Research On Optical Fiber Distributed Seismic Wave Signal Denoising Algorithm

Distributed Acoustic Sensing(DAS)technology based on Phase-sensitive Optical Time Domain Reflectometer(?-OTDR)has attracted much attention due to its advantages such as small channel spacing,easy long-term large-scale deployment,full well high-density seismic data acquisition and high sampling resolution.The actual production field tests have been carried out in many fields such as vertical seismic profile(VSP)monitoring and imaging,surface seismic data acquisition,and 4D monitoring of deepwater oilfields,and satisfactory results have been achieved.When using DAS system to collect VSP signals,although the influence of surface noise can be reduced,such as downhole spring wave noise,casing wave noise,wellbore wave noise can still deteriorate the signal quality,affecting subsequent analysis such as horizon tracking and fault recognition.In this thesis,based on the feature of effective wave and noise in the actual acquired DAS VSP signal,three denoising methods for the actual acquired DAS VSP signal are proposed.The work of this thesis is summarized as follows:(1)Taking encoder-decoder as the framework,the DAS VSP signal noise reduction network based on cross-shaped multi-head self-attention is constructed.The actual acquired DAS VSP signal is converted into token by embedding vector generation method of Token token as the input of network encoder.The jump connector encoder is introduced to make full use of local details to enhance the noise reduction ability of DAS VSP signal.The cross-shaped multi-head self-attention is used to calculate the attention weight distribution of the network.The decoder with the same internal structure as the backbone sub-encoder is used to fit the data distribution of DAS VSP clean signal layer by layer.A convolution module is used to output the denoised DAS VSP signal.(2)Taking the generative confrontation network as the framework,the DAS VSP signal noise reduction network based on self-attention and generative confrontation is constructed.The cross shape multi-head self-attention is used to construct the feature extractor.The input structure of the feature extractor is consistent with the work(1).The automatic encoder is used as the generator of the network,and the discriminant is used to judge the DAS VSP noise reduction signal output by the generator.Feature extractor,generator,discriminator update parameters alternately,and finally complete the training.(3)In order to optimize and upgrade the performance of the feature extractor,the network in work(2)is modified,and the structure of the generator and the discriminator is retained.The residual block and LSTM network are used to form a new feature extractor,and the DAS VSP signal denoising network based on residual LSTM and generative adversarial is constructed.In this network,the actual acquired DAS VSP signal and noise residual signal are input into the network in the form of original twodimensional matrix,and the feature extractor composed of residual block and LSTM is used to extract feature and generate dependence between long distance data.In this thesis,the research on the noise reduction of DAS VSP signals collected in practice effectively boosts the application of DAS system in oil and gas exploration and development,which has important engineering practical significance.