Key Technology Research On Speech Signal Enhancement Based On Distributed Optical Fiber Acoustic Sensing System

The Distributed fiber-optic Acoustic Sensing(DAS)system is a novel acoustic detection technology that has attracted much attention in practical applications due to its high spatiotemporal resolution,long-distance continuous monitoring,and no need for external interference sources.However,the presence of random noise in the DAS system can seriously affect the quality of weak speech signals,limiting the ability to extract information from speech signals.To address this issue,this thesis proposes a neural network algorithm based on complex spectrum mapping from theoretical research to improve the quality and information recognition ability of DAS speech signals.Experimental verification shows that the algorithm can effectively suppress random noise,improve the quality of DAS speech signals,and achieve high-fidelity and high-quality speech signal enhancement under weak signals.The research results provide strong support for the efficient recognition of speech information in noisy environments for DAS systems,and take an important step towards the practical application of high-performance DAS systems.The main research contents of this thesis are as follows:(1)Starting from the Rayleigh scattering of optical fibers,this thesis elucidates the specific process of DAS technology detecting acoustic signals through the Rayleigh scattering of optical fibers and using demodulation algorithms to demodulate and restore them into speech signals.Through the derivation of the theoretical basis of distributed fiber-optic acoustic sensing technology,a deeper understanding of the principles and implementation methods of this technology is achieved.(2)A four-channel acquisition method based on a 3×3 coupler DAS system was designed,and the DAS speech signal dataset required for this thesis was constructed.The self-alignment of noisy speech signals and clean speech signals was achieved to avoid the difficulty of manual alignment.In addition,through the feature analysis of the timefrequency domain,the specific requirements for speech signal enhancement in distributed fiber-optic acoustic sensing systems were introduced.(3)Classic and neural network speech signal enhancement algorithms were used to enhance DAS speech signals in experiments.The results showed that the simple scheme of enhancing speech signals based on noise estimation from non-speech segments in the classic algorithm did not produce good results.In order to better extract speech information,this thesis introduced smooth dilated convolution and attention mechanism based on convolutional recurrent networks to construct a convolutional recurrent network algorithm based on complex spectrum mapping to improve the algorithm’s enhancement performance.Experimental results showed that the proposed method can successfully suppress random noise in the DAS system,reduce the noise intensity of the speech signal by approximately 19.25 d B,and improve the average scale-invariant signal-to-noise ratio by approximately 50.26 d B.Compared with other methods,using the proposed method in this thesis can obtain higher quality speech signals.The research results are expected to promote the integration and development of sensing technology and communication technology,improve the speech communication quality based on DAS technology,and provide technical support for communication in various complex environments.