Semi-supervised Learning of Non-stationary Acoustic Signals Using Time-Frequency Energy Maps

Non-stationary signals are time-varying signals that represent various real-world phenomena, such as biomedical signals, vibrating machinery, and acoustic signals, among others. Their accurate classifying is crucial to enhance diagnostic capabilities and predict critical events. A widely used method of characterization is based on time-frequency mapping techniques, which capture both temporal and frequency information. However, this approach introduces redundant information and difficult the extraction of relevant features for accurate classification. This paper presents a robust classification method for non-stationary signals using time-frequency maps focusing on relevant information while minimizing redundancy. The proposed methodology extracts non-stationary features from STFT maps using a background subtraction technique based on singular value decomposition (SVD). Then, principal component analysis (PCA) is implemented for dimensionality reduction, clustering, and classification using vectorized foreground maps. The method was tested using non-stationary ultrasonic signals. The results demonstrate the effectiveness of the proposed method in extracting features from the STFT maps, resulting in a significant separation between clusters of classes. The experimental results highlight the robustness and effectiveness of the method.