With the expansion of location-based services over the last decade, indoor localization in different networks has gained a lot of study attention and makes it crucial to have the best position prediction for users based on the received signal strength (RSS). Several localization techniques have been established in recent years that leverage the devices' transmitted signal to locate people. In this thesis, we aim to first show the possibility of transfer learning in indoor localization from 2.4 to 5.2, and secondly utilize different algorithms to do the position prediction with RSS and compare their performance. To start the analysis, due to data generalizability, we used the dataset gathered by the group of researchers at the University of Victoria, containing five AP sending signals on 2.4 and 5.2GHz. The fingerprint localization contains two phases; offline (testing) and online (training), which work with inter or intra- technology methods. We look at three different methods in this thesis; LN-T, RF- T, and R-IDW, and compare them to traditional approaches based on basic interpolation, using alternative methods such as hyper-parameter optimization and experimenting with different portions of the datasets for training and testing based on k-fold in varied availability ratios of 25%, 50%, 75%, and 100% to simulate the scenario. The results demonstrate that transfer learning from 2.4 to 5.2GHz is possible and inter technology works better than intra technology. Different methods for availability ratios can have various results based on the distribution of the positions for the training phase, one can work better for some and the worst for others. RF-T generally has the best performance with a maximum of 0.8767m as the localization error compared to the BL. R-IDW performs better than LN-T but not as well as RF-T in all testing sets and the LN-T has almost the same prediction as to the BL. In addition, as the availability ratio increased, the error decreases, showing that the data has more power than the algorithms.
Con l'espansione dei servizi basati sulla posizione nell'ultimo decennio, la localizzazione indoor in diverse reti ha guadagnato molta attenzione da parte degli studi e rende fondamentale avere la migliore previsione della posizione per gli utenti in base alla potenza del segnale ricevuto (RSS). Negli ultimi anni sono state stabilite diverse tecniche di localizzazione che sfruttano il segnale trasmesso dai dispositivi per localizzare le persone. In questa tesi, ci proponiamo di mostrare prima la possibilità di trasferire l'apprendimento nella localizzazione indoor da 2.4 a 5.2, e in secondo luogo utilizzare diversi algoritmi per fare la previsione della posizione con RSS e confrontare le loro prestazioni. Per avviare l'analisi, vista la generalizzabilità dei dati, abbiamo utilizzato il dataset raccolto dal gruppo di ricercatori dell'Università di Victoria, contenente cinque AP che inviano segnali su 2.4 e 5.2GHz. La localizzazione delle impronte digitali contiene due fasi; offline (test) e online (formazione), che funzionano con metodi inter o intra- tecnologici. In questa tesi esaminiamo tre diversi metodi; LN-T, RF-T e R-IDW e confrontarli con approcci tradizionali basati sull'interpolazione di base, utilizzando metodi alternativi come l'ottimizzazione iper-parametrica e sperimentando diverse porzioni dei set di dati per l'addestramento e il test basati su k-fold in vari rapporti di disponibilità del 25%, 50%, 75% e 100% per simulare lo scenario. I risultati dimostrano che è possibile trasferire l'apprendimento da 2,4 a 5,2 GHz e che la tecnologia inter funziona meglio della tecnologia intra. Diversi metodi per i rapporti di disponibilità possono avere risultati diversi in base alla distribuzione delle posizioni per la fase di formazione, si può lavorare meglio per alcuni e peggio per altri. Per essere precisi, RF-T ha generalmente le migliori prestazioni con un massimo di 0.8767m come errore di localizzazione rispetto al BL. R-IDW si comporta meglio di LN-T ma non così come RF-T in tutti i set di test e LN-T ha quasi la stessa previsione del BL. Inoltre, all'aumentare del rapporto di disponibilità, l'errore diminuisce, dimostrando che i dati hanno più potenza degli algoritmi.
Machine learning based indoor localization via inter-frequency knowledge transfer
JAVADZAD, AYSA
2020/2021
Abstract
With the expansion of location-based services over the last decade, indoor localization in different networks has gained a lot of study attention and makes it crucial to have the best position prediction for users based on the received signal strength (RSS). Several localization techniques have been established in recent years that leverage the devices' transmitted signal to locate people. In this thesis, we aim to first show the possibility of transfer learning in indoor localization from 2.4 to 5.2, and secondly utilize different algorithms to do the position prediction with RSS and compare their performance. To start the analysis, due to data generalizability, we used the dataset gathered by the group of researchers at the University of Victoria, containing five AP sending signals on 2.4 and 5.2GHz. The fingerprint localization contains two phases; offline (testing) and online (training), which work with inter or intra- technology methods. We look at three different methods in this thesis; LN-T, RF- T, and R-IDW, and compare them to traditional approaches based on basic interpolation, using alternative methods such as hyper-parameter optimization and experimenting with different portions of the datasets for training and testing based on k-fold in varied availability ratios of 25%, 50%, 75%, and 100% to simulate the scenario. The results demonstrate that transfer learning from 2.4 to 5.2GHz is possible and inter technology works better than intra technology. Different methods for availability ratios can have various results based on the distribution of the positions for the training phase, one can work better for some and the worst for others. RF-T generally has the best performance with a maximum of 0.8767m as the localization error compared to the BL. R-IDW performs better than LN-T but not as well as RF-T in all testing sets and the LN-T has almost the same prediction as to the BL. In addition, as the availability ratio increased, the error decreases, showing that the data has more power than the algorithms.File | Dimensione | Formato | |
---|---|---|---|
Aysa_Javadzad.pdf
Open Access dal 06/07/2022
Descrizione: Machine Learning based Indoor Localization via Inter-Frequency Knowledge Transfer
Dimensione
7.53 MB
Formato
Adobe PDF
|
7.53 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/178113