, the fuzzy transform is introduced into the context of existing neural network-based localization algorithms. The fuzzy transform is proposed as a replacement for dimensionality reduction methods in the algorithms. Before feeding the data into the neural network for the purpose of training, its dimensionality needs to be reduced as it is the main factor of performance. The performance of the fuzzy transform compared to the full dataset and a common reduction method principal component analysis is evaluated by testing on simulated data. The obtained results show that the proposed fuzzy transform based approach outperforms principal component analysis in localization accuracy while being comparable in neural network training time. The fuzzy transform performed comparably or slightly worse with the full dataset in localization accuracy, but showed much shorter training times.
Graduation Thesis language
English
Graduation Thesis type
Bachelor - Computer Science
Supervisor(s)
Stefania Tomasiello
Defence year
2023