MFE-UNet-Based Seismic Random Noise Suppression Method

Abstract

The present invention discloses a method for suppressing seismic random noise based on MFE-UNet. The method steps are as follows: constructing an MFE-UNet denoising model; using the seismic noisy data image as the input data of the MFE-UNet denoising model to obtain the seismic denoised data image. The method for suppressing seismic random noise based on MFE-UNet proposed by the present invention can directly train the initial denoising model on the test area and obtain the MFE-UNet denoising model when the training is not fully converged, making full use of the convergence performance of the convolutional neural network, thereby greatly reducing the training cost of the MFE-UNet denoising model and enhancing the denoising effect. At the same time, the denoising principle involved in this method will minimize the loss of effective data in the noisy seismic data.

Description

Technical Field

[0001] The present invention relates to the field of seismic data preprocessing, and more specifically, to a method for suppressing seismic random noise based on a new neural network of unsupervised learning. Background Art

[0002] With the continuous increase in the difficulty of oil and gas exploration and development and the increasing maturity of seismic technology, the main task of seismic exploration has evolved from simple structural exploration in the past to the search for complex and subtle oil and gas reservoirs. However, the seismic data obtained through seismic exploration will inevitably be contaminated by random noise, which in turn affects subsequent seismic data processing and interpretation. High-precision seismic exploration also requires a high signal-to-noise ratio of seismic data. By suppressing the random noise in seismic data, not only the contamination of random noise to seismic data is reduced, but also the signal-to-noise ratio of seismic data is increased, and the signal accuracy of seismic data is improved, which is of great significance for subsequent seismic data processing.

[0003] In the prior art, the seismic data denoising method is based on the UNet model and realizes denoising through paired downsampling and upsampling structures. The ability to remove random noise is limited, and there is still room for improvement in the denoising effect of the model. Moreover, the traditional neural network denoising method has problems such as high training cost, high demand for training set samples, long training time, and not very ideal denoising effect. Therefore, how to solve the problems of limited denoising ability of the seismic data denoising model in the prior art, high training cost, high demand for training set samples, long training time, and not ideal denoising effect of the traditional neural network denoising model has become an important technical problem to be solved by those skilled in the art. Summary of the Invention

[0004] To solve the above technical problems existing in the prior art, the present invention provides a method for suppressing seismic random noise based on MFE-UNet.

[0005] To achieve the above object, the present invention provides the following technical solution: A method for suppressing seismic random noise based on MFE-UNet, and the method steps are as follows:

[0006] Step1. Construct an MFE-UNet denoising model;

[0007] Step2. Use the seismic noisy data image as the input data of the MFE-UNet denoising model to obtain a seismic denoised data image.

[0008] Preferably, in Step1, the method for constructing the MFE-UNet denoising model includes:

[0009] Step101. Processing of seismic noisy data: Select a preset trace gather area in the seismic noisy data as the data area, and form corresponding training sets and test sets through an improved training method;

[0010] Step102. Constructing an initial denoising model: The initial denoising model includes a U-shaped structure, a feature enhancement structure, and a multi-layer trace reduction structure; The U-shaped structure is established based on convolutional block modules, max pooling layers, transposed convolutional layers, and depth connection layers, and the feature enhancement structure is established based on convolutional block modules and transposed convolutional layers; The multi-layer trace reduction structure is established based on convolutional layers.

[0011] Step103. Training the initial denoising model: Use the training set as the input data of the initial denoising model for model training. After the model training is completed, an MFE-UNet denoising model is obtained. The MFE-UNet denoising model performs denoising processing on the seismic noisy data in the test set for model testing.

[0012] Preferably, the improved training method in Step101 specifically includes:

[0013] S1. Constructing a new trace gather: Divide the seismic traces involved in the trace gather area into two sub-trace gathers, an odd trace set and an even trace set, according to the parity of the trace numbers, where the positions of each seismic trace in the sub-trace gathers remain relatively unchanged;

[0014] S2. Generating processing blocks: Randomly select a position within the range of the new trace gather, and generate two corresponding processing blocks in the two sub-trace gathers at this position, which are used as the input and output of the denoising model respectively;

[0015] S3. Constructing a set of processing blocks: Repeat Step S2 to generate multiple groups of processing block combinations. Among them, the processing blocks generated in the odd trace set form a group a of the set of processing blocks, which is used as the input of the denoising model; The processing blocks generated in the even trace set form a group b of the set of processing blocks, which is used as the output of the denoising model; The group a and group b of the set of processing blocks together constitute the training set.

[0016] Preferably, the size of the processing block in Step S2 is 80*80.

[0017] Preferably, in Step101, processing blocks are divided and a test set is generated within the preset trace gather area, and the data area involved in the test set is consistent with the data area involved in the training set.

[0018] Preferably, the U-shaped structure includes a decoder and an encoder. Each layer of the encoder includes two convolutional block modules and a 2×2 max pooling layer, and each layer of the decoder includes a transposed convolutional layer with a stride of 2 and a convolutional kernel size of 2×2 and two convolutional block modules; the output of each transposed convolutional layer and the input of the max pooling layer corresponding to the corresponding layer of the encoder are connected through a deep connection layer and then used as the input of the corresponding layer of the decoder.

[0019] Preferably, the feature enhancement structure is used to enlarge the feature maps of each layer of the decoder and keep their sizes consistent, and then these feature maps are connected through a deep connection layer.

[0020] Preferably, the multi-layer downsampling structure is used to output a multi-channel input image as a single-channel image, which is composed of multiple convolutional layers, and the number of filters in each convolutional layer decreases sequentially from front to back.

[0021] Preferably, the training of the initial denoising model in Step103 includes the following steps:

[0022] D1. Use the seismic noisy data in set a of the processing blocks as the input data of the U-shaped structure to obtain multiple groups of seismic noisy data images representing different features;

[0023] D2. Use the multiple groups of seismic noisy data images obtained in D1 as the input data of the feature enhancement structure to obtain a group of seismic noisy data images representing different features in different channel intervals;

[0024] D3. Use the seismic noisy data images representing different features in different channel intervals obtained in D2 as the input data of the multi-layer downsampling structure to obtain a group of single-channel seismic noisy data images;

[0025] D4. Use the single-channel seismic noisy data images obtained in D3 and the seismic noisy data in set b of the processing blocks as the input data of the Adam optimizer respectively to obtain two groups of feature data. The feature data is calculated through a loss function to obtain optimized model parameters, and the optimized model parameters are substituted into the initial denoising model for model optimization to obtain the MFE-UNet denoising model.

[0026] Preferably, the loss function Loss in D4 is:

[0027]

[0028] where H, W, and C are the height, width, and number of channels of the seismic noisy data image respectively, y p is the p-th pixel value of the seismic noisy data image output by the denoising model, and t p is the p-th pixel value of the corresponding seismic noisy data image in set b of the processing blocks.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] The present invention adopts an improved training method. First, a new seismic gather is constructed, then processing blocks are generated, and finally a set of processing blocks is constructed to build a training set required for training a denoising model. The constructed initial denoising model is trained using this training set to obtain an MFE-UNet denoising model, and the seismic noisy data image is denoised by the established MFE-UNet denoising model. The present invention can directly train the initial denoising model on the test area and obtain the MFE-UNet denoising model when the training is not fully converged, which makes full use of the convergence performance of the convolutional neural network, thereby greatly reducing the training cost (training time and the quality requirement of the training set samples) of the MFE-UNet denoising model and enhancing the denoising effect. In addition, the method proposed by the present invention enables the MFE-UNet denoising model to accelerate the mastery of effective data and reduce the learning of random noise, and thus can use the MFE-UNet denoising model to denoise based on the difference in the learning difficulty between the two. By increasing the size of the processing block to increase the learning difficulty of random noise, the network can better master the global features of the effective signal. And the random selection method conforms to the uniform distribution, enabling the network to more evenly master the effective data in each small area of the entire architecture, avoiding local deviation in the learning effect of the network due to uneven distribution. At the same time, the denoising principle involved in this method can minimize the loss of effective data in the noisy seismic data. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0032] Figure 1 is a flowchart of the improved training method in the embodiment of the present invention;

[0033] Figure 2 is a flowchart of the method for establishing the denoising model of the new convolutional neural network MFE-UNet in the embodiment of the present invention;

[0034] Figure 3 is a schematic diagram of the neural network structure of the MFE-UNet denoising model in the embodiment of the present invention;

[0035] Figure 4 is a schematic diagram of the structure of the U-shaped structure module in the embodiment of the present invention;

[0036] Figure 5 is a schematic diagram of the structure of the feature enhancement structure module in the embodiment of the present invention;

[0037] Figure 6It is a schematic structural diagram of a multi-layer descending channel structure module in an embodiment of the present invention;

[0038] Figure 7 It is a schematic structural diagram of a convolutional block module in an embodiment of the present invention;

[0039] Figure 8 It is a schematic diagram of model training for an initial denoising model in an embodiment of the present invention;

[0040] Figure 9 It is a synthetic seismic data image before and after denoising processing by the MFE-UNet denoising model in an embodiment of the present invention;

[0041] Figure 10 It is the denoising situation of the synthetic seismic data image in an embodiment of the present invention: denoising results: (a) MFE-Unet, snr = 18.96 (b) fx deconvolution, snr = 7 (c) multi-channel singular spectrum analysis MSSA, snr = 4.3 (d) Dncnn, snr = 0.59 (e) UNet, snr = 9.93; noise removal: (f) MFE-UNet (g) fx deconvolution (h) MSSA (i) Dncnn (j) UNet;

[0042] Figure 11 It is the FK analysis of the synthetic seismic data image in an embodiment of the present invention: (a) noisy data (b) clean data (c) MFE-UNet (d) fx deconvolution (e) MSSA (f) Dncnn (g) Unet;

[0043] Figure 12 It is an actual seismic data image before and after denoising processing by the MFE-UNet denoising model in an embodiment of the present invention;

[0044] Figure 13 It is the denoising situation of the actual seismic data image in an embodiment of the present invention: (a) actual seismic data, denoising results: (b) MFE-UNet (c) fx deconvolution (d) MSSA (e) Dncnn (f) UNet.

[0045] Figure 14 It is the noise removal situation of the actual seismic data image in an embodiment of the present invention: (a) MFE-UNet (b) fx deconvolution (c) MSSA (d) Dncnn. Detailed implementation manners

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. The components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but only represents the selected embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0047] The purpose of this specific implementation manner is to provide a method for suppressing seismic random noise based on MFE-UNet, and obtain a denoising model with stronger denoising ability, better denoising effect, lower training cost, lower training set sample requirements, and shorter training time for seismic noisy data images compared with traditional training methods. In particular, this method can directly train the initial denoising model on the test area and obtain the MFE-UNet denoising model when the training is not fully converged, which makes full use of the convergence performance of the convolutional neural network, thereby greatly reducing the training cost (training time and training set sample quality requirements) of the MFE-UNet denoising model and enhancing the denoising effect. At the same time, the denoising principle involved in this method will minimize the loss of effective data in the noisy seismic data.

[0048] Hereinafter, the embodiments will be described with reference to the drawings. In addition, the embodiments shown below do not limit the content of the invention described in the claims in any way. In addition, all the contents of the configurations shown in the following embodiments are not limited to those necessary for the solution of the invention described in the claims.

[0049] The following will be combined with Figures 1-14 , and specifically describe the method for suppressing seismic random noise based on MFE-UNet provided by the present invention. The method for suppressing seismic random noise based on MFE-UNet includes an improved training method and a method for establishing a denoising model of the new convolutional neural network MFE-UNet. In this embodiment, the improved training method is as shown in Figure 1 , and includes the following steps:

[0050] S01: Construct a new trace gather: Divide the seismic traces involved in the selected area into two sub-trace gathers according to the parity of the trace numbers, namely the odd-trace set and the even-trace set. The positions of each seismic trace in the two sub-trace gathers should remain relatively unchanged.

[0051] It should be noted that the selected gather area should meet the following requirements: noisy seismic data; continuous area gathers; the total number of seismic traces involved is even and not too large. In this embodiment, a region containing 400 traces and 240 sampling points in the pre-stack depth migration data of the Marmousi2 model is selected as the synthetic seismic data, and 0 dB of Gaussian white noise is added for training and testing. The actual seismic data selects a gather area containing 400 traces in the actual marine seismic data of the United States.

[0052] S02: Generate processing blocks: Randomly select a position within the range of the new gather area, and generate two corresponding processing blocks at this position from the two sub-gathers. These two processing blocks are respectively used as the input and output of the denoising model. The size of the processing block should be set to be relatively large to increase the learning difficulty of the denoising model for random noise, and at the same time enable the denoising model to better master the global characteristics of the effective signal. In this embodiment, the size of the processing block is set to 80*80.

[0053] It should be noted that the random selection method used should conform to the uniform distribution, so that the denoising model can more evenly master the effective data in each small area of the entire gather area, and avoid local deviation in the learning effect of the denoising model due to uneven distribution. At the same time, the seismic traces involved in the above two corresponding processing blocks were originally adjacent seismic traces, with similar seismic phase characteristics and seismic interface characteristics, and the two are spatially correlated. When the two processing blocks are respectively used as the input and output of the denoising model, the effective signals contained in them are similar and are more easily mastered by the denoising model. The random noise contained in the two is unpredictable and lacks spatial correlation as a kind of disordered time series, so it is difficult for the denoising model to master compared with the effective signal. We use this difference in learning difficulty for denoising.

[0054] S03: Construct a set of processing blocks: Repeat the step of generating processing blocks to generate multiple groups of processing block combinations. Among them, the processing blocks generated in the odd trace set form set a of processing blocks and are used as the input of the denoising model; the processing blocks generated in the even trace set form set b of processing blocks and are used as the output of the denoising model. Set a of processing blocks and set b of processing blocks together constitute the training set. In this embodiment, both the synthetic seismic data and the actual seismic data construct a training set containing 384 groups of samples.

[0055] Furthermore, a test set is divided on the selected gather area. It should be noted that the data area involved in the test set is the same as or slightly smaller than the data area involved in the training set. Only in this way can the convergence performance of the convolutional neural network be fully utilized, reduce the training cost (training time and requirements for the quality of training set samples) of the denoising model, and improve the denoising effect of the denoising model. In this embodiment, the areas involved in the training set and the test set of the synthetic seismic data and the actual seismic data are the same.

[0056] In this embodiment, the method for establishing the denoising model of the new convolutional neural network MFE-UNet is as Figure 2 shown, and includes the following steps:

[0057] S01: Processing of noisy seismic data: Select a preset trace gather area in the noisy seismic data as the data area, and form corresponding training sets and test sets through an improved training method, where the improved training method is the above-mentioned improved training method.

[0058] S02: Constructing an initial denoising model: The initial denoising model includes three basic secondary structures, namely, a U-shaped structure, a feature enhancement structure, and a multi-layer trace reduction structure, as Figure 3 shown; The U-shaped structure is established based on a convolutional block module, a max pooling layer, a transposed convolutional layer, and a depth connection layer. The feature enhancement structure is established based on a convolutional block module and a transposed convolutional layer. The multi-layer trace reduction structure is established based on a convolutional layer.

[0059] Specifically, please refer to Figure 4 , the U-shaped structure includes two parts, a decoder and an encoder. Each layer of the encoder includes two convolutional block modules and a 2×2 max pooling layer. Each layer of the decoder includes a transposed convolutional layer with a stride of 2 and a convolutional kernel size of 2×2 and two convolutional block modules. The output of each transposed convolutional layer is connected to the input of the max pooling layer of the corresponding layer of the encoder through a depth connection layer, and then used as the input of the corresponding layer of the decoder.

[0060] It should be noted that different from the U-shaped structure of UNet which only outputs at the shallowest layer, the decoder part of the U-shaped structure proposed in the present invention can output at each layer. Specifically, the output after the last operation of the shallowest layer and the output after transposed convolution of non-shallowest layers can all or selectively be input into the next structure. The function of such a setting is that the deep features at different depths of the decoder and the shallow features lost due to convolutional pooling operations at different depths can all be processed subsequently. In this embodiment, the depth of the U-shaped structure is 3 layers, and the output of each layer of the decoder is all input into the next structure, that is, the feature enhancement structure. The number of convolutional kernels of the transposed convolutional layer is set to 128 and 64 in sequence from the deepest layer to the shallowest layer.

[0061] The structure of the convolutional block module is as Figure 7As shown, it includes two consecutive convolutional layers with a step size of 1 and a convolutional kernel size of 3×3, a normalization layer (BN), and a LeakyReLU activation function layer. Among them, LeakyReLU = max(0, x) + a * min(0, x), and the slope a ∈ (0, 1). It should be noted that the double convolutional structure composed of two consecutive convolutional layers can accelerate the convergence speed of the denoising model training and reduce the training cost; the characteristic that the LeakyRelu activation function layer is not 0 in the negative interval enables the denoising model to avoid the problem of neuron death when updating parameters, but reduces the convergence speed of the denoising model training. In this embodiment, the slope a of the LeakyReLU activation function is set to 0.01. With this setting, both the problem of neuron death is avoided and the influence of the LeakyReLU activation function on the convergence speed of the denoising model is reduced.

[0062] The calculation process of the convolutional block module is as follows: the input data first undergoes the calculation of a convolutional layer with a convolutional kernel size of 3×3, then undergoes the calculation of a convolutional layer with a convolutional kernel size of 3×3 again, and then successively undergoes the operations of a normalization layer (BN layer) and a LeakyReLU activation function layer to obtain the output data. It should be noted that the number of convolutional kernels in the convolutional layer of the convolutional block module varies with different layers, but is the same in the same layer and its corresponding layer. In this embodiment, the number of convolutional kernels in the convolutional layer of the convolutional block module is set to 256, 128, and 64 in sequence from the deepest layer to the shallowest layer.

[0063] The feature enhancement structure is as Figure 5 shown. The feature enhancement structure can accept multiple inputs, enlarge the size of the feature maps from each input to make them consistent, and finally connect these feature maps through a depth connection layer. In this embodiment, the feature enhancement structure enlarges the feature maps of each layer of the decoder and then connects them through a depth connection layer. With this setting, the output feature map of the feature enhancement structure simultaneously has deep features from different depths and shallow features that are lost due to pooling operations at different depths, that is, different channel intervals of the output feature map represent the features of each layer of the decoder.

[0064] It should be noted that the output of the MFE-UNet denoising model is a single-channel image; therefore, the feature enhancement structure can control the channel number ratio of the channels representing different features by setting the number of convolutional kernels in the convolutional layer in different layers, that is, the ratio of the total number of channels representing the current feature to the total number of output channels of the depth connection layer, so as to control the bias of the output image of the MFE-UNet denoising model towards a certain layer of features.

[0065] In this embodiment, the magnification process is implemented through a transposed convolutional layer and a double convolutional block module structure. The feature enhancement structure has three layers, which correspond one by one to each layer of the U-shaped structure. The number of convolutional kernels involved in the convolutional layers of the convolutional block modules at the same layer is the same, and is set to 128, 64, and 64 in sequence from the deepest layer to the shallowest layer, that is, the channel number ratios of each layer are 2:1:1 in sequence. Such a setting makes the output image of the MFE-UNet denoising model more biased towards the bottom layer features while taking into account the features of each layer. Because under the above improved training method, the deep layer features are basically the features of valid data, while the shallow layer features contain the features of both valid data and random noise.

[0066] The multi-layer downsampling structure is as Figure 6 shown, and is composed of multiple convolutional layers. The number of filters of each convolutional layer decreases sequentially from front to back. Its function is to output a single-channel image from a multi-channel input image. Such a setting, on the one hand, the multi-layer and gradually downsampling structure can retain the features from the feature enhancement structure as much as possible and reduce the degree of feature loss caused by convolutional operations; on the other hand, the multi-layer structure makes the number of selected parameters increase, which makes the structure have a larger range and more flexibility in parameter selection, so as to play a greater role. In this embodiment, the multi-layer downsampling structure is composed of four convolutional layers with a stride of 1 and a convolutional kernel size of 1*1, and the corresponding number of convolutional kernels is set to 64, 16, 4, and 1 in sequence from front to back.

[0067] S03: Training the initial denoising model: Using the training set as the input data of the initial denoising model for model training. After the model training is completed, the MFE-UNet denoising model is obtained. The MFE-UNet denoising model performs denoising processing on the noisy seismic data in the test set for model testing.

[0068] Furthermore, please refer to Figure 8 , the specific steps of model training include: using the seismic noisy data in the a-group of the processing block set as the input data of the U-shaped structure to obtain multiple groups of seismic noisy data images representing different features; using the multiple groups of seismic noisy data images as the input data of the feature enhancement structure to obtain a group of seismic noisy data images representing different features in different channel intervals; using the seismic noisy data images representing different features in different channel intervals as the input data of the multi-layer downsampling structure to obtain a group of single-channel seismic noisy data images; using the single-channel seismic noisy data images and the seismic noisy data in the b-group of the processing block set as the input data of the Adam optimizer respectively to obtain two groups of feature images. The feature data is calculated through the loss function to obtain the optimized model parameters, and the optimized model parameters are substituted into the initial denoising model for model optimization to obtain the MFE-UNet denoising model. It should be noted that the sizes of the seismic noisy data images in the training set and the test set are both 80*80*1.

[0069] Further, when training the initial denoising model, a suitable loss function needs to be selected to optimize the model's parameters.

[0070] In a specific embodiment, the used loss function Loss is:

[0071]

[0072] where H, W, and C are the height, width, and number of channels of the seismic noisy data image, respectively, and y p is the p-th pixel value of the seismic noisy data image output by the denoising model, and t p is the p-th pixel value of the expected output image, that is, the corresponding seismic noisy data image in the b-th group of processing blocks.

[0073] It should be noted that the loss function used in the training of this denoising model does not perform an average calculation to prevent rounding errors in the computer for the averaged value. At the same time, the denoising model does not need to be trained until it is fully convergent, and the training can be ended when it is initially convergent.

[0074] The number of training rounds g should satisfy:

[0075] g = 3a + 1 (2)

[0076] where a is the round number at which the loss curve first shows a stable turn.

[0077] In this embodiment, the hyperparameters in the model training process are set as follows: the batch size is set to 32, and a total of 10 epochs are trained. The parameter optimization algorithm for the MFE-UNet denoising model is Adam, and the learning rate is set to 10 -3 ; when the number of times the model is trained reaches 10, the training is stopped to obtain the optimized model parameters.

[0078] After the model training is completed, the seismic noisy data is used as the input data of the MFE-UNet denoising model, and the seismic denoised data image can be output. In this embodiment, as Figure 9 shown, are the seismic data images before and after the denoising process of the MFE-UNet denoising model. Among them, the left side is the synthetic seismic noisy data image before processing, and the right side is the synthetic seismic denoised data image after processing. The time taken for this training is 66 seconds, and the number of samples in the training set used is 384.

[0079] In this embodiment, the denoising situations of various denoising methods for the synthetic seismic noisy data image are as Figure 10As shown, the denoising results are as follows: (a) MFE-UNet, snr = 18.96; (b) fx deconvolution, snr = 7; (c) multi-channel singular spectrum analysis (MSSA), snr = 4.3; (d) Dncnn, snr = 0.59; (e) UNet, snr = 9.93. The noise removal methods are: (f) MFE-UNet; (g) fx deconvolution; (h) MSSA; (i) Dncnn; (j) UNet. From Figure 10 It can be seen that the proposed MFE-UNet denoising model has the best denoising effect and the least loss of effective data during the denoising process.

[0080] In this embodiment, the FK analysis of the denoised images of the synthetic seismic noisy data images after denoising by various denoising methods is as Figure 11 shown. Among them, (a) noisy data; (b) clean data; (c) MFE-UNet; (d) fx deconvolution; (e) MSSA; (f) Dncnn; (g) Unet. It can be seen that the FK analysis image of the proposed MFE-UNet denoising model is the most similar to the effective data.

[0081] In this embodiment, the actual seismic noisy data image is used as the input data of the MFE-UNet denoising model, and the actual seismic denoised data image can be output. The actual seismic data images before and after the denoising process of the MFE-UNet denoising model are as Figure 12 shown. Among them, the left side is the actual seismic noisy data image before processing, and the right side is the actual seismic denoised data image after processing. The training time for this time is 67 seconds, the number of samples in the training set used is 384, and the hyperparameter settings for model training are the same as above.

[0082] In this embodiment, the denoising situations of various denoising methods for the actual seismic noisy data image are as Figure 13 shown, where (a) actual seismic data, and the denoising results are: (b) MFE-UNet; (c) fx deconvolution; (d) MSSA; (e) Dncnn; (f) UNet. It can be seen that the proposed MFE-UNet denoising model has the best denoising effect.

[0083] In this embodiment, the random noise removed by various denoising methods from the actual seismic noisy data image is as Figure 14 shown, where (a) MFE-UNet; (b) fx deconvolution; (c) MSSA; (d) Dncnn. It can be seen that the proposed MFE-UNet denoising model has the least loss of effective data during the denoising process.

[0084] The present invention also provides a method for suppressing seismic random noise based on MFE-UNet, comprising the following steps: using the seismic noisy data image as the input data of the MFE-UNet denoising model, and then obtaining the seismic denoised data image, where the MFE-UNet denoising model is the above-mentioned MFE-UNet denoising model. With such a setting, the denoising effect of the seismic noisy data image is better, and the computational amount in the denoising process is reduced.

[0085] The applicant declares that the present invention uses the above examples to illustrate the detailed method of the present invention, but the present invention is not limited to the above detailed method, that is, it does not mean that the present invention must rely on the above detailed method to be implemented. Those skilled in the art should understand that any improvement to the present invention, the equivalent transformation of the raw materials of the present invention, the addition of auxiliary components, and the selection of specific conditions and methods, etc., all fall within the protection scope and the disclosure scope of the present invention.

Claims

1. A seismic random noise suppression method based on MFE-UNet, characterized in that The method steps are as follows: Step1. Construct an MFE-UNet denoising model; Step2. Use the seismic noisy data image as the input data of the MFE-UNet denoising model to obtain a seismic denoised data image.

2. The seismic random noise suppression method based on MFE-UNet according to claim 1, wherein, In Step1, the method for constructing the MFE-UNet denoising model includes: Step101. Seismic noisy data processing: Select a preset trace gather area in the seismic noisy data as the data area, and form corresponding training sets and test sets through an improved training method; Step102. Construct an initial denoising model: The initial denoising model includes a U-shaped structure, a feature enhancement structure, and a multi-layer downsampling structure; The U-shaped structure is established based on convolutional block modules, max pooling layers, transposed convolutional layers, and skip connection layers; The feature enhancement structure is established based on convolutional block modules and transposed convolutional layers; The multi-layer downsampling structure is established based on convolutional layers; Step103. Train the initial denoising model: Use the training set as the input data of the initial denoising model for model training. After the model training is completed, an MFE-UNet denoising model is obtained, and the MFE-UNet denoising model performs denoising processing on the seismic noisy data in the test set for model testing.

3. The method for suppressing seismic random noise based on MFE-UNet according to claim 2, wherein, The specific improved training method in Step101 includes: S1. Construct a new trace gather: Divide the seismic traces involved in the trace gather area into two sub-trace gathers, an odd trace set and an even trace set, according to the parity of the trace numbers, where the positions of each seismic trace in the sub-trace gathers remain relatively unchanged; S2. Generate processing blocks: Randomly select a position within the range of the new trace gather, and generate two corresponding processing blocks in the two sub-trace gathers at this position, which are used as the input and output of the denoising model respectively; S3. Construct a set of processing blocks: Repeat Step S2 to generate multiple groups of processing block combinations, where the processing blocks generated in the odd trace set form set a of the processing block set, which is used as the input of the denoising model; The processing blocks generated in the even trace set form set b of the processing block set, which is used as the output of the denoising model; Set a and set b of the processing block set together form the training set.

4. The method for suppressing seismic random noise based on MFE-UNet according to claim 3, wherein The size of the processing block in Step S2 is 80*80.

5. The method for suppressing seismic random noise based on MFE-UNet according to claim 3 or 4, characterized in that In Step101, divide processing blocks and generate a test set within the preset trace gather area, and the data area involved in the test set is the same as the data area involved in the training set.

6. The method for suppressing seismic random noise based on MFE-UNet according to claim 2, wherein The U-shaped structure includes two parts, a decoder and an encoder. Each layer of the encoder includes two convolutional block modules and a 2*2 max pooling layer, and each layer of the decoder includes a transposed convolutional layer with a stride of 2 and a convolutional kernel size of 2*2 and two convolutional block modules; The output of each transposed convolutional layer and the input of the corresponding max pooling layer of the encoder are connected through a skip connection layer and then used as the input of the corresponding layer of the decoder.

7. The method for suppressing seismic random noise based on MFE-UNet according to claim 6, characterized in that, The feature enhancement structure is used to enlarge the feature maps of each layer of the decoder and keep their sizes consistent, and then connect these feature maps through a skip connection layer.

8. The method for suppressing seismic random noise based on MFE-UNet according to claim 7, characterized in that, The multi-layer downsampling structure is used to output a single-channel image from a multi-channel input image, which consists of multiple convolutional layers, and the number of filters in each convolutional layer decreases sequentially from front to back.

9. The method for suppressing seismic random noise based on MFE-UNet according to claim 3, wherein The training of the initial denoising model in Step103 includes the following steps: D1. Use the seismic noisy data in set a of processing blocks as the input data of the U-shaped structure to obtain multiple sets of seismic noisy data images representing different features; D2. Use the multiple sets of seismic noisy data images obtained in D1 as the input data of the feature enhancement structure to obtain a set of seismic noisy data images representing different features in different channel intervals; D3. Use the seismic noisy data images representing different features in different channel intervals obtained in D2 as the input data of the multi-layer channel reduction structure to obtain a set of single-channel seismic noisy data images; D4. Use the single-channel seismic noisy data images obtained in D3 and the seismic noisy data in set b of processing blocks as the input data of the Adam optimizer respectively to obtain two sets of feature data. The feature data is calculated by the loss function to obtain the optimized model parameters, and the optimized model parameters are substituted into the initial denoising model for model optimization to obtain the MFE-UNet denoising model.

10. The method for suppressing seismic random noise based on MFE-UNet according to claim 9, wherein, The loss function Loss in D4 is: Where H, W, and C are the height, width, and number of channels of the seismic noisy data image, respectively, and y p is the p-th pixel value of the seismic noisy data image output by the denoising model, and t p is the p-th pixel value of the corresponding seismic noisy data image in the b-th group of processing blocks.

Images