Research On STFT-LLE Manifold Learning Method And Its Application In Motion Acoustic Feature Extraction

With the acceleration of urbanization,urban traffic noise pollution has become more and more prominent,and it has become a social problem that needs to be solved urgently.Traffic noise is characterized by time-varying,superimposing and space-time coupling of sound signals,and the sound data is characterized by high dimensionality and nonlinearity,which makes it difficult to extract key acoustic features.The method of sound feature extraction has high complexity,large numerical calculation and poor validity.Therefore,how to effectively extract acoustic features and reduce the complexity of the extraction method has become an important scientific problem for the accurate identification of multi-source acoustic sources.In this paper,the STFT-LLE manifold learning method is proposed.It combined with short-time Fourier transform(STFT)and local linear embedding algorithm(LLE).This method is applied to the feature extraction of motion acousticfield.The key algorithm and the core process are verified by numerical simulation and experimental test.The thesis is selected from the National Natural Science Foundation of China(61671262,61871447).The main research works of this paper are as follows:(1)Based on the in-depth study of manifold learning algorithm and subsonic motion acoustic features,the local linear embedding algorithm(LLE)is proposed as the core manifold learning algorithm for motion acoustic feature extraction,and the key factors affecting the manifold structure characteristics and the construction method of high-dimensional feature matrix are discussed in detail.(2)The short-time Fourier transform(STFT)is used as the first step of acoustic feature extraction.The extraction result is used as the input vector of LLE algorithm to construct the high-dimensional feature matrix of LLE algorithm,which solves the problem of high-dimensional feature matrix construction in this algorithm.The STFT-LLE manifold learning method is further integrated by STFT and LLE algorithm,and the acoustic feature manifold learning is carried out.The specific implementation flow of the method is given.(3)The acoustic feature extraction of STFT-LLE manifold learning method is studied by numerical simulation.Aiming at the subsonic linear uniform motion acoustic signals with multiple Mach numbers less than 1,this method is applied to extract the moving sound features to clarify the feasibility of the key algorithm of STFT-LLE method in motion sound feature extraction.(4)The experimental analysis and theoretical verification of the STFT-LLE manifold learning method and its motion sound feature extraction effectiveness are carried out by experimental tests.For the collected car(Kai Yue 1.6LX-MT)and motorcycle(Medical XB125T-11F)acoustic signals,the acoustic feature extraction is carried out by four joint feature extraction methods(PSR-PCA,PSR-LLE,STFT-PCA,STFT-LLE)and each feature vector is sent to Support Vector Machine(SVM)classifier.The vehicle identification experiment verifies the effectiveness of the STFT-LLE manifold learning method.