| Recently,a large amount of research on acoustic logging tools and corresponding signal processing methods contributes to the significant development of borehole acoustic reflection imaging technology,which has been able to detect the complex geological structures tens of meter away from borehole,such as fractures,caves,and so on.The technology fills the gap between exploring resolution of well logging and seismic exploration.Therefore,it has been widely applied for the exploration and development of unconventional oil and gas and produces remarkable results.In this paper,the field data processing and interpretation of acoustic reflection imaging are systematically studied,composed by four parts: field waveform data preprocessing,migration algorithm,image enhancement and segmentation.The first part is the field waveform data preprocessing.There are two problems of the measured acoustic logging waveform data: one is the corrupted waveform due to the complicated borehole environment,the other is the dispersion of the direct mode wave of the measured waveform.First,an unsupervised reconstruction strategy for the preprocessing of measured acoustic logging waveform based on the deep prior is proposed to restore the corrupted waveform.The structure of convolutional neural network acts as an implicit prior for constraining the ill-posed interpolation problem.Then the ideal waveform data is iterative generated by the Multi Res UNet,which captures the inner data features at different scales only from the raw waveform.The velocity of formation extracted by the traditional methods is lower than the true velocity due to the dispersion of direct mode wave.To solve the problem,we propose a data-driven dispersive processing method.The dispersion curve is first estimated using the above reconstructed waveform data.Then,the slowness probability density is calculated from the dispersion data and is subsequently fitted with a proper analytical function.By finding the edge of the function,the true formation velocity can be determined and applied to build the precise velocity model for the following migration.Second,the prestack gaussian beam migration algorithm is present,which is applied to image the geologic structure outside the borehole using the above precise velocity model.The direct and back propagation wavefield are first approximated by a summation of Gaussian beams.The subsurface image is obtained by calculating the coherence between the direct and back propagetaed wavefield.Finally,a procedure of stacking the images that correspond to different common source waveform is present become it allows the removal of areas of a casual coherence and enhances the signal-to-noise ratio of reflectors.The simulation and field examples show that using the precise velocity field,the diffraction wave are accurately located by the Gaussian beam migration and the geological structures are accurate and clear in the image.However,the random noise presents in the other part of image.The third part corresponds to the image enhancement methods.After the migration,the weak amplitude reflected wave becomes the continuous coherent cuve in the image,while the other kind of the wave,such as directed mode wave and random noise,in measured waveform will present as small-scale and discontinuous background noise,which strongly interfere the distinguishment of wanted reflectors.An anisotropic diffusion filtering method based on structure tensor is first proposed.The structure tensor,allowing both orientation estimation and image structure analysis,is applied to construct the diffusion matrix.Then the anisotropic diffusion equation is solved by the semi-implicit difference scheme due to its stabilization.The method can smooth the small-scaled random noise while preserve the reflectors.The paper also proposes another data-driven discriminative denoising method capturing more semantic information by the convolutional neural network.The core of the network is the multiresolution residual block,composed of several parallel multi-resolution convolution streams for extracting multi-scale features,and the self-attention mechanisms for aggregation of multi-scale features and filter of contextual information.The network learns an enriched set of feature that combines contextual information at multiple scales,while simultaneously preserves the highresolution spatial details.In order to control the degree of denoising of the network,the noise map acts as the input of the network.Meantime,the reversible down-sampling operation is applied to the input image for reducing the memory usage and training time.Using the synthetic image dataset,the trained network is applied to the simulated and field image.The results demonstrate the effective removal of the background noise and enhancement of the reflectors.The last part is the intelligent interpretation method of the image by a feature pyramid network,composed of encoder,decoder and segmentation modules.The Efficientnet adopts the depthwise separable convolutions and channel-wise attention mechanism,acting as the encoder.The network greatly reduces the computation while keep the power for extracting the multiscale features from the input image.The decoder gradually transforms the low-resolution feature from encoder to its original size,at the same time,aggregating the low-level features directly transmitted from the shallow layer of encoder.Finally,the multi-resolution feature processed by the decoder is passed to the segmentation module for the prediction of segmentation map.The CNN could capture abundant low-level features and detect the amplitude anomaly,after trained on the labeled synthetic dataset built by the method based on Born approximation.Then the acquired knowledge is transferred to the field image domain by the transfer learning strategy.The field examples show that the reflectors have been accurately detected under the interference from the random noise and direct wave residual.The position and dip of the reflectors could be estimated based on the above segmented results,which are provided for the reservoir evaluation or other field application.In this paper,a comprehensive field data processing and application methods are present based on the detailed research on the numerical simulation,migration and field application of acoustic reflection imaging.The deep learning algorithm is introduced into the field of acoustic reflection imaging and applied for the waveform reconstruction,image enhancement and segmentation,promoting the application and development of acoustic reflection imaging technology in the oilfield. |
