A method for separating and interpolating and synchronizing multi-source data based on iterative training

Abstract

The present invention discloses a method for multi-source data separation and interpolation synchronization processing based on iterative training, comprising the following steps: S1. Construct a neural network model; S2. Perform iterative training on the neural network model; S3. Use the trained neural network model to predict the input, and perform iterative update of the neural network, and thereby implement multi-source data separation and interpolation synchronization processing. The present invention uses an iterative strategy to train multiple sets of network models, captures the mapping relationship between data with different aliasing degrees and different missing trace situations and label data from a set of training data, thereby significantly improving the network model training process, fully considering the efficiency issue, and ensuring the data quality after processing while ensuring the efficient implementation of the algorithm.

Description

Technical Field

[0001] The present invention relates to the processing of multi-source data, and particularly to a method for separating and interpolating and synchronizing multi-source data based on iterative training. Background Art

[0002] In a multi-source observation system, by using a time-delay coding strategy during the acquisition process, the aliasing noise is programmed as incoherent noise in the common receiver gather, common midpoint gather or common offset gather, while the effective signal is coherent. In order to suppress the aliasing noise in multi-source data, algorithms based on denoising and inversion ideas have been developed. The algorithms based on denoising mainly regard the aliasing noise as incoherent random noise and directly suppress the noise in the aliased record. The algorithms based on inversion use time-delay coding and sparse transform to iteratively estimate the aliasing noise. Compared with direct denoising, the algorithms based on the inversion idea have better separation effect, but the computational complexity is significantly greater than that of the denoising algorithms. For the phenomenon of missing traces or irregular sampling in seismic data, data interpolation technology has always been an important means for seismic data reconstruction. At present, a variety of interpolation methods have been developed, mainly divided into two categories: the first category is the prediction interpolation method based on the wave equation, which predicts and interpolates the missing traces according to the propagation law of seismic waves in the underground medium. The second category is the interpolation method based on sparse transform, such as wavelet transform, curvelet transform, Radon transform and Fourier transform, etc. This kind of method is not restricted by the geological model and has good interpolation effect, but it requires continuous iterative adjustment of the threshold and addition of constraints manually, consuming a lot of time and energy. In addition, in recent years, the interpolation method based on dictionary learning of sparse coding has also achieved good results in seismic data processing. By training to establish a complete dictionary, sparse representation and prediction interpolation of seismic data are carried out, and certain results have also been obtained.

[0003] For irregular aliased seismic records, irregular sampling will affect the separation effect of the aliasing noise, and the aliasing noise will in turn affect the accuracy of interpolation prediction. Therefore, separately processing the aliasing noise and interpolation may not achieve satisfactory results. Compared with separately processing interpolation and de-aliasing, it is possible to synchronously separate the aliasing noise and interpolate the seismic traces. For example, the method for synchronously processing noise separation and interpolation based on inversion requires a large amount of computational time during the iterative optimization and noise suppression process, and also relies on frequent parameter adjustment of some constraint terms or thresholds manually to achieve satisfactory results. However, it will prolong the data processing cycle, thus affecting the development of subsequent seismic data interpretation work. The method for directly estimating the aliasing noise and interpolation based on deep learning technology reduces the effect of noise separation and interpolation reconstruction while improving the processing efficiency. This non-iterative processing method will damage the effective signal. Summary of the Invention

[0004] The object of the present invention is to overcome the deficiencies of the prior art and provide a method for separating and interpolating and synchronizing multi-source data based on iterative training, which fully considers the efficiency issue and ensures the quality of the processed data while efficiently implementing the algorithm.

[0005] The object of the present invention is achieved by the following technical solutions: A method for separating and interpolating and synchronizing multi-source data based on iterative training, comprising the following steps:

[0006] S1. Construct a neural network model;

[0007] S2. Perform iterative training on the neural network model;

[0008] S3. Use the trained neural network model to predict the input, and perform iterative update of the neural network, and accordingly achieve the separation and interpolation and synchronization processing of multi-source data.

[0009] The neural network parameters included in the neural network model constructed in step S1 are θ.

[0010] The said step S2 includes:

[0011] S201. If X represents the neural network input and Y represents the target output of the neural network, the loss function is defined as:

[0012]

[0013] where represents the square of the L2 norm; represents the network's prediction of the input X. When the prediction result is closer to Y, the loss is smaller. The neural network gradually updates the network parameter θ by minimizing the loss function;

[0014] By minimizing the loss function, iterative training is performed on the parameter θ, and the parameter θ is updated to the parameter that makes L(θ) minimum;

[0015] Then use the estimated parameter [[ID=at]]

[0016]

[0017] represents the approximate value close to M obtained by neural network prediction, where M represents the unaliased seismic records of all sources. When n sources are excited "synchronously", then where d k represents the conventional seismic data of the k-th source, d ble represents the aliased data of the missing trace gather collected in the field, Denote the time-delay coding inverse operator of the k-th seismic source.

[0018] S202. When there are significant differences in the characteristics or distributions of the training data and the test data, the effective signals will be damaged during the processing of the test data. However, when separating the aliased noise, not only the incoherent signals need to be suppressed, but also the effective signals should be retained as much as possible. Therefore, the parameters obtained from the i-th iterative update are denoted as:

[0019]

[0020] where i ∈ (1, 2, 3, …, m) represents the i-th iteration. At the first iteration, i.e., i = 0, then it is equivalent to the case of . At the next iteration, the input data becomes where the incoherent aliased noise is estimated by the network and subtracted from the mixed record. After multiple iterations, it gradually converges to the clean seismic record M, where m k represents the sampling matrix of the k-th seismic source, which controls the data missing situation, and its element values are 0 and 1. 0 corresponds to the missing trace gather, and 1 corresponds to the acquired trace gather.

[0021] S203. When the input data changes and the same set of network parameters are used to estimate the aliased noise and interpolation, after multiple iterations, the signal-to-noise ratio of the prediction result will decrease. To adapt to this change in the input, the iterative update of is applied to the network training, and the final loss function is obtained as follows

[0022]

[0023] where θ i-1 represents the network parameters obtained from the (i - 1)-th training, and θ i represents the network parameters obtained from the i-th training of the network. When i = 1, at this time, the calculation method of L(θ i ) is the same as that of L(θ) in step S201, that is

[0024] The said step S3 includes:

[0025] S301. In each iteration process, the neural network parameters are trained by minimizing the loss function L(θ i ), that is, θ i-1 is updated to the θ i that minimizes L(θ i );

[0026] S302. After each iteration is completed, bring the iteration result to obtain the result after removing aliasing noise and interpolation;

[0027] S303. After each iteration, the input data will be closer to the target data. At this time, the network will converge quickly, and a higher signal-to-noise ratio and lower loss will be obtained. After multiple iterations, a better iterative separation and interpolation processing effect is obtained.

[0028] The beneficial effects of the present invention are as follows: The present invention uses an iterative strategy to train multiple sets of network models, captures the mapping relationship between data with different aliasing degrees and different missing trace situations and label data from a set of training data, thereby significantly improving the network model training process, fully considering the efficiency issue, and ensuring the data quality after processing while ensuring the efficient implementation of the algorithm. Brief Description of the Drawings

[0029] Figure 1 is a schematic diagram of the method principle of the present invention;

[0030] Figure 2 is a comparison chart of the training effects of two different training strategies;

[0031] Figure 3 is the iterative prediction result of the traditional deep convolutional neural network in the embodiment;

[0032] Figure 4 is the prediction result of the deep convolutional neural network based on iterative training in the embodiment. Detailed Embodiments

[0033] The technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings, but the protection scope of the present invention is not limited to the following.

[0034] The present invention introduces an inversion formula on the basis of a deep neural network, thereby realizing the synchronous processing of high-efficiency and high-quality aliasing noise separation and interpolation reconstruction, and avoiding the disadvantages existing in the previous technologies, that is, the synchronous processing method of the present invention can ensure the efficient processing of irregular aliased records and can ensure the quality of the processed seismic data. Specifically:

[0035] As Figure 1 shown, a method for synchronous processing of multi-source data separation and interpolation based on iterative training includes the following steps:

[0036] S1. Construct a neural network model; the neural network parameters included in the neural network model constructed in step S1 are Θ.

[0037] S2. Perform iterative training on the neural network model;

[0038] The step S2 includes:

[0039] S201. If X represents the neural network input and Y represents the target output of the neural network, the loss function is defined as:

[0040]

[0041] where represents the square of the L2 norm; represents the network's prediction of the input X. When the prediction result is closer to Y, the loss is smaller. The neural network gradually updates the network parameters θ by minimizing the loss function;

[0042] For image processing tasks, if the gap between the input and the target to be predicted is relatively large, the network requires more epochs to converge. For different epochs, the network input remains unchanged, but the network parameters are continuously iteratively updated. By minimizing the loss function, the parameter θ is iteratively trained, that is, the parameter θ is updated to the parameter that minimizes L(θ);

[0043] Then, using the estimated parameter of the updated θ, the estimation result is:

[0044]

[0045] represents an approximation value close to M obtained by neural network prediction, where M represents the unaliased seismic records of all seismic sources. When n seismic sources are excited "synchronously", then where d k represents the conventional seismic data of the k-th seismic source, d ble represents the aliased data of the missing trace gather collected in the field, represents the time-delay coding inverse operator of the k-th seismic source.

[0046] S202. When the characteristics or distributions of the training data and the test data differ greatly, the effective signals will be damaged when processing the test data; however, when separating the aliased noise, not only non-coherent signals need to be suppressed, but also effective signals need to be retained as much as possible. Therefore, the parameter obtained by the i-th iterative update is denoted as:

[0047]

[0048] where i ∈ (1, 2, 3,..., m) represents the i-th iteration. At the first iteration, that is, when i = 0, then At this time, it is equivalent to In the case where, in the next iteration, the input data becomes the incoherent aliasing noise therein is estimated by the network and subtracted from the mixed record; after multiple iterations, it gradually converges to the clean seismic record M. where m k represents the sampling matrix of the k-th source, which controls the missing situation of the data, and its element values are 0 and 1, where 0 corresponds to the missing trace gather and 1 corresponds to the acquired trace gather;

[0049] S203. When the input data changes and the same set of network parameters are used to estimate the aliasing noise and interpolation, after multiple iterations, the signal-to-noise ratio of the prediction result will decrease. To adapt to this change in the input, the iterative update of

[0050]

[0051] is applied to the network training to obtain the final loss function as follows i-1 where θ i represents the network parameters obtained from the (i - 1)-th training, and θ L(θ i ) is calculated in the same way as L(θ) in step S201, that is

[0052] S3. Use the trained neural network model to predict the input and perform iterative updates of the neural network, and accordingly achieve synchronous processing of multi-source data separation and interpolation.

[0053] The said step S3 includes:

[0054] S301. In each iteration process, by minimizing the loss function L(θ i ), train the neural network parameters, that is, update θ i-1 to the θ i that makes L(θ i ) minimum;

[0055] S302. After each iteration is completed, substitute the iteration result into to obtain the result after removing the aliasing noise and interpolation;

[0056] S303. After each iteration, the input data will be closer to the target data. At this time, the network will converge quickly and obtain a higher signal-to-noise ratio and a lower loss. After multiple iterations, a better iterative separation and interpolation processing effect is obtained.

[0057] In the embodiments of the present application, an example of a lightweight network is used for illustration. For fair comparison, during the iterative training process, the number of iterations is set to 3, and each iteration is set to train for 4 epochs, that is, a total of three sets of network parameters need to be trained, and the total training epochs are the same as those in Technique 2.

[0058] Figure 2 Figure for comparing the training effects of two different training strategies, where Figure 2 (a) is the loss function graph, Figure 2 (b) is the signal-to-noise ratio graph. To avoid the uncertainty brought by random initialization, we set a random seed to ensure that the initial values of the networks with two different training strategies are the same. From Figure 2 it can be clearly seen that the convergence situations of the two training strategies are the same from 1 to 4 epochs, and at this time the dotted line and the solid line completely overlap. When the epoch is 5, the iterative training framework starts the second round of iterative training. Since the network parameters will be randomly initialized again, this leads to an increase in the loss at the beginning, but the loss decreases very quickly later and drops lower than that of the traditional training method. The change situation of the signal-to-noise ratio is consistent with the loss function, and the signal-to-noise ratio of the iterative framework training is higher than that of the traditional training method. By the third round of iteration, the network based on the iterative training framework converges better, which fully demonstrates the significant advantages of the present invention.

[0059] The synthetic seismic record and the network model are used to observe the noise separation and interpolation synchronization processing effects of the deep convolutional neural network based on iterative training. And for fair comparison, we substitute the parameters of the traditional deep neural network into the iterative formula to conduct a comparative evaluation experiment. As Figure 3 shown, Figure 3 (a) is the processing result of the traditional deep convolutional neural network after three iterations, and its finally restored signal-to-noise ratio is 14.80 dB, Figure 3 (b) is the corresponding residual of its processing result. Figure 4 (a) is the synchronization processing result of the deep convolutional neural network using the iterative training framework, and its signal-to-noise ratio is 17.46 dB, Figure 4 (b) is the corresponding residual. From Figure 3 and Figure 4 it can be seen that the deep convolutional neural network based on the iterative training framework has better de-aliasing and interpolation reconstruction effects.

[0060] The present invention mainly solves the problems of efficiently and high-quality separating aliased noise and interpolating and reconstructing in irregularly aliased seismic data, and proposes an iterative training framework based on a deep neural network that can adapt to the noise and interpolation iteration rules. At present, the multi-source efficient acquisition technology is a hot topic in seismic data acquisition. Accurately separating aliased noise and missing trace interpolation can lay a solid foundation for subsequent seismic data processing and prestack inversion:

[0061] (1) High - efficiency and high - quality synchronous processing of aliasing noise and interpolation for seismic data processing provides support for subsequent seismic data interpretation. Irregular aliased seismic records are relatively complex due to the inclusion of aliasing noise and missing seismic traces. During seismic data interpretation, the residue of aliasing noise or inaccurate prediction of seismic traces causes inconvenience to the correct understanding of underground geological laws. Therefore, in seismic data processing, accurately and efficiently separating aliasing noise and reconstructing seismic traces can enable us to more clearly understand underground geological information.

[0062] (2) As an efficient implementation method of "two - wide and one - high" data acquisition, the aliasing noise and irregularity of the multi - source technology are exacerbated by complex structures and underground medium conditions. Therefore, accurate and efficient separation of aliasing noise and interpolation reconstruction provides guarantee for the accuracy of seismic wave imaging.

[0063] The above description shows and describes a preferred embodiment of the present invention. However, as previously mentioned, it should be understood that the present invention is not limited to the form disclosed herein, should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be changed within the scope of the inventive concept described herein through the above teachings or the technology or knowledge in related fields. And any changes and modifications made by those skilled in the art that do not depart from the spirit and scope of the present invention should fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for separating and interpolating and synchronizing multi-source data based on iterative training, characterized in that: It includes the following steps: S1. Construct a neural network model; S2. Conduct iterative training on the neural network model; The step S2 includes: S201. If X represents the neural network input and Y represents the target output of the neural network, the loss function is defined as: Among them represents the square of the L2 norm; represents the network's prediction of the input X. When the prediction result is closer to Y, the loss is smaller. The neural network gradually updates the network parameters θ by minimizing the loss function; By minimizing the loss function, iterative training is performed on the parameter θ, that is, updating the parameter θ to the parameter that minimizes L(θ); Then use the updated estimate of θ The estimation result is: represents the approximate value close to M predicted by the neural network, where M represents the unaliased seismic records of all seismic sources; when n seismic sources are excited "synchronously", then where d k represents the conventional seismic data of the k-th seismic source, d ble represents the aliased data of the missing trace gather collected in the field, represents the time-delay coding inverse operator of the k-th seismic source; S202. When the features or distributions of the training data and the test data differ significantly, the effective signals will be damaged during the processing of the test data; however, when separating the aliased noise, not only the incoherent signals need to be suppressed, but also the effective signals need to be retained as much as possible. Therefore, the parameters obtained by the i-th iterative update are denoted as: where \(i\in(1,2,3,\cdots,m)\) represents the \(i\)-th iteration. At the first iteration, then it is equivalent to this situation; at the next iteration, the input data becomes where the incoherent aliasing noise is estimated by the network and subtracted from the mixed record; after multiple iterations, it gradually converges to the clean seismic record \(M\); where \(m\) k represents the sampling matrix of the \(k\)-th source, which controls the data missing situation. Its element values are 0 and 1, where 0 corresponds to the missing trace gather and 1 corresponds to the acquired trace gather; S203. When the input data changes and the same set of network parameters are used to estimate the aliasing noise and interpolation, after multiple iterations, the signal-to-noise ratio of the prediction result will decrease. To adapt to this change in the input, the iterative update of is applied to network training, and the final loss function is as follows Among them, θ i-1 represents the network parameters obtained from the (i - 1)-th training, and θ i represents the network parameters obtained from the i-th training of the network. When i = 1, at this time, the calculation method of L(θ i ) is the same as that of L(θ) in step S201, that is S3. Use the trained neural network model to predict the input, and perform iterative update of the neural network, and accordingly achieve multi-source data separation and interpolation synchronization processing.

2. The multi-source data separation and interpolation synchronization processing method based on iterative training according to claim 1, wherein: The neural network parameters included in the neural network model constructed in step S1 are Θ.

3. A method for multi-source data separation and interpolation synchronization processing based on iterative training according to claim 1, characterized in that: The step S3 includes: S301. In each iteration, the neural network parameters are trained by minimizing the loss function L(θ i ), that is, updating θ i-1 to the θ i that minimizes L(θ i ); S302. After each iteration is completed, substitute the iteration result into to obtain the result after removing aliasing noise and interpolation; S303. After each iteration, the input data will be closer to the target data. At this time, the network will converge quickly, and a higher signal-to-noise ratio and lower loss will be obtained. After multiple iterations, a better iterative separation and interpolation processing effect is obtained.

Images