Emergency environmental sound classification using TrianglePat: a self-organized feature engineering framework

Abstract

Automated environmental sound classification (ESC) has become a key research area in advanced signal processing. This study focuses on detecting emergency events using a novel ESC model. A new ESC emergency dataset was developed, containing 4,841 audio samples categorized into four classes: (i) crying/complaining, (ii) low intensity/normal, (iii) high intensity, and (iv) violence. To automate ESC, we introduced a feature extraction function based on basic statistical moments: maximum, minimum, and median. A triangular representation was created from these moments, and the edge lengths were computed to generate a feature map. Histograms of the feature map were used as feature vectors. To evaluate the classification potential of the proposed feature extraction method, Triangle Pattern (TrianglePat), we developed a feature engineering framework with three key steps: (i) tunable q-factor wavelet transform (TQWT), (ii) TrianglePat-driven multilevel feature extraction, and (iii) feature selection via Cumulative Weighted Iterative Neighborhood Component Analysis (CWINCA). The classification step used a t-algorithm-based support vector machine (tSVM), making the model self-organized, thus forming a Self-organized Feature Engineering (SOFE) model. The TrianglePat-based SOFE model was validated using four cases within the dataset, consistently achieving over 92% classification accuracy. These findings confirm that the TrianglePat-driven SOFE model is a highly effective feature engineering method. The results also demonstrate that the proposed model can accurately detect emergency-related events through environmental sound analysis, making it a valuable tool for real-world applications in emergency services.