Marine seismic exploration data ghost wave suppression method based on machine learning

Abstract

The invention relates to the technical field of oil and gas geophysical exploration, in particular to an offshore seismic exploration data ghost wave suppression method based on machine learning. According to the marine seismic exploration data ghost wave suppression method based on machine learning, a deepened towrope training set only needs to be combined from conventional horizontal towrope data, and various seismic data containing ghost waves can be realized without too much extra calculation cost; a deepened towing cable data set is used as training data of a network, so that a more real offshore seismic data acquisition environment can be reflected, and ghost wave characteristics in actual towing cable seismic data can be simulated more accurately; the network model obtained through training can well achieve ghost wave intelligent suppression and is successfully applied to actual marine towing cable data, and the resolution ratio and the frequency band width of the marine towing cable data are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas geophysical exploration, and more specifically, to a method for suppressing peg-leg multiples in marine seismic exploration data based on machine learning. Background Art

[0002] In order to improve the signal-to-noise ratio of seismic data, in the acquisition of towed streamer seismic data in marine seismic exploration, air guns and hydrophones are usually placed below the sea level to excite and receive seismic wave fields. Since the sea level is in contact with air and there is a large difference in wave impedance between seawater and air, with its reflection coefficient approaching -1, peg-leg multiples, i.e., waves formed by reflection from the sea level, will be generated in marine towed streamer data. Peg-leg multiples will be formed on the seismic record at both the excitation and reception points. The peg-leg multiples generated by the reflection of the up-going wave at the excitation point from the sea level are called shot peg-leg multiples, while the peg-leg multiples formed after reflection from the sea level at the reception point are called receiver peg-leg multiples. Therefore, there are three types of peg-leg multiples in marine towed streamer data: shot peg-leg multiples, receiver peg-leg multiples, and shot-receiver peg-leg multiples. Since the travel time of peg-leg multiples is not much different from that of effective reflected waves and they usually overlap behind the effective reflected waves, the existence of peg-leg multiples will greatly reduce the resolution of seismic data. A very important task in the processing of marine towed streamer seismic data is to suppress peg-leg multiples and improve the resolution of seismic data.

[0003] The suppression of peg-leg multiples mainly involves two approaches: acquisition and processing. In acquisition methods, improvements are mainly made in terms of receivers, the arrangement of source arrays, or receiver types. In 1989, Bearnth et al. proposed an acquisition method using slant cables to reduce the impact of peg-leg multiples. In a slant cable, the receiver depth varies linearly with the offset, resulting in different notch positions for each trace, thereby increasing the bandwidth of post-stack data. In 2007, Moldoveanu et al. proposed a method using upper and lower dual cables for reception, suppressing peg-leg multiples by recording the up-going and down-going wavefields. However, the upper and lower dual cables require twice the number of receivers. In 2010, Kragh et al. proposed a shallow and deep dual cable acquisition method, where the shallow cable is used for conventional acquisition and the deep cable uses a sparse array to record the low-frequency information lost by the shallow cable, thus achieving the purpose of restoring the bandwidth. With the continuous progress of acquisition equipment technology, in 2007, Carlson et al. combined the idea of dual cable acquisition, applied a dual-sensor cable to simultaneously measure the pressure component and particle velocity, and used the opposite polarity of peg-leg multiples to suppress them. At the same time, there are also methods for suppressing peg-leg multiples in single-component data (conventional horizontal cables and slant cables) through pure processing means. Most of these methods are based on wave theory or use signal processing techniques. In wave theory methods, they are mainly based on wavefield separation driven by Green's function, which has been well described and proven in a series of papers, such as Zhang and Weglein (2005), Mayhan and Weglein (2013). Additionally, there are filtering methods based on signal processing, where the inverse function of the peg-leg multiple function is constructed as a de-ghosting filter, and the filter is usually represented in the intercept-slowness domain. This method is sensitive to the signal-to-noise ratio, but it can usually provide effective de-ghosting under some assumptions, such as assumptions about the characteristics of the sea surface and water column (Wang et al., 2016). At the same time, Amundsen and Zhou (2013) derived a low-frequency de-ghosting method, which is a filter-based de-ghosting method aimed at improving the low frequency by filling the first ghost notch near 0 Hz. In 2017, Wang et al. improved this method, which can be removed trace by trace in the spatio-temporal domain, making it adaptable to towed cable data with mild depth variations. However, traditional methods still have many problems. Acquisition means require high-cost acquisition systems, and processing means are affected by engineers' experience and computing speed, and the suppression process is relatively cumbersome. Summary of the Invention

[0004] The object of the present invention is to overcome the deficiencies of the relatively cumbersome, low-precision, and low-efficiency methods for suppressing peg-leg multiples in the prior art, and to provide a method for suppressing peg-leg multiples in marine seismic exploration data based on machine learning, which can more accurately simulate the peg-leg multiple characteristics in actual towed cable seismic data, has a high suppression ability, and also improves the processing efficiency.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A marine seismic exploration data ghost wave suppression method based on machine learning is provided, including the following steps:

[0007] S1. Collect a ghost wave-containing seismic data set and a ghost-free label: First, through wave equation forward modeling, horizontal streamer acquisition at different depths is carried out; parameters are adjusted to produce a large number of streamer curves with different forms; the horizontal streamer data are combined using the streamer curves to obtain ghost wave-containing seismic data as the training data set; the air gun source and the streamer are placed at a seawater depth of 0 to produce the corresponding ghost-free label;

[0008] S2. Input the training data set into a neural network model for training, output ghost-free data, calculate the loss by taking the difference from the ghost-free label, and update the parameters of the neural network model until the model converges; obtain a neural network model for suppressing ghost waves;

[0009] S3. Input the to-be-tested marine seismic exploration data into the neural network model for suppressing ghost waves to obtain ghost-free data.

[0010] For the marine seismic exploration data ghost wave suppression method based on machine learning provided by the present invention, the variable-depth streamer training set only needs to be combined from conventional horizontal streamer data, and diverse ghost wave-containing seismic data can be achieved without excessive additional computational costs; by using the variable-depth streamer data set as the training data of the network, it can reflect a more real marine seismic data acquisition environment and can more accurately simulate the ghost wave characteristics in actual streamer seismic data; the network model trained by the present invention can well achieve intelligent ghost wave suppression and be successfully applied to actual marine streamer data, effectively improving the resolution and bandwidth of marine streamer data.

[0011] Preferably, in step S1, while keeping the source depth unchanged, a streamer depth curve equation is defined to simulate the morphological changes of the streamer, and the depth expression of the i-th geophone is:

[0012] H(i) = A 1 sin(w 1 x(i) + b 1 ) + A 2 cos(w 2 x(i) + b 2 )(1)

[0013] In the formula: H(i) is the depth of the i-th geophone in the same shot gather. By defining the values of parameters A, w, and b, the morphology of the streamer is changed. By |A 1 + A 2 | and |A 1 - A 2 |, the maximum depth and minimum depth of the streamer are controlled, w 1 、w2 is the undulation frequency of the towing cable controlled by the angular frequency, b 1 , b 2 is the curve phase information.

[0014] Preferably, in step S1, in order to control the computational cost, to achieve diverse ghost wave acquisitions at the same location, a variable-depth towing cable acquisition is implemented using a combination method of conventional horizontal towing cables at different depths. First, acquisitions are performed using horizontal towing cables at multiple depths, with the speed range being m, 6m ≤ m ≤ 14m, to obtain the data set D(m, :), and through the combination of the depth curve forms, the variable-depth towing cable D var is obtained, where the expression of the i-th trace is:

[0015] D var (i) = D(Round[H(i)], i)(2)

[0016] In the formula, Round() is the rounding operation. In this way, a large number of ghost wave-containing data can be produced by changing different curve forms for the shot gather data at the same location, avoiding the need to perform a wave equation calculation for each towing cable form. This scheme can effectively reduce the computational cost while improving the richness of the ghost wave-containing data.

[0017] Preferably, step S1 specifically includes the following steps:

[0018] S11. Based on two Marmousi sub-models, first, forward modeling is performed using the wave equation for conventional horizontal towing cable acquisitions at different depths, with the depth being 6m to 14m and the depth interval being 1m;

[0019] S12. Adjust the parameters A 1 , A 2 , w 1 , w 2 , b 1 , b 2 to produce a large number of towing cable curves with different forms;

[0020] S13. Apply formula (2) to combine the horizontal towing cable data to obtain variable-depth towing cable data, which is used as the ghost wave-containing seismic data in the training set. At the same time, the air gun source and the towing cable are placed at a seawater depth of 0 to produce the corresponding ghost-free label data.

[0021] Machine learning has shown great advantages in the field of extracting image features. The network can obtain the effective wave features through an end-to-end learning process, thereby achieving the purpose of suppression.

[0022] Preferably, in step S2, the network output can be expressed by the following formula:

[0023] sout f(t) = f θ (ε, s(t)) (3)

[0024] where f is the network model, ε is the editable parameter in the network, θ is the weight parameter for training, and s(t) is the seismic data containing peg-leg multiples;

[0025] The training process of the model is based on the seismic data containing peg-leg multiples and the data without peg-leg multiples labels made, and the optimal weight θ is obtained by learning the mapping relationship through minimizing the loss function * , finally making the network output s out (t) approximate the real data without peg-leg multiples:

[0026]

[0027] And the quality of the entire suppression process depends on the network and the production of training data. A high-quality network has a stronger approximation ability, and the mapping relationship of the training set needs to meet the requirements of the actual problem.

[0028] The network model built in the present invention is the AttUNet model. The model encoder consists of four consecutive convolutional layers, which respectively process the input structure data of five different spatial scales related to the downsampling rates of 2, 4, 8, 16, and 32. The number of channels changes from 1 at the input to 32, 64, 128, 512, and 1024 in turn; each convolutional layer consists of two groups of convolutional operations (3*3) + BN + ReLU, followed by a max pooling layer (2*2) for downsampling. The decoder part includes five spatial scales consistent with the encoder, and integrates the upsampling and the skip connection from the encoder through the attention mechanism module. The advantages of the model are that there are enough trainable parameters, and the addition of the attention mechanism ignores the irrelevant image features and improves the model convergence speed.

[0029] Preferably, the loss function is defined as:

[0030] Loss = ||f θ (ε, s(t)) - s l (t)|| 1

[0031] where s l (t) is the label corresponding to s(t).

[0032] Preferably, the training data is normalized to the range of -1 to 1, and the normalization is achieved by dividing by the maximum value of the absolute value of the gather data; after normalizing the data, the profile is split into pictures of size 512*512, and the batch size is set to 40; within each epoch, the loss function is calculated; the adam optimizer and the adaptive learning step size are used to accelerate the network optimization, and finally the trained network model for suppressing peg-leg multiples is obtained.

[0033] The present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned method are implemented.

[0034] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the above-mentioned method are implemented.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] For the ghost wave suppression method of marine seismic exploration data based on machine learning of the present invention, the variable-depth streamer training set only needs to be combined from conventional horizontal streamer data, and diverse ghost wave-containing seismic data can be achieved without excessive additional computational costs; by using the variable-depth streamer data set as the training data of the network, the more real marine seismic data acquisition environment can be reflected, and the ghost wave characteristics in the actual streamer seismic data can be more accurately simulated; the network model trained by the present invention can well achieve intelligent ghost wave suppression and is successfully applied to actual marine streamer data, effectively improving the resolution and bandwidth of marine streamer data. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic flowchart of a ghost wave suppression method for marine seismic exploration data based on machine learning.

[0038] Figure 2 It is a schematic flowchart of the training process of the network model of the present invention.

[0039] Figure 3 It is a schematic diagram of the AttUNet network structure of the present invention.

[0040] Figure 4 It is a training data set model based on the Marmousi sub-model in Embodiment 1.

[0041] Figure 5 It is a test data set model based on the Marmousi sub-model in Embodiment 1.

[0042] Figures 6 to 8 It is a production process of the training set data; wherein, Figure 6 represents a schematic diagram of conventional horizontal streamer data; Figure 7 represents the streamer curve shape and the corresponding schematic diagram of the streamer data; Figure 8 represents a schematic diagram of ghost wave-free data.

[0043] Figure 9 It is a schematic diagram of the test result using the test data set in Embodiment 1.

[0044] Figure 10 Schematic diagram of the original input ghost wave-containing data in Example 1

[0045] Figure 11 Schematic diagram of the network output ghost wave-free data in Example 1

[0046] Figure 12 Schematic diagram of the comparison of amplitude spectra before and after suppressing the actual data in Example 1 Detailed implementation manners

[0047] The present invention will be further described below in conjunction with the detailed implementation manners. Among them, the attached drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to this patent; in order to better illustrate the embodiments of the present invention, some components in the attached drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the attached drawings may be omitted

[0048] In the attached drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the attached drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the attached drawings are only for illustrative purposes and should not be construed as a limitation to this patent. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances

[0049] Example 1

[0050] As Figure 1 shown, this embodiment is the first embodiment of a method for suppressing ghost waves in marine seismic exploration data based on machine learning, including the following steps:

[0051] S1. Collect the seismic data set containing ghost waves and the ghost wave-free label: First, through wave equation forward modeling, horizontal streamer acquisitions at different depths are carried out; parameters are adjusted to produce a large number of streamer curves with different forms; the horizontal streamer data are combined using the streamer curves to obtain the seismic data containing ghost waves as the training data set; the air gun source and the streamer are placed at the position where the seawater depth is 0 to produce the corresponding ghost wave-free label

[0052] S2. Input the training data set into the neural network model for training, output the ghost wave-free data, calculate the loss by taking the difference with the ghost wave-free label, and update the parameters of the neural network model until the model converges; obtain the neural network model for suppressing ghost waves

[0053] S3. Input the offshore seismic exploration data to be measured into the neural network model for suppressing ghost waves to obtain ghost-wave-free data.

[0054] In step S1, keeping the source depth unchanged, define a cable depth curve equation to simulate the morphological changes of the cable. The depth expression of the i-th geophone is:

[0055] H(i) = A 1 sin(w 1 x(i) + b 1 ) + A 2 cos(w 2 x(i) + b 2 )(1)

[0056] Where: H(i) is the depth of the i-th geophone in the same shot gather. By defining the values of parameters A, w, and b, the morphology of the cable can be changed. The maximum and minimum depths of the cable are controlled by |A 1 + A 2 | and |A 1 - A 2 |. w 1 , w 2 are angular frequencies that control the undulation frequency of the cable morphology, and b 1 , b 2 are curve phase information.

[0057] In step S1, to control the calculation cost and achieve diverse ghost wave acquisitions at the same location, a variable-depth cable acquisition is implemented using a combination method of conventional horizontal cables at different depths. First, collect data through horizontal cables at multiple depths. The speed range is m, 6m ≤ m ≤ 14m, to obtain the data set D(m, :), and then combine them through the cable depth curve morphology to obtain the variable-depth cable D var , where the expression of the i-th trace is:

[0058] D var (i) = D(Round[H(i)], i)(2)

[0059] Where Round() is the rounding operation. In this way, a large number of ghost-wave-containing data can be produced by changing different curve morphologies for the shot gather data at the same location, avoiding the need to perform a wave equation calculation for each cable morphology. This scheme can effectively reduce the calculation cost while improving the richness of ghost-wave-containing data.

[0060] The specific steps of step S1 include the following steps:

[0061] S11. Based on two Marmousi sub-models, first, through forward modeling of the wave equation, conventional horizontal streamer acquisitions at different depths are carried out, with the depth ranging from 6 m to 14 m and the depth interval being 1 m;

[0062] S12. Adjust the parameter A through formula (1) 1 , A 2 , w 1 , w 2 , b 1 , b 2 to produce a large number of streamer curves with different shapes;

[0063] S13. Apply formula (2) to combine the horizontal streamer data to obtain variable-depth streamer data, which is used as the ghost-wave-containing seismic data in the training set. At the same time, the air gun source and the streamer are placed at the position where the seawater depth is 0 to produce the corresponding ghost-free label data.

[0064] Machine learning has shown great advantages in the field of extracting image features. The network can obtain effective wave features through an end-to-end learning process, so as to achieve the purpose of suppression.

[0065] In step S2, the network output can be expressed by the following formula:

[0066] s out (t) = f θ (ε, s(t)) (3)

[0067] where f is the network model, ε is the editable parameter in the network, θ is the training weight parameter, and s(t) is the ghost-wave-containing seismic data;

[0068] The training process of the model is based on the produced ghost-wave-containing seismic data and ghost-free label data. By minimizing the loss function, the mapping relationship is learned to obtain the optimal weight θ * , and finally make the network output s out (t) approximate the real ghost-free data:

[0069]

[0070] And the quality of the entire suppression process depends on the network and the production of training data. The high-quality network has a stronger approximation ability, and the mapping relationship of the training set needs to meet the actual problem requirements.

[0071] The network model established in this invention is the AttUNet model. The model encoder consists of four consecutive convolutional layers, which respectively process the input structural data of five different spatial scales related to downsampling rates of 2, 4, 8, 16, and 32. The number of channels changes from 1 input to 32, 64, 128, 512, and 1024 in sequence. Each convolutional layer consists of two sets of convolutional operations (3*3)+BN+ReLU, followed by a max-pooling layer (2*2) for downsampling. The decoder part includes five spatial scales consistent with the encoder, and integrates upsampling and skip connections from the encoder through an attention mechanism module. The advantages of the model are that there are enough trainable parameters, and the addition of the attention mechanism ignores irrelevant image features and improves the model convergence speed.

[0072] Considering the gap in data size between actual data for better application to the network, the training data is normalized to the range of -1 to 1, which is achieved by dividing by the maximum absolute value of the shot gather data. After normalizing the data, the profile is split into images of size 512*512, and the batch size is set to 40. In each epoch, the loss function is calculated. The loss function is defined as:

[0073] Loss=||f θ (ε,s(t))-s l (t)|| 1

[0074] where s l (t) is the label corresponding to s(t).

[0075] The adam optimizer and adaptive learning step size are used to accelerate network optimization. The learning rate is set to 0.001, and it is trained for 200 epochs to finally obtain a trained network model for suppressing ghost waves.

[0076] Figure 2 It shows the entire data production and training process of the invention. ① is the production process of the variable-depth training set. Based on the conventional horizontal streamer forward modeling, the diversity of ghost waves is determined by the generated streamer curves, without the need for additional wave equation forward modeling. In the case of being able to restore the offshore streamer morphology, it greatly saves computational costs. ② is the training process of the invention. The model is AttUNet. Compared with the conventional UNet model, the addition of the attention mechanism module greatly improves the network ability.

[0077] Figure 3 Based on the UNet network architecture, skip connections are connected through the attention mechanism. The attention mechanism enables the neural network to ignore unimportant feature vectors and focus on calculating useful vectors. While abandoning the interference of useless features on the fitting result, it also improves the operation speed.

[0078] Figures 6 to 8 shows the data of the conventional horizontal streamer, variable-depth streamer, and 0-depth streamer. The production process of the variable-depth streamer is to first forward model the horizontal streamer data at different depths, as shown in (6), and then arrange and combine them through the defined depth curve ( Figure 7 the black curve in it) to obtain the corresponding variable-depth streamer data. The quantity and richness of the data can be completed by generating different curves without excessive forward modeling calculations; Figure 8 is the output label of the network, which is obtained by forward modeling with the air gun source and the streamer depth both set to 0.

[0079] Before applying the present invention to actual data, it is first tested in the synthetic test set data. Figure 9 a in it is an input profile in the test data. It can be seen that as the streamer changes, the resolution of different traces is also different, and the peg-leg multiple interference is serious. Figure 9 b in it is used as the suppression output of the network, which is similar to the data without peg-leg multiples ( Figure 9 c in it). The 450th and 500th trace records are extracted from it for analysis, as shown in Figure 9 d and e in it. The resolution is greatly improved; the suppression result has a high degree of coincidence with the morphology of the data without peg-leg multiples.

[0080] A total of 600 shot gather data are processed in the actual application. Figure 5 is the common offset gather profile extracted by us. It can be seen from the original data ( Figure 10 ) that due to the existence of peg-leg multiples, the event axis is thicker and the resolution is low. Especially in the area where the formation is thinner, the accurate reflection interface cannot be distinguished. After peg-leg multiple suppression ( Figure 11 ), the resolution is significantly improved; as shown in Figure 12 , it is the amplitude spectrum comparison of the actual data. The bandwidth of the suppressed data is effectively improved, and it is effectively restored at the 0 Hz notch and 90 Hz notch respectively.

[0081] For the method for suppressing peg-leg multiples in marine seismic exploration data based on machine learning in this embodiment, the variable-depth streamer training set only needs to be combined from the conventional horizontal streamer data, and diverse peg-leg multiple-containing seismic data can be achieved without excessive additional computational costs; by using the variable-depth streamer data set as the training data of the network, it can reflect the more real marine seismic data acquisition environment and can more accurately simulate the peg-leg multiple characteristics in the actual streamer seismic data; the network model trained by the present invention can well achieve intelligent peg-leg multiple suppression and is successfully applied to the actual marine streamer data, effectively improving the resolution and bandwidth of the marine streamer data.

[0082] Embodiment 2

[0083] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the method described in Embodiment 1 are implemented.

[0084] Embodiment Three

[0085] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the method described in Embodiment 1 are implemented.

[0086] In the specific content of the above specific implementation manner, each technical feature can be combined arbitrarily without contradiction. For the sake of concise description, not all possible combinations of the above technical features are described. However, as long as the combinations of these technical features do not exist in contradiction, they should be considered as the scope described in this specification.

[0087] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, rather than limitations on the implementation manners of the present invention. For those of ordinary skill in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to enumerate all the implementation manners here. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for suppressing ghost waves in offshore seismic exploration data based on machine learning, characterized in that: The following steps are involved: S1. Collecting seismic data sets containing ghost waves and ghost-free labels: First, perform horizontal streamer acquisition at different depths through wave equation forward modeling; adjust parameters to produce a large number of streamer curves with different shapes; combine horizontal streamer data using streamer curves to obtain seismic data containing ghost waves as training data sets; place airgun seismic sources and streamers at a seawater depth of 0 to produce corresponding ghost-free labels; S2. Input the training data set into the neural network model for training, output the ghost-free data, and calculate the loss by difference with the ghost-free label, and update the parameters of the neural network model until the model converges; obtain the neural network model for suppressing ghost waves; S3. Input the offshore seismic exploration data to be tested into the neural network model for suppressing ghost waves to obtain ghost-free data.

2. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 1, characterized in that: In step S1, the focal depth is kept constant, and a streamer depth curve equation is defined to simulate the morphological changes of the streamer, where the depth expression of the i-th detector is: H(i)=A1sin(w1x(i)+b1)+A2cos(w2x(i)+b2)(1) Where: H(i) is the depth of the i-th detector in the same shot set. The shape of the streamer is changed by defining the values ​​of A, w and b parameters. The maximum and minimum depths of the streamer are controlled by |A1+A2| and |A1-A2|. w1 and w2 are the angular frequencies that control the fluctuation frequency of the streamer shape. b1 and b2 are the curve phase information.

3. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 2, characterized in that: In step S1, the data is collected by horizontal streamers at multiple depths, with a speed range of m, 6m≤m≤14m, to obtain a data set D(m,:), and then combined with the depth curve to obtain a depth streamer D var , where the i-th formula is expressed as: D var (i)=D(Round[H(i)],i)(2) In the formula, Round() is a rounding operation.

4. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 3 is characterized in that: The step S1 specifically includes the following steps: S11. Based on the two Marmousi sub-models, conventional horizontal streamer acquisition at different depths is first performed by forward modeling of the wave equation, ranging from 6m to 14m with a depth interval of 1m; S12. By using formula (1), adjust the parameters A1, A2, w1, w2, b1, b2 to produce a large number of towline curves of different shapes; S13. Apply formula (2) to combine the horizontal streamer data to obtain variable depth streamer data, which are used as the ghost wave seismic data in the training set. At the same time, the airgun seismic source and the streamer are placed at a position where the sea water depth is 0 to produce corresponding ghost wave label data.

5. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 1, characterized in that: In step S2, the network output can be expressed by the following formula: s out (t)=f θ (ε,s(t))(3) Where f is the network model, ε is the editable parameter in the network, θ is the training weight parameter, and s(t) is the seismic data containing ghost waves; The training process of the model is based on the seismic data with ghost waves and the label data without ghost waves. The mapping relationship is learned by minimizing the loss function to obtain the optimal weight θ * , and finally the network outputs s out (t) Approximation to real ghost-free wave data:

6. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 5, characterized in that: The network model is the AttUNet model. The model encoder consists of four consecutive convolutional layers, which process input structure data of five different spatial scales related to the downsampling rates of 2, 4, 8, 16 and 32 respectively. The number of channels changes from input 1 to 32, 64, 128, 512 and 1024 respectively; each convolutional layer consists of two sets of convolution operations (3*3)+BN+ReLU, followed by a maximum pooling layer (2*2) for downsampling. The decoder part includes five spatial scales consistent with the encoder, and integrates upsampling with jump links from the encoder through the attention mechanism module.

7. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 5, characterized in that: The loss function is defined as: Loss=||f θ (ε,s(t))-s l (t)||1 where s l (t) is the label corresponding to s(t).

8. The method for suppressing ghost waves in offshore seismic exploration data based on machine learning according to claim 7, characterized in that: The training data is normalized in the range of -1 to 1 by dividing by the maximum absolute value of the shot data. After the data is normalized, the profile is split into 512*512 images, and the batch size is set to 40. In each cycle, the loss function is calculated. The adam optimizer and adaptive learning step size are used to accelerate network optimization, and finally a trained network model for suppressing ghost waves is obtained.

9. A computer device comprising a memory and a processor, characterized in that: The memory stores a computer program, and the processor implements the steps of the method described in any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 8 are implemented.

Images