Material Classification from Non-Line-of-Sight Acoustic Echoes Using Wavelet-Acoustic Hybrid Feature Fusion.

Area of Science:

• Acoustics
• Signal Processing
• Machine Learning

Background:

• Non-line-of-sight (NLOS) acoustic material classification is challenging due to signal degradation.
• Existing methods struggle with complex acoustic environments and obstructed sound paths.

Purpose of the Study:

• To improve NLOS material recognition using a hybrid feature fusion and deep recurrent neural networks.
• To develop a robust and interpretable framework for real-time NLOS acoustic sensing.

Main Methods:

• Collected echo signals from nine materials using the ANLOS-R dataset simulating NLOS environments.
• Extracted time-domain acoustic features and multi-scale wavelet energy/entropy statistics.
• Trained deep learning models including LSTM, BiLSTM, GRU, and CNN-LSTM on a 70-dimensional hybrid feature set.

Main Results:

• The Convolutional Neural Network-LSTM (CNN-LSTM) architecture achieved the highest balanced accuracy and macro-F1 score of 0.99.
• SHapley Additive exPlanations (SHAP) analysis revealed complementary roles of Mel-Frequency Cepstral Coefficients (MFCCs) and wavelet features.
• The model demonstrated strong generalization and convergence performance.

Conclusions:

• The proposed wavelet-acoustic hybrid feature fusion with CNN-LSTM effectively addresses NLOS material classification challenges.
• The approach offers a robust, interpretable, and data-driven solution for real-time acoustic sensing in obstructed environments.