Heterogeneous Transfer Learning for Wi-Fi Indoor Positioning Based Hybrid Feature Selection

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• Indoor positioning systems (IPS) face challenges with accuracy and calibration due to noisy data from multiple Wi-Fi access points.
• Heterogeneous transfer learning (HetTL) offers a promising approach to adapt models across different domains, but requires effective feature selection for Wi-Fi fingerprinting.

Purpose of the Study:

• To develop and evaluate novel algorithms for feature selection in fingerprint-based indoor positioning problems (IPP) using HetTL.
• To reduce training calibration effort and noise from duplicate Wi-Fi fingerprints.
• To enhance positioning performance in the target domain.

Main Methods:

• Application of heterogeneous transfer learning (HetTL) methods.
• Development of two novel feature selection algorithms: a Principal Component Analysis-based (PCA-based) approach and a hybrid approach combining PCA with correlation effect analysis.
• Construction of a new feature vector retaining significant predictors and determination of efficient feature dimensions.
• Cramer-Rao Lower Bound (CRLB) analysis to estimate the lower limit of positioning error variance.

Main Results:

• The hybrid-based feature selection algorithm achieved the minimum mean absolute error in indoor positioning.
• CRLB analysis indicated that the number of Wi-Fi access points impacts location estimation error bounds.
• Identifying significant predictors effectively improves positioning performance.

Conclusions:

• The proposed hybrid feature selection approach within HetTL significantly enhances accuracy for Wi-Fi fingerprint-based indoor positioning.
• Optimizing feature selection is crucial for mitigating noise and improving the reliability of IPS.
• CRLB analysis provides valuable insights into the fundamental limits of positioning accuracy based on available Wi-Fi infrastructure.