Autoencoder-based Feature Learning for Wi-Fi Fingerpring Indoor Localization
Mensio, Markus (2026)
2026
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2026-05-29
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202605286465
https://urn.fi/URN:NBN:fi:tuni-202605286465
Tiivistelmä
Indoor positioning has become increasingly important due to the growing demand for location-aware services in environments where Global Navigation Satellite System (GNSS) signals are unreliable. Among indoor localization approaches, Wi-Fi fingerprinting is widely used because it utilizes existing wireless infrastructure without requiring additional hardware. However, Received Signal Strength Indicator (RSSI) measurements are often noisy, unstable, and highdimensional, which can reduce localization accuracy.
This thesis investigates whether autoencoder-based feature learning can improve Wi-Fi fingerprint indoor localization performance. The study evaluates the effect of learned feature representations on localization accuracy using two machine learning methods: k-nearest neighbors (kNN) and Random Forest (RF). Experiments were conducted using multiple publicly available datasets representing different indoor environments and temporal conditions.
The methodology consists of preprocessing RSSI measurements, learning compact latent feature representations using an autoencoder, and applying localization algorithms both to raw RSSI data and to learned features. Performance was evaluated using mean localization error.
The results show that the impact of autoencoder-based feature learning is strongly datasetdependent. In some datasets, learned feature representations improved localization accuracy, particularly for kNN-based localization. However, in other datasets, feature compression degraded performance, especially under temporal distribution shifts. Random Forest generally provided the most stable overall performance across datasets.
The findings indicate that autoencoder-based feature learning can improve RSSI representation quality and localization accuracy in certain scenarios, but its effectiveness depends on the stability and characteristics of the underlying signal environment.
This thesis investigates whether autoencoder-based feature learning can improve Wi-Fi fingerprint indoor localization performance. The study evaluates the effect of learned feature representations on localization accuracy using two machine learning methods: k-nearest neighbors (kNN) and Random Forest (RF). Experiments were conducted using multiple publicly available datasets representing different indoor environments and temporal conditions.
The methodology consists of preprocessing RSSI measurements, learning compact latent feature representations using an autoencoder, and applying localization algorithms both to raw RSSI data and to learned features. Performance was evaluated using mean localization error.
The results show that the impact of autoencoder-based feature learning is strongly datasetdependent. In some datasets, learned feature representations improved localization accuracy, particularly for kNN-based localization. However, in other datasets, feature compression degraded performance, especially under temporal distribution shifts. Random Forest generally provided the most stable overall performance across datasets.
The findings indicate that autoencoder-based feature learning can improve RSSI representation quality and localization accuracy in certain scenarios, but its effectiveness depends on the stability and characteristics of the underlying signal environment.