Abnormal Audio Detection Based On Deep Learning

With the development of artificial intelligence technology,the application scenarios of deep neural networks are becoming more and more extensive,and the introduction of deep learning methods in audio data processing has become one of the research hotspots.Due to the monitoring of dead angles and object occlusion and bad weather in traditional video detection methods,video surveillance of abnormal events is prone to problems such as missed detection and false detection.In this thesis,the method of abnormal audio feature extraction is analyzed and the deep feedforward neural network is introduced into the abnormal audio detection system.The main work of this thesis is as follows:(1)The relevant thesis and literatures related to audio data processing and deep neural networks are reviewed,the research status and future development trends of related fields are analyzed and the research plan and content of this paper are clarified.(2)The audio data preprocessing method is studied.The audio data is collected from the audio acquisition hardware,and the analog signal is converted into a digital signal,and the digital signal is pre-emphasized by the audio data,and the audio noise reduction algorithms such as wavelet denoising method,homomorphic filtering method and Wiener filtering method are compared and analyzed.After the audio data is subjected to noise reduction processing,endpoint detection is performed on the continuous noise reduction frequency data,and the effective audio segment is separated from the continuous audio signal.(3)The audio data feature extraction method is studied.In the process of audio feature extraction,the pre-processed effective audio segment is first framed and windowed to have stable short-term features,and then the time domain characteristics of the audio data and the frequency of the audio data are compared and analyzed.The advantages and disadvantages of domain features and methods of feature extraction using artificial neural networks are compared and analyzed.After comparative analysis,this thesis uses the frequency domain features of audio data to extract the features of the audio signal.After the audio signal data is framed and windowed,the corresponding Mel frequency Cepstral coefficients are calculated for each audio signal frame,which has the advantages of low computational complexity and good audio data feature extraction,and classifies the extracted low latitude features to obtain the final classification result.(4)The method of classification of abnormal audio data is studied.The deep feed forward neural network is used as the classifier in this thesis.The basic unit and architecture of the deep feedforward neural network are analyzed in detail.The advantages and disadvantages between different activation functions,loss functions and optimization algorithms are compared and analyzed.Two deep feedforward neural network model architectures are designed in this thesis,,which are applied to high-performance servers with strong computational power and low-power embedded devices.Compared with the traditional algorithm,the results show that the proposed algorithm has a good classification effect.