Feature Extraction of Mixed Sound Events Based on Improved VMD and Multiscale Permutation Entropy

The sound event recognition technology has been applied in many important fields, and the progress of feature extraction of sound event can improve the performance of the sound recognition system under the noise background. The variational modal decomposition（VMD） algorithm, which was first used in fault diagnosis field, was applied to the feature extraction of mixed sound events. Particle swarm optimization (PSO) was used to improve the VMD algorithm, and the empirical mode decomposition（EMD） algorithm was compared. Firstly, the signals with simple mixed sound events are decomposed by VMD algorithm and EMD algorithm to obtain multiple intrinsic modal components. Then its correlation coefficient was calculated, and each component was synthesized and spliced according to the principle of maximum correlation to reconstruct the signal, and the type of component was determined. The multiscale permutation entropy (MPE) was used to calculate the MPE values of each component, and the starter signal was extracted successfully. By comparison, the result of VMD algorithm is better than that of EMD algorithm in the processing of simple mixed sound signals. Then the PSO-VMD algorithm, VMD algorithm and EMD algorithm are applied to the more complex signal decomposition process, and combined with MPE to complete the feature extraction of sound signals. Finally, by comparing the MPE distribution map, it can be seen that VMD algorithm and PSO-VMD algorithm are better than EMD algorithm, more accurate in the decomposition of signals, and easier to distinguish features extracted by combining with MPE.