A Wavelet Scattering Feature Extraction Approach for Deep Neural Network Based Indoor Fingerprinting Localization

Area of Science:

• Signal Processing
• Machine Learning
• Indoor Localization

Background:

• Artificial Neural Network (ANN) performance in indoor fingerprinting is hindered by unstable Received Signal Strength Indicator (RSSI) variations.
• Existing feature extraction methods inadequately address RSSI fluctuations, degrading ANN-based indoor localization accuracy.

Purpose of the Study:

• To develop a robust feature extraction technique for ANN-based indoor localization that mitigates RSSI variation.
• To improve the stability and performance of deep neural network (DNN) models for indoor positioning.

Main Methods:

• Employed wavelet scattering transform for feature extraction, ensuring stability against minor deformations and rotation invariance.
• Utilized zero-th and first layer decomposition coefficients, concatenating scattering path coefficients as input features for a DNN model.
• Validated the proposed algorithm using real-world indoor measurement data.

Main Results:

• The proposed feature extraction method demonstrated stability in the face of RSSI variations.
• The deep neural network model, using the extracted features, achieved good performance in indoor localization.
• Experimental results confirm the effectiveness of the wavelet scattering transform for robust feature extraction in this context.

Conclusions:

• Wavelet scattering transform offers a reliable feature extraction solution for ANN-based indoor localization systems.
• The developed method significantly enhances the resilience of indoor localization algorithms to signal instability.
• This approach paves the way for more accurate and dependable indoor positioning solutions.