Abstarct: Abstract Cardiovascular diseases are among the leading causes of mortality and disability worldwide. Rapid and accurate diagnosis of these diseases plays a key role in reducing treatment costs and improving patients' quality of life. However, conventional methods baxsed on heart signals still face challenges such as the need for precise segmentation of signal components, computational complexity, and limitations in extracting effective features. In this thesis, a novel frxamework for classification of phonocardiogram (PCG) signals is presented, which, for the first time, employs Deep Orthogonal Non-negative Matrix Factorization (Deep ONMF) combined with feature engineering and fusion protocols in the domain of heart signals. The innovation of this study lies in the purposeful use of Deep ONMF for extracting high-level features from PCG signals and designing four protocols to fuse the output of laxyers into latent feature maps. The proposed structure consists of three Deep ONMF laxyers, each extracting high-level time-frequency representations. Subsequently, classification is performed using various feature selection algorithms and classifiers. Evaluation results on two reputable datasets, PhysioNet and Yaseen, demonstrate that protocols two and three, baxsed on temporal feature fusion, yield superior class separation. Furthermore, the Random Forest (RF) classifier outperforms others across all protocols, achieving up to 100% classification accuracy on both datasets. These findings indicate the high efficiency of the proposed frxamework in extracting rich features, eliminating the need for signal segmentation, reducing computational complexity, and enhancing class separability, making it a powerful tool for developing intelligent assistive diagnostic systems in cardiovascular disease screening.