Research On Sound Signal Feature Extraction And Recognition Of Complex System

With the rapid development of science and technology and the diversification of people’s needs,more and more complex systems are encountered and applied in people’s production and life.The internal structure and operation mechanism of complex system are complicated,it is difficult for people to realize its running state law from the internal structure,the sound signal sent by the system reflects the real-time internal running state of the system.Therefore,people can take advantage of this property,through the recognition of the sound signal generated by the complex system,the internal state of the system can be monitored in real-time,improved people to better understand the internal cognition of the system,it also can guide people to use complex system better to achieve efficient production and life targets.The main purpose of this thesis is to achieve high recognition accuracy of sound signals generated by complex systems.Therefore,the main research work of this thesis includes the following parts.(1)This thesis discusses the basic research of sound signal feature extraction and recognition.The research work is divided into two parts:feature extraction and recognition of sound signal.The time-frequency domain method of sound signal feature extraction,traditional machine learning method and neural network in sound signal recognition are studied respectively.The characteristics and main influencing factors of sound signal are analyzed,and the general steps of sound signal feature extraction and recognition are summarized.(2)Aiming at the noise problem existing in the sound signal data set,and in order to extract richer and more regular features containing the original sound signal,better training data set is provided for the subsequent sound signal recognition.In this thesis,Variational Mode Decomposition algorithm has been adopted to decompose and de-noising sound signals and extract features.Because VMD algorithm requires artificial preset K value,therefore in this thesis,the Improved Variational Mode Decomposition has been proposed.This improved algorithm uses the center frequency ratio to find the optimal K value to achieve the best result of VMD decomposition.At the same time,the T values of the decomposed eigenmode components are calculated by combining the permutation entropy and correlation coefficient,and the noise and characteristic components are determined according to the T values.The experimental simulation shows that IVMD can decompose the main frequency components of the signal,and has strong denoising ability,which can provide better sound signal characteristics for the sound signal recognition model.(3)In order to realize the accurate recognition of sound signal characteristics.According to the feature of sound signal is a sequential signal,this thesis selects the Long Short Term Memory neural network in the recurrent neural network.LSTM can predict and recognize one-dimensional time sequence signals,and solve the problem of gradient disappearing or explosion to a certain extent.In this thesis,a sound signal feature extraction and recognition model of IVMD-LSTM network was established based on IVMD and LSTM.The model was initially set with two layers of LSTM,but the model could not reach the fitting state.With the number of layers increased to five,the recognition accuracy of the model reached 84.34%,but the accuracy curve of the training process was not smooth.There are still large fluctuations,and the recognition effect of the model is not stable.(4)For the problem that the sound signal recognition effect of the model established by LSTM network is not ideal.The Residual Neural Network in the Convolutional Neural Network is selected in this thesis to recognize and train the features of sound signals.ResNet network not only has the advantage of traditional convolutional neural networks,therefore,it can train more data with fewer parameters,but also can overcome the problem of gradient disappearance or explosion caused by the increase of layers due to the existence of residual units.Since ResNet network is generally trained to recognize two-dimensional Image data,the Sound signal is converted into Sound Image in this thesis.At the same time,this thesis compares the performance of single feature and multi-feature,IVMD and traditional EMD algorithm,and different layers of ResNet network in this model.Finally,IVMD-SI-ResNet50 model has the best effect,which can achieve 99.47% recognition accuracy under fewer network parameters,and the network fitting speed is fast.With strong generalization ability,it achieves the research goal of high recognition accuracy of sound signal.