Research And Application Of Audio Event Classification Based On Deep Network

As a common information carrier,audio signal is an important means for people to perceive the environment and exchange information.Compared with other signals,it has the characteristics of easy collection,easy transmission,and little environmental interference.Related research has also been concerned by scholars favored.Audio event classification is one of the important research fields.It uses machine equipment to analyze the collected audio signals,intelligently perceives changes in the current environment,and determines the corresponding events to assist users in decision-making.The role of this technology is also used in security monitoring,smart city traffic,autonomous driving and other fields.In recent years,deep learning algorithms have developed rapidly,and various deep network frameworks have emerged in an endless stream.They have shown great advantages in solving multi-input strong coupling scenarios.This also provides an idea for the breakthrough of audio event classification system performance.Therefore,this paper mainly explores the classification tasks of audio events around deep neural network frameworks,improves system performance by building network models of different architectures,and applies algorithms to tool wear monitoring problems.The main research contents of this article are as follows:First,the audio event classification method based on the classic deep network is explored.The structure of convolutional network and recurrent network and the corresponding training algorithms are introduced.Based on the characteristics of audio signals,the spectral characteristics are used as the network input,and the audio event classification system with their deep network as the recognition algorithm is built,several groups of comparative experiments were carried out on the ESC-10 universal data set.Experimental results show that the classification performance of deep networks is generally better than that of shallow machine learning methods.Among them,the recognition system with the filter bank as the feature input and the main body network as the convolution neural network has the best classification effect.Secondly,a deep network model based on data fusion is proposed and used in audio event classification tasks.In the traditional deep network model,researchers often need to adjust the structure and parameters to avoid the loss of key information.In response to this problem,this paper adds multiple information sources to the network design,do data fusion in dimensions of feature,structure and model,constructs a multi-feature input,multi-scale field of view convolution recurrent deep network architecture,using algorithms Strong decoupling ability,adaptively select parameter operations,and introduce attention mechanism to filter out unnecessary information.The experimental results show that the audio event classification system based on the fusion algorithm shows higher accuracy and F1 score.Finally,from the perspective of application,the effectiveness of the deep fusion network algorithm classification system is verified in the tool wear scenario.The experiment builds an audio database during the machining of tools with different degrees of wear,and uses a deep fusion network for the audio signal classification to achieve a high accuracy.At the same time,using the Raspberry Pi 4B embedded motherboard,the sound sensor and the external USB sound card,a set of system is designed to detect the wear degree of cutting tools on-line.which further illustrates the ability of deep network algorithms to solve engineering problems,and provides a reference for the auto parts processing application of intelligent machining.