Combination Of Singular Value Decomposition And Local Linear Embedding Manifold Learning Method Application

With the rapid development of social economy,people are demanding higher and higher environmental comfort.In large industrial and mining enterprises,the noise of the environment seriously affects the physical and mental health of workers,but also seriously affects the life and efficiency of equipment.Effective control of noisy environment sound has become an urgent problem to be solved.Noisy environment sound is produced by the joint action of multiple sound sources,and the sound characteristics of each sound source are different.Feature extraction and identification of noisy environment sound are the premise and basis of effective control.Manifold learning,as a machine learning method,is a front-end technology of intelligent data processing,which can quickly and efficiently mine data feature information.In this paper,a manifold learning method based on singular value decomposition and local linear embedding is proposed,which can efficiently extract useful acoustic information,separate background noise,improve the speed and accuracy of feature extraction in noisy environment,using machine learning for noise recognition and noise feature extraction.The source feature is identified,and the method is verified by the feature recognizer.The method is further applied to extract the noise characteristics of the oxygen plant compressor.This paper is supported by the National Natural Science Foundation of China(61671262,61871447).The main research contents are as follows:(1)In this paper,the principle of manifold learning method,manifold formation mechanism and the physical meaning of manifold are discussed,the key factors and core parameters that affect manifold formation are clarified,the common algorithm of manifold learning is analyzed,and the manifold learning process is formed.(2)The process of singular value decomposition(SVD),the basis of singular value selection and the key factors affecting the size of singular value are studied in detail.SVD is applied to construct a high-dimensional data set based on short-time Fourier transform,which improves the validity of the data set by removing redundant information and noise components.(3)The local linear embedding algorithm is studied,and the key factors affecting the accuracy of the algorithm,which are the value of the nearest neighbor point and its determination,are found.The optimal selection of the nearest neighbor point is carried out by constructing the correlation function,and the principle of adaptive nearest neighbor selection is given.While ensuring the accuracy of manifold structure,the automatic optimal selection of the nearest neighbor point value is realized.(4)A manifold learning method based on SVD-ALLE is proposed.Firstly,SVD denoising is applied to the high-dimensional data set based on short-time Fourier transform.Secondly,adaptive LLE algorithm is applied to extract the sound source features in the noisy environment to jointly realize the high-dimensional data processing and feature extraction.(5)The support vector machine(SVM),which is a feature recognizer,verifies the theoretical validity of the proposed method,and applies it to multi-source coupling,strong background noise and non-linear noise feature extraction of compressor in oxygen plant.The analysis results further show the reliability of the method.