Multiple suppression method and system based on unsupervised learning algorithm
Multiple wave suppression is performed by U-net neural network model based on unsupervised learning algorithm, and the problems of low multi-wave suppression efficiency and large calculation amount in the prior art are solved, and a more efficient and accurate multi-wave suppression effect is achieved.
Patent Information
- Application Number
- CN202410200139.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-02-23
- Publication Date
- 2025-08-26
AI Technical Summary
The existing multi-wave suppression method is inefficient in seismic data processing in complex geological areas. It depends on a priori assumptions, has a large calculation amount and is costly, and it is difficult to effectively remove multi-waves of far-off distances.
The multi-wave suppression method based on unsupervised learning algorithm is adopted, and the U-net neural network model is constructed using traditional multi-wave prediction methods. Multi-wave suppression is performed through training and generalization neural network models, and the matching accuracy and efficiency are improved in combination with the attention mechanism.
The accuracy and efficiency of multiple wave suppression are improved, the assumptional conditional limitations in traditional methods are avoided, the problem of insufficient tag data is solved, and more efficient multiple wave suppression is achieved.
Smart Images
Figure CN120538892A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a field of multi-wave suppression, and in particular to a method and system based on an unsupervised learning algorithm. Background Art
[0002] Existing multiple suppression methods all have limitations, and their efficiency and successful application depend largely on corresponding prior assumptions, especially for seismic data acquired from complex geological regions. For example, predictive deconvolution can achieve good results for data at close offsets and zero offsets, but the periodicity of multiples cannot be guaranteed for data at large offsets. Radon transforms, FK filters, CMP stacking, and KL filters are effective for suppressing multiples at far offsets, but removing multiples at close offsets is more difficult. The inverse scattering series (ISS) method, based on scattering theory, does not require prior information about the subsurface medium and can generate multiples of the same order associated with all interfaces in a single prediction. In the absence of effective means to distinguish between significant waves and multiples, ISS is the most effective multiple suppression method. However, this method is computationally intensive and costly, making it difficult to predict multiples at far offsets. The Estimation of Primaries by Sparse Inversion (EPSI) method, based on inversion theory, directly obtains primary waves through iterative inversion, eliminating the multiple prediction and adaptive matching subtraction steps required in the surface-related multiple elimination (SRME) method. This avoids the damage to the primary waves caused by adaptive matching subtraction. However, the EPSI iteration process is computationally intensive, and the estimation of unstable source wavelets makes its application more challenging. The common-focus-point (CFP) method is relatively adaptable to more complex geological conditions, but it only produces multiples related to a specific interface in a single prediction, and relies to a certain extent on the initial velocity model to obtain an accurate focusing operator.
[0003] The virtual event method can accurately predict multiples, but it places high demands on the observation system. It requires sequentially extracting information from different primary waves in the seismic data to construct multiples for the relevant horizons, a process that relies on manual operation. This process is difficult to implement for actual seismic data from complex regions and is computationally intensive. The Marchenko multiple suppression method is theoretically novel and effective, having been applied to model data and some actual data. However, this method places high demands on the observation system and signal-to-noise ratio of the seismic data, and the input data is deconvolved, resulting in a high computational load. Further research is needed to improve its applicability. Summary of the Invention
[0004] In view of the above problems, the present invention is proposed to provide a multiple wave suppression method and system based on an unsupervised learning algorithm to overcome the above problems or at least partially solve the above problems.
[0005] According to one aspect of the present invention, a multiple wave suppression method based on an unsupervised learning algorithm is provided, the suppression method comprising:
[0006] Step S1: Predicting the multiple wave model using a traditional multiple wave prediction method;
[0007] Step S2: constructing a neural network model;
[0008] Step S3: inputting the predicted multiple wave model into the constructed neural network for training until the training results are stable;
[0009] Step S4: Repeat steps S1 and S3 using different earthquake data to train the neural network model with the strongest generalization ability;
[0010] Step S5: Perform prediction using the trained neural network model obtained in step S1;
[0011] Step S6: Output the seismic data after multiple wave suppression.
[0012] Optionally, the step S1: predicting the multiple wave model using a traditional multiple wave prediction method specifically includes:
[0013] The multiple waves predicted by the SRME method are obtained by convolution of the total wave field and the primary wave field, that is,
[0014] M=P′0P (1)
[0015] Where P is the total wave field, P′0 is the current estimated primary wave, and M is the predicted multiple wave;
[0016] Based on the current estimated primary wave and full wavefield data, the iterative prediction of multiple waves is expressed as
[0017] M (n+1) =(P′0) n P (2)
[0018] Among them, (P′0) 0 =P,(P′0) n is the estimated wave, n is the number of iterations;
[0019] Optionally, the step S1: predicting the multiple wave model using a traditional multiple wave prediction method further includes:
[0020] The predicted multiple waves are corrected using an adaptive matched filter to make the predicted multiple waves consistent with the actual multiple waves. The matched multiple waves are expressed as
[0021] M′0=f(t)M (3)
[0022] The effective wave after suppressing the multiple waves is
[0023]
[0024] Where, f(t) is the filter;
[0025] Using the iterative method, formula (4) can be written as
[0026]
[0027] The final seismic wave field after suppressing the multiple waves is
[0028] Optionally, the step S2: constructing a neural network model specifically includes:
[0029] It adopts the U-net network structure, including encoder, decoder and connection bridge.
[0030] Optionally, the encoder and the decoder both include two convolutional layers and one dropout layer, downsampling is a pooling layer, and upsampling is a deconvolution layer;
[0031] The dropout layer prevents overfitting, and upsampling maps the low-dimensional data contained in the primary wave to a high-dimensional space and reconstructs the primary wave.
[0032] In addition, each convolutional layer and deconvolution layer is followed by an activation function. The bridge is similar to the encoder, but does not include the maximum pooling layer and dropout layer.
[0033] Optionally, the depth of the U-net network structure is 4.
[0034] Optionally, the step S3: inputting the predicted multiple wave model into the constructed neural network for training until the training result is stable specifically includes:
[0035] Normalize the input data set;
[0036] The smooth L1 loss function eliminates the sensitivity problem of the L2 loss function and has all the advantages of the L1 loss function, with faster convergence speed and smaller gradient changes for network training;
[0037] Its expression is
[0038]
[0039] Wherein, x=P0-P″0 is the difference between the actual primary wave P0 and the predicted primary wave P″0.
[0040] Select the Aadm optimizer to iteratively update each parameter.
[0041] Optionally, the normalizing the input data set specifically includes:
[0042] Since the neural network uses the ELU activation function, the amplitude of the data set is normalized to [-1, 1]. The calculation formula is as follows
[0043]
[0044] Among them, x is the amplitude value of the data set, x normal is the normalized value, and max|x| is the global maximum absolute amplitude of the data set.
[0045] Optionally, after selecting the Aadm optimizer to iteratively update each parameter, the method further includes:
[0046] During the initialization phase, a small random number is assigned to each network connection, and a random number is used to initialize the bias of each neuron.
[0047] Optionally, the Aadm optimizer is selected to iteratively update the learning rate of each parameter from 10 -3 start.
[0048] The present invention further provides a multiple wave suppression system based on an unsupervised learning algorithm, which applies the above-mentioned multiple wave suppression method based on an unsupervised learning algorithm. The suppression system specifically includes:
[0049] A multiple wave model prediction module is used to predict the multiple wave model using a traditional multiple wave prediction method;
[0050] Network model building module, used to build neural network models;
[0051] The neural network training module is used to input the predicted multiple wave model into the constructed neural network for training until the training results are stable;
[0052] A neural network model screening module is used to repeat steps S1 and S3 using different earthquake data to train a neural network model with the strongest generalization ability;
[0053] A model prediction module, used to make predictions based on the trained neural network model obtained in step S1;
[0054] The suppressed seismic data output module is used to output the seismic data after multiple wave suppression.
[0055] The present invention provides a multiple wave suppression method and system based on an unsupervised learning algorithm. The suppression method includes: step S1: predicting a multiple wave model using a traditional multiple wave prediction method; step S2: constructing a neural network model; step S3: inputting the predicted multiple wave model into the constructed neural network for training until the training results are stable; step S4: repeating steps S1 and S3 using different seismic data to train the neural network model with the strongest generalization ability; step S5: performing prediction using the trained neural network model obtained in step S1; and step S6: outputting the seismic data after multiple wave suppression. Using deep learning methods, the seismic data itself can be used to mine the characteristics of the relationship between data for multiple wave suppression research, thereby avoiding the hypothetical conditions in traditional methods, better matching the predicted multiple wave model with the actual multiple waves in the original data, and improving the accuracy of the multiple wave suppression results. At the same time, the introduction of the attention mechanism significantly improves the efficiency of multiple wave suppression.
[0056] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0058] Figure 1 A schematic diagram of the flow of the model-driven deep learning multiple wave suppression method provided by the present invention;
[0059] Figure 2 The U-net network model structure diagram for deep learning in this invention;
[0060] Figure 3 A network framework diagram of the attention-based mechanism for deep learning in this invention. DETAILED DESCRIPTION
[0061] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.
[0062] The terms "comprises" and "comprising" and any variations thereof in the description, embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.
[0063] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.
[0064] like Figure 1 As shown, the purpose of the present invention is achieved through the following technical measures: a model-driven deep neural network multiple wave suppression method, comprising: (1) using a traditional multiple wave prediction method to predict a multiple wave model; (2) constructing a neural network model; (3) inputting the predicted multiple wave model into the constructed neural network for training until the training result is stable; (4) repeating (1) and (3) using different seismic data to train a neural network model with strong generalization ability; (5) inputting other multiple wave models obtained in (1) into the trained neural network model for prediction; (6) outputting the seismic data after the multiple waves are suppressed.
[0065] Traditional multiple prediction methods
[0066] The multiple waves predicted by the SRME method are obtained by convolution of the total wave field and the primary wave field, that is,
[0067] M=P′0P (1)
[0068] Where P is the total wavefield, P′0 is the current estimated primary wave, and M is the predicted multiple wave. Based on the current estimated primary wave and the full wavefield data, the iterative prediction of the multiple wave is expressed as
[0069] M (n+1) =(P′0) n P (2)
[0070] Among them, (P′0) 0 =P,(P′0) n is the estimated wave, and n is the number of iterations.
[0071] Compared with the actual multiple waves in the wave field, the predicted multiple waves have errors in amplitude, frequency and phase. The adaptive matched filter is used to correct them so that the predicted multiple waves are consistent with the actual multiple waves. The matched multiple waves can be expressed as
[0072] M′0=f(t)M (3)
[0073] The effective wave after suppressing the multiple waves is
[0074] P′0=PM′0
[0075] =Pf(t)M (4)
[0076] Where f(t) is the filter.
[0077] Using the iterative method, formula (4) can be written as
[0078]
[0079] The final seismic wave field after suppressing the multiple waves is
[0080] Network model construction
[0081] The present invention mainly adopts U-net network structure, such as Figure 2 As shown, the network structure has a depth of 4 and consists of an encoder, a decoder, and a connecting bridge. Both the encoder and decoder contain two convolutional layers and a dropout layer. Downsampling is performed using a pooling layer, and upsampling is performed using a deconvolution layer. The dropout layer prevents overfitting, while upsampling maps the low-dimensional data contained in the primary wave to a higher-dimensional space and reconstructs the primary wave. Furthermore, each convolutional and deconvolution layer is followed by an activation function (exponential linear units, ELU). The bridge is similar to the encoder, but does not include the max pooling and dropout layers.
[0082] In order to solve the problem of large amount of computation in actual data, the attention mechanism is introduced. Therefore, the overall network model constructed by the present invention is composed of U-net and attention mechanism model.
[0083] When the training of the U-net network model reaches stability, the matched multiple wave model is obtained. The initial matched multiple wave model is merged with the multiple wave model results output by the attention model to generate the final matched multiple wave model. The final matched multiple wave model is then subtracted from the original data to finally obtain the data after multiple wave suppression.
[0084] The present invention mainly addresses the problem that traditional multiple wave suppression methods are restricted by assumptions and the obtained filters usually damage effective waves. The present invention uses deep learning methods to explore the characteristics of the relationship between data and conduct multiple wave suppression research based on the seismic data itself, thereby avoiding the assumptions in traditional methods and better matching the predicted multiple wave model with the actual multiple waves in the original data, thereby improving the accuracy of the multiple wave suppression results. At the same time, the introduction of the attention mechanism greatly improves the efficiency of multiple wave suppression.
[0085] Network model training
[0086] like Figure 3As shown in the figure, the network framework diagram of the attention-based mechanism for deep learning is shown. In order to improve the convergence speed of network model training, the input data set needs to be normalized, especially for field seismic data. Since the neural network uses the ELU activation function, the amplitude of the data set is normalized to [-1, 1]. The calculation formula is as follows
[0087]
[0088] Among them, x is the amplitude value of the data set, x normal is the normalized value, and max|x| is the global maximum absolute amplitude of the data set. The smooth L1 loss function eliminates the sensitivity problem of the L2 loss function and has all the advantages of the L1 loss function, with faster convergence speed and smaller gradient changes for network training. Its expression is
[0089]
[0090] Wherein, x=P0-P″0 is the difference between the actual primary wave P0 and the predicted primary wave P″0.
[0091] In network training, the more complex the network structure and the more training data required, the longer the training time. Therefore, selecting the appropriate optimizer is particularly important. Common optimizers include batch gradient descent, stochastic gradient descent (SGD), mini-batch gradient descent, momentum, Adagrad, RMSProp, Adadelta, and Adam. This paper selects the Adaptive Moment Estimation (Aadm) optimizer to iteratively update each parameter, starting with a learning rate of 10⁻³. It is suitable for optimizing parameters in large-scale data and for problems with strong noise or sparse gradients, and is particularly advantageous for large-scale field data. Furthermore, during the initialization phase, a small random number is assigned to each network connection, and a random number is used to initialize the bias of each neuron.
[0092] The process of multiple wave suppression method based on model-driven deep neural network is attached. Figure 1 , specifically including the following:
[0093] (1) Use traditional multiple wave prediction methods to predict the initial multiple wave model.
[0094] (2) Use the tensorflow framework to build a neural network model based on the attention mechanism and test the network training parameters. Figure 2 , 3.
[0095] (3) The predicted multiple wave model is input into the neural network for training to obtain a network model with high generalization ability.
[0096] (4) Input other predicted multiple wave models into the trained network model, perform multiple wave matching processing, and output the earthquake record after multiple wave suppression.
[0097] Beneficial effects: The deep learning-based multiple suppression method can avoid the assumptions in traditional methods and improve the accuracy of multiple suppression.
[0098] The attention mechanism model introduced in this invention improves the training of neural network models and the multiple wave suppression rate;
[0099] By using traditional methods to predict the multiple wave model and constraining the network training, a multiple wave suppression method based on model-driven unsupervised learning is formed, which solves the problem of insufficient labeled data in supervised learning algorithms.
[0100] By combining deep learning with seismic data processing, a solution for suppressing multiple waves has been established, which helps to effectively suppress multiple waves in seismic data processing;
[0101] The model of the present invention is driven by an unsupervised learning algorithm, which not only avoids the limitations of the assumptions of traditional methods but also effectively solves the problem of insufficient labeled data in actual data in the multiple wave suppression method based on supervised learning algorithms, thereby greatly improving the accuracy and efficiency of multiple wave suppression.
[0102] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A multiple wave suppression method based on an unsupervised learning algorithm, characterized in that: The pressing method comprises: Step S1: Predicting the multiple wave model using a traditional multiple wave prediction method; Step S2: constructing a neural network model; Step S3: inputting the predicted multiple wave model into the constructed neural network for training until the training results are stable; Step S4: Repeat steps S1 and S3 using different earthquake data to train the neural network model with the strongest generalization ability; Step S5: Perform prediction using the trained neural network model obtained in step S1; Step S6: Output the seismic data after multiple wave suppression.
2. The multiple wave suppression method based on an unsupervised learning algorithm according to claim 1, characterized in that: The step S1: using the traditional multiple wave prediction method to predict the multiple wave model specifically includes: The multiple waves predicted by the SRME method are obtained by convolution of the total wave field and the primary wave field, that is, M=P0'P (1) Where P is the total wave field, P0' is the current estimated primary wave, and M is the predicted multiple wave; Based on the current estimated primary wave and full wavefield data, the iterative prediction of multiple waves is expressed as M (n+1) =(P0') n P (2) Among them, (P0') 0 =P,(P0') n is the estimated wave, and n is the number of iterations.
3. The multiple wave suppression method based on unsupervised learning algorithm according to claim 1, characterized in that: The step S1: predicting the multiple wave model using the traditional multiple wave prediction method further includes: The predicted multiple waves are corrected using an adaptive matched filter to make the predicted multiple waves consistent with the actual multiple waves. The matched multiple waves are expressed as M'0=f(t)M (3) The effective wave after suppressing the multiple waves is Where, f(t) is the filter; Using the iterative method, formula (4) can be written as The final seismic wave field after suppressing the multiple waves is 4. The multiple wave suppression method based on an unsupervised learning algorithm according to claim 1, characterized in that: The step S2: constructing a neural network model specifically includes: It adopts the U-net network structure, including encoder, decoder and connection bridge.
5. The multiple wave suppression method based on unsupervised learning algorithm according to claim 4, characterized in that: The encoder and decoder both include two convolutional layers and a dropout layer, downsampling is a pooling layer, and upsampling is a deconvolution layer; The dropout layer prevents overfitting, and upsampling maps the low-dimensional data contained in the primary wave to a high-dimensional space and reconstructs the primary wave. In addition, each convolutional layer and deconvolution layer is followed by an activation function. The bridge is similar to the encoder, but does not include the maximum pooling layer and dropout layer.
6. The multiple wave suppression method based on unsupervised learning algorithm according to claim 4, characterized in that: The depth of the U-net network structure is 4.
7. The multiple wave suppression method based on unsupervised learning algorithm according to claim 1, characterized in that: The step S3: inputting the predicted multiple wave model into the constructed neural network for training until the training result is stable specifically includes: Normalize the input data set; The smooth L1 loss function eliminates the sensitivity problem of the L2 loss function and has all the advantages of the L1 loss function, with faster convergence speed and smaller gradient changes for network training; Its expression is Wherein, x=P0-P0" is the difference between the actual primary wave P0 and the predicted primary wave P0". Select the Aadm optimizer to iteratively update each parameter.
8. The method for multiple wave suppression based on an unsupervised learning algorithm according to claim 7, wherein: The normalization processing of the input data set specifically includes: Since the neural network uses the ELU activation function, the amplitude of the data set is normalized to [-1, 1]. The calculation formula is as follows Among them, x is the amplitude value of the data set, x normal is the normalized value, and max|x| is the global maximum absolute amplitude of the data set.
9. The multiple wave suppression method based on unsupervised learning algorithm according to claim 7, characterized in that: After selecting the Aadm optimizer to iteratively update each parameter, the following steps are also included: During the initialization phase, a small random number is assigned to each network connection, and a random number is used to initialize the bias of each neuron.
10. The multiple wave suppression method based on unsupervised learning algorithm according to claim 7, characterized in that: The Aadm optimizer is selected to iteratively update the learning rate of each parameter from 10 -3 start.
11. A multiple suppression system based on an unsupervised learning algorithm, applying the multiple suppression method based on an unsupervised learning algorithm according to any one of claims 1 to 10, characterized in that: The pressing system specifically includes: A multiple wave model prediction module is used to predict the multiple wave model using a traditional multiple wave prediction method; Network model building module, used to build neural network models; The neural network training module is used to input the predicted multiple wave model into the constructed neural network for training until the training results are stable; A neural network model screening module is used to repeat steps S1 and S3 using different earthquake data to train a neural network model with the strongest generalization ability; A model prediction module, used to make predictions based on the trained neural network model obtained in step S1; The suppressed seismic data output module is used to output the seismic data after multiple wave suppression.