Improved Local Mean Decomposition Based on the T-Distribution for Feature Extraction of Abnormal Sounds in Public Places

Weihong Li,Bingxin Zhao,Shuyong Peng,Weiguo Gong

DOI: https://doi.org/10.17743/jaes.2017.0031

IF: 1.155

2017-01-01

Journal of the Audio Engineering Society

Abstract:Background noise in public places can affect feature extraction for abnormal sounds when Local Mean Decomposition (LMD) is used as a signal-processing tool. In this paper we first confirm that the background noise in public places obeys a T-distribution using Kolmogorov-Smirnov hypothesis testing. Second, we propose an improved LMD method based on the T-distribution for effectively extracting the features of abnormal sounds in public places. We add particular production function components of inhomogeneous random noise obeying a T-distribution to the abnormal sound in a nested manner and then take the ensemble means of the obtained production functions as the decomposition results to alleviate the mode mixing inherent in LMD. Additionally, we replace the moving average with a linear spline to reduce the iteration required in LMD from triple-loop iteration to double-loop iteration, and we set a maximum sifting time restriction to decrease the time consumption. Finally, we use the proposed method as a tool for decomposing an abnormal sound and use the energy ratio of each production function with respect to the original signal as the features for classification. Experimental results demonstrate that the proposed method outperforms commonly used methods in terms of the classification rate and computational cost.