OHetTLAL: An Online Transfer Learning Method for Fingerprint-Based Indoor Positioning.

Area of Science:

• Computer Science
• Signal Processing
• Machine Learning

Background:

• Indoor positioning systems (IPS) often rely on transfer learning (TL) but struggle with dynamic environments and varying signal distributions.
• Traditional TL methods face challenges like costly data collection, online data arrival, differing feature spaces, and negative knowledge transfer.

Purpose of the Study:

• To propose an online heterogeneous transfer learning algorithm (OHetTLAL) for RSS fingerprinting in IPS.
• To enhance positioning performance by effectively fusing knowledge from source and target domains while mitigating negative transfer.

Main Methods:

• Developed an OHetTLAL algorithm for IPS-based RSS fingerprinting.
• Refined the source domain based on the target domain to prevent negative knowledge transfer.
• Utilized the co-occurrence measure of feature spaces (Cmip) to create homogeneous feature spaces and selected high-weight features for classifier training.

Main Results:

• The proposed algorithm demonstrated superior performance compared to state-of-the-art methods in real-world indoor positioning scenarios.
• The OHetTLAL algorithm proved robust to changing and fluctuating indoor environments.
• The method effectively mitigated the model's overfitting problem, enhancing generalization.

Conclusions:

• OHetTLAL offers a robust and accurate solution for RSS fingerprint-based indoor positioning, especially in dynamic environments.
• The algorithm successfully addresses limitations of traditional TL by adapting to new data and avoiding negative knowledge transfer.