SASMOTE: A Self-Attention Oversampling Method for Imbalanced CSI Fingerprints in Indoor Positioning Systems

Area of Science:

• Computer Science
• Electrical Engineering
• Signal Processing

Background:

• WiFi fingerprinting is a key indoor positioning method.
• Building fingerprint databases is crucial but faces challenges like data imbalance.
• Data imbalance hinders the accuracy of WiFi-based localization systems.

Purpose of the Study:

• To propose a novel deep learning-based oversampling method, SASMOTE, to address data imbalance in CSI fingerprint databases.
• To enhance the accuracy of indoor positioning systems by improving the quality of training data.
• To validate the effectiveness of SASMOTE through comprehensive experiments.

Main Methods:

• Developed a self-attention encoder-decoder for feature extraction and dimensionality reduction.
• Applied Synthetic Minority Oversampling Technique (SMOTE) to balance the CSI fingerprint dataset.
• Constructed a CSI fingerprinting dataset for model training and evaluation.
• Utilized an improved 1D-MobileNet model for localization on the balanced dataset.

Main Results:

• SASMOTE effectively resolves the data imbalance issue in CSI fingerprint databases.
• The proposed method significantly improves localization accuracy compared to traditional approaches.
• Experiments demonstrate the robustness and performance of SASMOTE across different datasets.

Conclusions:

• SASMOTE is a viable deep learning solution for data imbalance in WiFi fingerprinting.
• Balancing the fingerprint database is critical for achieving high indoor localization accuracy.
• The integration of SASMOTE with advanced localization models like 1D-MobileNet shows promising results.