Denoising convolutional neural network local strong noise suppression method based on passive data

Abstract

The invention provides a passive data-based denoising convolutional neural network local strong noise suppression method. The method comprises the steps of analyzing time-space-frequency distribution characteristics of signals and noise in full-node seismic data signals; performing active and passive data separation, and dividing local strong noise; constructing a data sample conforming to real noise features; performing data expansion and manual screening; constructing an effective signal sample; building a denoising neural network; performing neural network optimization training; denoising evaluation is carried out, and network model parameters are stored; and carrying out local strong noise identification to obtain data with a high signal-to-noise ratio. According to the passive data-based denoising convolutional neural network local strong noise suppression method, effective signals of active data can be more effectively protected, the signal-to-noise ratio of seismic data is improved, intelligent recognition and suppression of local strong noise are realized, and the application capability of passive data in full-node seismic data processing is verified.

Description

Technical Field

[0001] The invention relates to the technical field of oilfield development, and in particular to a method for suppressing local strong noise of a denoising convolutional neural network based on passive data. Background Art

[0002] Seismic nodes can collect data in all time periods and spaces. In addition to the commonly used reflection wave information, there are also a large amount of passive data in the collected data. How to "turn waste into treasure" and make full use of the signal to obtain near-surface structure information has become a powerful auxiliary in seismic exploration, which has become the key to reflecting the advantages of full-node acquisition. When performing full-node signal acquisition, the source and the detector are randomly distributed, and the signal acquisition time is long. Compared with conventional cable acquisition, a large amount of passive data is also recorded. The characteristics of passive data are different from those of conventional passive data. How to "turn waste into treasure" and make full use of the signal has become the key to reflecting the advantages of full-node acquisition. The seismic exploration noise is not a stationary random process in the traditional sense. Its stability is affected by the duration of the noise and the complexity of the acquisition environment. Zhang et al. (2017) proposed to use denoising convolutional neural networks for image denoising. They used a convolutional neural network with 17 convolutional layers, abandoned the traditional way of learning noise-free data, and used noise as output, that is, residual learning, which can speed up the training process and improve the noise reduction performance. Yu et al. (2019) used CNN to automatically suppress noise in various seismic data and achieved good random noise denoising effects; Saad et al. (2020) proposed a deep denoising auto-encoding network for denoising, encoding seismic data into multiple abstract domains for learning, effectively suppressing random noise. Traditional denoising algorithms cannot effectively remove non-stationary noise or damage effective signals, which affects the further processing of passive data. It is necessary to study the identification and elimination algorithms of non-stationary noise.

[0003] Denoising methods based on the characteristic differences between signals and noise in different transform domains can usually effectively suppress noise with different signal distribution characteristics, such as random noise and surface wave noise in seismic data. When the difference between signal and noise is not obvious enough, such as strong background noise interference close to the signal frequency band, this method often cannot effectively separate them.

[0004] In the Chinese patent application with application number: CN202211250180.X, a seismic denoising method based on convolutional neural network and visual transformation neural network is involved, which relates to the field of information processing technology. The invention collects noisy seismic data, denoises the noisy seismic data, obtains multiple denoised seismic data blocks, constructs a seismic data set including a training set, a validation set and a test set, and constructs a seismic data denoising network based on convolutional neural network and visual transformation neural network. After the seismic data denoising network is trained with the training set, the denoising effect of the trained seismic data denoising network is debugged with the validation set. After the seismic data denoising network with the best denoising effect is obtained, the trained seismic data denoising network is used to remove the noise of the noisy seismic data blocks in the test set to obtain the seismic data after the noise is removed. The invention effectively improves the denoising performance of seismic data, reduces the training cost, and is conducive to the rapid and accurate acquisition of underground medium structural information.

[0005] In the Chinese patent application with application number: CN202210097478.5, a method for suppressing random noise in 3D seismic data based on self-supervision of 3D convolutional neural network is involved, which includes: 1. Designing a 3D convolutional neural network architecture; 2. Preparing 3D seismic data containing random noise; 3. Designing self-supervision auxiliary tasks; 4. Designing strategies to avoid overfitting; 5. Running the 3D convolutional neural network; 6. Determining the optimal noise suppression result. Compared with the existing deep learning-based methods, this method solves the problem that the diversity of samples in the data set or the quality of labels limits the denoising and generalization capabilities of the neural network through a new 3D convolutional neural network assisted by a self-supervision strategy; considering the spatial correlation of effective signals in 3D seismic data, 3D convolution is used to overcome the limitations of 2D convolution in extracting 3D spatial structural features, improve the ability of the neural network to protect effective signals, and improve the signal-to-noise ratio of 3D seismic data.

[0006] In the Chinese patent application with application number: CN202210523847.2, a method for denoising seismic data based on convolutional neural networks is involved. First, an initial data set is constructed, and a cyclic generative adversarial convolutional neural network is constructed. The cyclic generative adversarial convolutional neural network uses the structure of a cyclic generative adversarial network as the main network. The structure of the cyclic generative adversarial network includes a generator and a discriminator. The non-local neural network is connected as a residual to the convolutional layers of the generator, and the PatchGAN output method is used in the discriminator instead of the original fully connected method. Then the ring-generated adversarial convolutional neural network is trained, and finally the noisy seismic data is denoised. Experiments have verified that the denoising effect of the invented method is good.

[0007] In the Chinese patent application with application number: CN201810886433.X, a method for suppressing random noise in seismic data based on intelligent deep learning is involved, including: generating multiple velocity models for forward modeling of seismic data; using the acoustic wave equation to obtain the corresponding seismic single shot record as the training output label for each velocity model, and adding noise to form a training data input set; and using a convolutional neural network algorithm and combining the training data input set to construct an autoencoder training model, complete model training, and use it for random noise suppression of seismic data.

[0008] The above existing technologies are significantly different from the present invention and fail to solve the technical problem we want to solve. For this reason, we have invented a new local strong noise suppression method of denoising convolutional neural network based on passive data. Summary of the invention

[0009] The purpose of the present invention is to provide a denoising convolutional neural network local strong noise suppression method based on passive data, which can extract the characteristics of background strong noise in passive data collected in all time and space, identify the background strong noise, protect the effective signal of active data, and improve the signal-to-noise ratio of seismic data.

[0010] The purpose of the present invention can be achieved by the following technical measures: a local strong noise suppression method of a denoising convolutional neural network based on passive data, the local strong noise suppression method of a denoising convolutional neural network based on passive data comprises:

[0011] Step 1, analyzing the time-space-frequency distribution characteristics of signals and noise in the full-node seismic data signal;

[0012] Step 2: Separate active and passive data to separate local strong noise;

[0013] Step 3, constructing a data sample that meets the real noise characteristics;

[0014] Step 4: data expansion and manual screening;

[0015] Step 5, constructing a valid signal sample;

[0016] Step 6: Build a denoising neural network;

[0017] Step 7, perform neural network optimization training;

[0018] Step 8, perform denoising evaluation and save network model parameters;

[0019] Step 9: perform local strong noise identification to obtain high signal-to-noise ratio data.

[0020] The purpose of the present invention can also be achieved by the following technical measures:

[0021] In step 1, the time-space-frequency distribution characteristics of the signal and noise in the full-node seismic data signal are first analyzed using signal analysis methods such as wavelet transform and FK transform.

[0022] In step 1, different types of noise are divided. The common random noise and surface wave noise in the full-node seismic data have time-frequency distribution characteristics that are significantly different from the effective signal and can be suppressed using conventional denoising techniques. The target local strong noise interference includes burst pulses, abnormal disturbances, and acoustic wave interference, which have the characteristics of long duration, concentration in the 0-25Hz frequency band, and high overlap with the effective signal. Conventional denoising methods based on time-frequency distribution cannot effectively remove local strong noise.

[0023] In step 2, according to the spatiotemporal distribution characteristics of the local strong noise in the target work area, which usually recurs in different regions and lasts for a long time, the active and passive data of the full-node seismic data are divided, and then the pure target noise signal segment is distinguished from the passive data.

[0024] In step 3, a random sliding operator is set to automatically pick up local strong noise in the pure target noise signal segment to construct a noise sample, thereby preliminarily constructing a data sample that conforms to the real noise characteristics for subsequent data expansion and training.

[0025] In step 4, random cropping, scaling, and other transformation methods are used to expand the data to enrich the local strong noise signal library, and through manual screening, signals with obvious noise characteristics are used as pre-training libraries, and a part of them are selected as verification libraries.

[0026] In step 5, effective signal samples are constructed through wave field forward simulation and using some actual acquisition data with high signal-to-noise ratio.

[0027] In step 5, through the preliminary analysis of the actual full-node seismic data, a model that conforms to the velocity characteristics and observation system of the work area is constructed to perform wave field forward simulation. In addition, some actual acquisition data with high signal-to-noise ratio are used to pick up effective signals. The two are combined to construct effective signal samples, thereby completing the data sample construction process and using it for neural network training.

[0028] In step 6, combined with the residual learning idea, a denoising convolutional neural network model is designed for the target local strong noise. A residual module is set in the network for encoding and decoding to avoid the gradient vanishing problem of the deep network and improve the generalization ability of the model.

[0029] In step 6, a denoising convolutional neural network model is designed for the target local strong noise. The network combines the ideas of residual learning and residual module encoding and decoding, sets the network output to the local strong noise of passive data, i.e., residual data, to reduce the sensitivity of abnormal points, and sets the residual module for encoding and decoding to avoid the problem of vanishing gradients in deep networks and improve the generalization ability of the model.

[0030] In step 7, the databases constructed in steps 4 and 5 are combined for network denoising training. Data enhancement and learning rate decay strategies are added during the training process to improve the optimal convergence speed of the network training process and the global minimum search accuracy, and obtain a neural network that can effectively identify local strong noise.

[0031] In step 8, the noisy signal in the verification library is used as input to verify the denoising performance of the neural network after training convergence. The denoising performance of the network is detected by indicators such as time-frequency analysis and signal-to-noise ratio. If the denoising performance meets the requirements, the training is terminated and the network model parameters are saved; otherwise, the learning rate is adjusted and the network training in step 7 is performed again.

[0032] In step 8, when saving the network model parameters, the network training is complete, and a neural network model capable of identifying target noise in noisy data is obtained, and the network model parameters are saved for subsequent actual data denoising processing.

[0033] In step 9, the full-node seismic data is input into the neural network for intelligent identification of target strong noise and denoising. The network learns from the rich noise samples in the passive data and can improve the signal-to-noise ratio while protecting the effective signal during the processing.

[0034] The method for suppressing local strong noise using a denoising convolutional neural network based on passive data in the present invention utilizes the passive data characteristics of full-time and space acquisition technology activities to construct a denoising convolutional neural network for noise identification and suppression. The method first uses wavelet transform, FK transform and other methods to analyze the characteristics of non-stationary noise signals in passive data and distinguish different characteristic noises; then, according to the spatiotemporal distribution characteristics of the target local strong noise, a noise sample library is established through wave field simulation and actual data picking; then, a denoising convolutional neural network model is designed, and through noise identification training, a network that can effectively identify local strong noise is obtained, and finally passive data noise suppression is performed, and the actual data processed has a higher signal-to-noise ratio. The present invention takes into account the feature extraction of local strong noise in passive data, and by extracting data features of local strong noise in passive data, it can protect the effective signal of active data and improve the quality of seismic data. The technical key points of the present invention are:

[0035] (1) The present invention targets the full-time and full-space acquisition characteristics of full-node seismic data and utilizes the local strong noise of passive data signals to construct target data samples for noise feature recognition training. This method can suppress the local strong noise that cannot be effectively processed by conventional methods, verifies the application capability of passive data in full-node seismic data, and opens up new ideas for noise suppression processing of full-node seismic data.

[0036] (2) The neural network model of the present invention combines the residual learning objective function and the residual module encoding and decoding characteristics, which can give full play to the performance of deep network mining target noise characteristics, avoid overfitting and have good generalization; through data enhancement and learning rate decay strategy for training, while increasing the network depth without reducing the convergence performance, the network training speed is improved. After training, the network can effectively identify and suppress local strong noise that has frequency band aliasing and propagation speed and direction overlap with the effective signal.

[0037] Compared with the existing noise suppression technology, the present invention has two main advantages:

[0038] (1) To learn the noise characteristics for local strong noise, a full-node observation system is required to collect all-time and space seismic data for passive data noise feature extraction. Conventional methods mostly focus on random noise or surface wave noise collected by geophones, and cannot obtain a high-quality database of local strong noise that is aliased with effective signal features. Therefore, this method can effectively suppress local strong noise that cannot be handled by conventional methods.

[0039] (2) After training, the invention can quickly and accurately identify local strong noise intelligently and suppress the strong background noise of seismic data of all nodes; by training a part of passive data, the invention can effectively suppress the local strong noise in the entire work area. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flowchart of a specific embodiment of the method for suppressing local strong noise by a denoising convolutional neural network based on passive data of the present invention;

[0041] Figure 2 A schematic diagram of a Wiger-Ville distribution of local strong noise in a specific embodiment of the present invention;

[0042] Figure 3 It is a schematic diagram of the time-space-frequency distribution characteristics of local strong noise in all-node seismic data in a specific embodiment of the present invention;

[0043] Figure 4 It is a schematic diagram of a denoising residual convolutional neural network built for full-node seismic data in a specific embodiment of the present invention;

[0044] Figure 5A schematic diagram of a denoising sample training library constructed using passive coda wave data in a specific embodiment of the present invention;

[0045] Figure 6 A schematic diagram of obtaining target data features by combining feature matrices and using feature vectors to identify local strong noise during prediction in a specific embodiment of the present invention;

[0046] Figure 7 A schematic diagram of a decaying learning rate optimization and data enhancement strategy training process in a specific embodiment of the present invention;

[0047] Figure 8 A schematic diagram of a data verification result using a fully trained network model in a specific embodiment of the present invention;

[0048] Fig. 9 It is a schematic diagram of frequency domain analysis and denoising index evaluation of verification data before and after denoising in a specific embodiment of the present invention;

[0049] Fig.10 It is a schematic diagram of the noise suppression result of actual single shot seismic data in a specific embodiment of the present invention;

[0050] Fig.11 It is a schematic diagram of a single-channel comparison of the denoised data with and without background strong noise interference in a specific embodiment of the present invention;

[0051] Fig.12 It is a schematic diagram of the original data of full-node seismic data and the processed data after surface wave suppression and local strong noise suppression in a specific embodiment of the present invention;

[0052] Fig.13 It is a schematic diagram of actual seismic data containing local strong noise in a specific embodiment of the present invention;

[0053] Fig.14 It is a schematic diagram of a local strong noise signal obtained by intelligent recognition of a neural network in a specific embodiment of the present invention;

[0054] Fig.15 It is a schematic diagram of a valid signal obtained by intelligent recognition of a neural network in a specific embodiment of the present invention;

[0055] Fig.16 It is a schematic diagram of the time-frequency spectrum analysis of seismic data before and after neural network denoising in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0056] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations and / or combinations thereof.

[0058] The passive data-based denoising convolutional neural network local strong noise suppression method of the present invention comprises the following steps:

[0059] (1) This method first uses signal analysis methods such as wavelet transform and FK transform to analyze the time-space-frequency distribution characteristics of signals and noise in the full-node seismic data signal;

[0060] (2) Different types of noise are divided into different types. The common random noise and surface wave noise in full-node seismic data have time-frequency distribution characteristics that are significantly different from the effective signal and can be suppressed using conventional denoising technology. The target local strong noise interference includes burst pulses, abnormal disturbances, acoustic wave interference, etc., which have the characteristics of long duration, concentration in the 0-25 Hz frequency band and high overlap with the effective signal. Conventional denoising methods based on time-frequency distribution cannot effectively remove local strong noise.

[0061] (3) Based on the temporal and spatial distribution characteristics of the local strong noise in the target work area, which usually recurs in different regions and lasts for a long time, the active and passive data of the full-node seismic data are divided, and then the pure target noise signal segment is distinguished from the passive data;

[0062] (4) Setting a random sliding operator to automatically pick up local strong noise and construct data samples that meet the real noise characteristics;

[0063] (5) Use random cropping, scaling, and other transformation methods to expand the data, enrich the local strong noise signal library, and through manual screening, use the signals with obvious noise characteristics as the pre-training library, and select a part as the verification library;

[0064] (6) Construct effective signal samples through wave field forward simulation and using some actual acquisition data with high signal-to-noise ratio;

[0065] (7) Combining the residual learning idea, a denoising convolutional neural network model is designed for the target local strong noise. A residual module is set in the network for encoding and decoding to avoid the gradient vanishing problem of the deep network and improve the generalization ability of the model;

[0066] (8) Conduct network denoising training, add data enhancement and learning rate decay strategies to improve the optimal convergence speed and global minimum search accuracy of the network training process, and obtain a neural network that can effectively identify local strong noise;

[0067] (9) The noisy signal in the validation library is used as input to verify the denoising performance of the neural network after training convergence. The denoising performance of the network is tested through time-frequency analysis, signal-to-noise ratio and other indicators. If the denoising performance meets the requirements, the training is terminated. Otherwise, the learning rate is adjusted and the training in step (8) is performed again.

[0068] (10) After training, the full-node seismic data is input into the neural network for intelligent identification of target strong noise and denoising. The network learns from the rich noise samples in the passive data and can improve the signal-to-noise ratio by more than 50% while protecting the effective signal during the processing.

[0069] The following are several specific embodiments of the present invention.

[0070] Example 1

[0071] In a specific embodiment 1 of the present invention, Figure 1 As shown, Figure 1 The flowchart of the method for suppressing local strong noise of a denoising convolutional neural network based on passive data of the present invention is as follows. The method for suppressing local strong noise of a denoising convolutional neural network based on passive data comprises the following steps:

[0072] In step 101, the time-space-frequency distribution characteristics of local strong noise in the passive data signal are first analyzed by wavelet transform, FK transform and other means. Different types of noise are divided. Among them, the common random noise, linear noise and surface wave noise in the full-node seismic data have characteristics that are significantly different from the effective signal in frequency band distribution and propagation speed, and can be suppressed by conventional denoising technology; while the target local strong noise interference includes burst pulses, abnormal disturbances, acoustic wave interference, etc., which have the characteristics of long duration, concentrated in the 0-25Hz frequency band and highly overlapping with the effective signal. Conventional denoising methods based on time-frequency distribution cannot effectively remove local strong noise. The process enters step 102.

[0073] In step 102, according to the temporal and spatial distribution characteristics of the local strong noise in the target work area, the noise usually recurs in different regions and lasts for a long time, the active and passive data of the full-node seismic data are divided, and then the pure target noise signal segment is distinguished from the passive data. The process enters step 103.

[0074] In step 103, a random sliding operator is set to automatically pick up local strong noise in the pure target noise signal segment to construct a noise sample. Through this step, a data sample that meets the characteristics of real noise can be initially constructed for subsequent data expansion and training. The process enters step 104.

[0075] In step 104, data expansion is performed by using random cropping, scaling and other transformation methods to enrich the local strong noise signal library, and signals with obvious noise characteristics are selected as pre-training libraries through manual screening, and a part is selected as verification libraries. The process enters step 105.

[0076] In step 105, through the analysis of the actual full-node seismic data in the early stage, a model that conforms to the velocity characteristics and observation system of the work area is constructed to perform wave field forward simulation. In addition, effective signals are picked up using some actual acquisition data with high signal-to-noise ratio, and the two are combined to construct effective signal samples. Through this step, the data sample construction process can be completed and used for neural network training. The process enters step 106.

[0077] In step 106, a denoising convolutional neural network model is designed for the target local strong noise. The network combines the ideas of residual learning and residual module encoding and decoding, sets the network output to the passive data local strong noise (residual data), reduces the sensitivity of abnormal points, and sets the residual module for encoding and decoding to avoid the gradient vanishing problem of the deep network and improve the generalization ability of the model. The process enters step 107.

[0078] In step 107, the databases constructed in steps 104 and 105 are combined for network denoising training, and data enhancement and learning rate decay strategies are added during the training process to improve the optimal convergence speed of the network training process and the global minimum search accuracy, so as to obtain a neural network that can effectively identify local strong noise. The process proceeds to step 108.

[0079] In step 108, the noisy signal in the verification library is used as input to verify the denoising performance of the neural network after training convergence. The denoising performance of the network is tested through time-frequency analysis, signal-to-noise ratio and other indicators. If the denoising performance meets the requirements, the training is terminated and the process enters step 109; otherwise, the learning rate is adjusted and the network training in step 107 is performed again.

[0080] In step 109 , the network training is complete, a neural network model capable of identifying target noise in noisy data is obtained, and the network model parameters are saved for subsequent actual data denoising processing. The process proceeds to step 110 .

[0081] In step 110, the full-node seismic data is input into a neural network for intelligent identification of target strong noise and denoising. The network learns from abundant noise samples in passive data and can improve the signal-to-noise ratio by more than 50% while protecting effective signals during the processing.

[0082] Through the above process, passive data after suppressing local strong noise can be obtained through a denoising convolutional neural network. The method of the present invention can realize a local strong noise suppression method based on a denoising convolutional neural network of passive data. The beneficial effects of the present invention are: by extracting the characteristics of the strong background noise in the passive data collected in full time and space, the strong background noise is quickly identified, and the effective signal of the active data can be more effectively protected to improve the signal-to-noise ratio of the seismic data, and the intelligent identification and suppression of local strong noise can be realized, which verifies the application ability of passive data in full-node seismic data processing.

[0083] Example 2

[0084] In a specific embodiment 2 of the present invention, Figure 2 is the Wiger-Ville distribution of local strong noise. Local strong noise has the characteristic of overlapping with the effective signal frequency band. Conventional noise suppression methods can no longer suppress local strong noise while protecting the effective signal.

[0085] Figure 3 The time-space-frequency distribution characteristics of local strong noise in all-node seismic data show that local strong noise signals are distributed throughout the entire acquisition time period; they appear repeatedly in different regions, and the noise characteristics under the same coordinates are similar (the noise source is the same). Data from some regions can be used for deep learning to suppress noise in other regions; the duration is long, and seismic data are distributed throughout the entire time period, which is convenient for labeling.

[0086] Figure 4 This is a schematic diagram of a denoising residual convolutional neural network built for full-node seismic data.

[0087] Figure 5 It is a denoising sample training library constructed using passive coda data.

[0088] Figure 6 The target data features are obtained by combining feature matrices, and the feature vectors are used to identify local strong noise during the prediction process.

[0089] Figure 7 It is a training process of attenuated learning rate optimization and data enhancement strategy. During the network training process, some selected data are first input for strong noise recognition training; when the loss function decreases, data enhancement training is performed. And the learning rate attenuation strategy is introduced during the training process to improve the accuracy of global minimum search. This training strategy can accelerate the network convergence speed.

[0090] Figure 8 The data verification results are obtained by using a well-trained network model. After being processed by the neural network, local strong noise that overlaps with the effective signal frequency band can be intelligently identified and suppressed.

[0091] Fig. 9 It verifies the frequency domain analysis and denoising index evaluation before and after data denoising. After denoising, the MSE error is significantly reduced, and the signal SNR and PSNR are effectively improved.

[0092] Fig.10 It is the noise suppression result of actual single-shot seismic data.

[0093] Fig.11 It is a single-channel comparison of the denoised data with and without background strong noise interference. It can be found that the time-frequency characteristics of the two data channels are similar after processing, and the local strong noise with full-time distribution has been effectively suppressed.

[0094] Fig.12 It is the original data of all-node seismic data and the processed data after surface wave suppression and local strong noise suppression.

[0095] pass Fig.10 , Fig.11 as well as Fig.12 It can be seen that the denoising convolutional neural network model can effectively identify local strong noise, separate and suppress the noise whose frequency band overlaps with the effective signal, and improve the signal-to-noise ratio of passive data.

[0096] Example 3

[0097] In the specific embodiment 3 of the present invention, noise suppression is performed on single-shot seismic data containing local strong noise in a certain work area.

[0098] Fig.13 This is the noisy single-shot seismic data of this work area. First, the signal analysis method is used to analyze the time-space-frequency distribution characteristics of local strong noise in the seismic data. The valid signal on the right side of the data is contaminated by local strong noise. The target local strong noise interference includes burst pulses, abnormal disturbances, acoustic interference, etc., which have the characteristics of long duration, concentrated in the 3-22Hz frequency band and highly overlapped with the effective signal. The conventional denoising method based on time-frequency distribution cannot effectively remove the local strong noise, and it is necessary to use the strong noise intelligent recognition technology for denoising.

[0099] According to the spatiotemporal distribution characteristics of the local strong noise in the target work area, the noise usually recurs in different regions and lasts for a long time. In order to verify the generalization ability of neural network denoising, the training sample labels are designed from the work area of ​​Example 2 for neural network training, and applied to the seismic data denoising in this work area. The pure target noise (local strong noise) signal segment is divided in the passive data containing local strong noise, and then a random sliding operator is set to automatically pick up the local strong noise in the pure target noise signal segment to construct a noise sample. Data expansion is carried out using random cropping, scaling and other transformation methods to enrich the local strong noise signal library, and through manual screening, the signal with obvious noise characteristics is used as a pre-training library, and a part of it is extracted as a verification library for verification during the training process. The signal is picked up from part of the actual acquisition data that is not contaminated by noise to construct a valid signal sample.

[0100] A denoising convolutional neural network model is designed for the target local strong noise. The network combines the ideas of residual learning and residual module encoding and decoding, sets the network output to the passive data local strong noise (residual data), reduces the sensitivity of abnormal points, and sets the residual module for encoding and decoding to avoid the problem of gradient disappearance in deep networks and improve the network training convergence performance. Then, the neural network training is carried out in combination with the signal label library constructed previously. During the training process, data enhancement and learning rate decay strategies are added to improve the optimal convergence speed and global minimum search accuracy, and a neural network that can effectively identify local strong noise is obtained.

[0101] The noisy signal in the verification library is used as input to verify the denoising performance of the neural network after training convergence. The denoising performance of the network is tested through time-frequency analysis, signal-to-noise ratio and other indicators. If the denoising performance meets the requirements, the training is terminated; otherwise, the learning rate is adjusted to execute the network training again.

[0102] After the network training is complete, a neural network model that can identify target noise in noisy data is obtained, and the network model parameters are saved for subsequent actual data denoising processing.

[0103] The single shot seismic data of this work area is used as input and imported into the neural network to intelligently identify the target strong noise and perform denoising. Fig.14 It is the noise signal identified by the neural network after processing the noisy seismic data; Fig.15 It is the effective signal recognized by the neural network after processing.

[0104] The wavelet transform is performed on the seismic data before and after denoising, and the following is obtained: Fig.16 The time-frequency analysis comparison diagram is shown. Figure 13 to Figure 16 By comparison, it can be seen that the denoising convolutional neural network can utilize the local strong noise in the passive data for feature learning and effectively identify the local strong noise in the active data. The quality of the seismic data is significantly improved after denoising, and the local strong noise overlapping with the effective signal frequency band is effectively suppressed.

[0105] The learning process of the strong noise of the background of the present invention adopts the denoising residual convolutional neural network, and other artificial intelligence algorithms can also be adopted or expanded, such as the attention mechanism module, the noise2noise noise learning algorithm, etc. The local strong noise suppression method of the present invention uses the passive data of the full-node seismic data for sample construction, and performs feature learning for the local strong noise that is difficult to separate and extract features. There is no other alternative solution.

[0106] Finally, it should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions recorded in the aforementioned embodiments or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0107] Except for the technical features described in the specification, all other technical features are known technologies to those skilled in the art.

Claims

1. A local strong noise suppression method based on passive data denoising convolutional neural network, characterized in that: The passive data-based denoising convolutional neural network local strong noise suppression method includes: Step 1, analyzing the time-space-frequency distribution characteristics of signals and noise in the full-node seismic data signal; Step 2: Separate active and passive data to separate local strong noise; Step 3, constructing a data sample that meets the real noise characteristics; Step 4: data expansion and manual screening; Step 5, constructing a valid signal sample; Step 6: Build a denoising neural network; Step 7, perform neural network optimization training; Step 8, perform denoising evaluation and save network model parameters; Step 9: perform local strong noise identification to obtain high signal-to-noise ratio data.

2. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 1, the time-space-frequency distribution characteristics of the signal and noise in the full-node seismic data signal are first analyzed using signal analysis methods such as wavelet transform and FK transform.

3. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 2, characterized in that: In step 1, different types of noise are divided. The common random noise and surface wave noise in the full-node seismic data have time-frequency distribution characteristics that are significantly different from the effective signal and can be suppressed using conventional denoising techniques. The target local strong noise interference includes burst pulses, abnormal disturbances, and acoustic wave interference, which have the characteristics of long duration, concentration in the 0-25Hz frequency band, and high overlap with the effective signal. Conventional denoising methods based on time-frequency distribution cannot effectively remove local strong noise.

4. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 2, according to the spatiotemporal distribution characteristics of the local strong noise in the target work area, which usually recurs in different regions and lasts for a long time, the active and passive data of the full-node seismic data are divided, and then the pure target noise signal segment is distinguished from the passive data.

5. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 3, a random sliding operator is set to automatically pick up local strong noise in the pure target noise signal segment to construct a noise sample, thereby preliminarily constructing a data sample that conforms to the real noise characteristics for subsequent data expansion and training.

6. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 4, random cropping, scaling, and other transformation methods are used to expand the data to enrich the local strong noise signal library, and through manual screening, signals with obvious noise characteristics are used as pre-training libraries, and a part of them are selected as verification libraries.

7. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 5, effective signal samples are constructed through wave field forward simulation and using some actual acquisition data with high signal-to-noise ratio.

8. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 7, characterized in that: In step 5, through the preliminary analysis of the actual full-node seismic data, a model that conforms to the velocity characteristics and observation system of the work area is constructed to perform wave field forward simulation. In addition, some actual acquisition data with high signal-to-noise ratio are used to pick up effective signals. The two are combined to construct effective signal samples, thereby completing the data sample construction process and using it for neural network training.

9. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 6, combined with the residual learning idea, a denoising convolutional neural network model is designed for the target local strong noise. A residual module is set in the network for encoding and decoding to avoid the gradient vanishing problem of the deep network and improve the generalization ability of the model.

10. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 9, characterized in that: In step 6, a denoising convolutional neural network model is designed for the target local strong noise. The network combines the ideas of residual learning and residual module encoding and decoding, sets the network output to the local strong noise of passive data, i.e., residual data, to reduce the sensitivity of abnormal points, and sets the residual module for encoding and decoding to avoid the problem of vanishing gradients in deep networks and improve the generalization ability of the model.

11. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 7, the databases constructed in steps 4 and 5 are combined for network denoising training. Data enhancement and learning rate decay strategies are added during the training process to improve the optimal convergence speed of the network training process and the global minimum search accuracy, and obtain a neural network that can effectively identify local strong noise.

12. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 8, the noisy signal in the verification library is used as input to verify the denoising performance of the neural network after training convergence. The denoising performance of the network is detected by indicators such as time-frequency analysis and signal-to-noise ratio. If the denoising performance meets the requirements, the training is terminated and the network model parameters are saved; otherwise, the learning rate is adjusted and the network training in step 7 is performed again.

13. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 12, characterized in that: In step 8, when saving the network model parameters, the network training is complete, and a neural network model capable of identifying target noise in noisy data is obtained, and the network model parameters are saved for subsequent actual data denoising processing.

14. The method for suppressing local strong noise using a denoising convolutional neural network based on passive data according to claim 1, characterized in that: In step 9, the full-node seismic data is input into the neural network for intelligent identification of target strong noise and denoising. The network learns from the rich noise samples in the passive data and can improve the signal-to-noise ratio while protecting the effective signal during the processing.

Images