Seismic data high-resolution reconstruction and random noise suppression method

Abstract

The invention belongs to the technical field of seismic exploration and artificial intelligence, and relates to a seismic data high-resolution reconstruction and random noise suppression method, which comprises the following steps of: performing feature extraction and reconstruction on low-resolution noisy seismic data through a classic generator to generate a high-resolution denoised seismic image; and discriminant learning is carried out on real seismic data and a generation result through a quantum enhanced double-path discriminator, and the reconstruction capability of the generator is continuously optimized. According to the quantum enhanced double-path structure designed by the invention, the modeling capability of the model on complex seismic signals can be improved, the reconstructed image is finer, the structure is clearer, the robustness of the network on random noise can be enhanced, and more effective noise suppression is realized. According to the invention, the residual learning mechanism enables the feature interaction between the layers to be more sufficient, and facilitates the transmission and fusion of the seismic features at different layers, thereby improving the overall reconstruction quality. According to the method, the suppression effect on random noise can be remarkably improved while high-resolution reconstruction of seismic data is realized.

Description

Technical Field

[0001] This invention belongs to the fields of seismic exploration and artificial intelligence technology, specifically relating to a method for seismic data resolution reconstruction and random noise suppression based on quantum generative adversarial networks. Background Technology

[0002] In today's society, oil and natural gas, as major traditional energy sources, constitute the core pillar of the global energy supply system. They not only support the raw material supply for industrial systems but also provide a crucial power source for transportation systems. However, given their strong regional distribution, the exploration and extraction of oil and gas resources typically face challenges such as high technological complexity and significant differences in geological conditions, playing a decisive role in the balance and fluctuations of the international energy market. Faced with ever-increasing global energy consumption demands, countries urgently need to rely on advanced technologies to improve the efficiency of oil and gas exploration and extraction, and promote the transformation of resource development towards intelligent and sustainable development.

[0003] Currently, seismic exploration technology is widely used in the identification and assessment of oil and gas resources, playing a crucial role as an important tool for analyzing subsurface structures. However, current seismic exploration practices still face many technical bottlenecks, especially in situations where data acquisition is sparse or incomplete, significantly limiting its effectiveness. Complex terrain conditions, limitations of acquisition equipment, and high costs can all lead to insufficient data coverage, thus restricting the complete modeling of subsurface geological information. Low sampling density not only weakens the clarity and resolution of seismic images but also increases the uncertainty of the interpretation process, potentially leading to misjudgments in oil and gas reservoir identification. This is particularly true for deep or structurally complex oil and gas targets, where incomplete data results in more severe imaging distortion. Therefore, leveraging advanced information processing technologies such as artificial intelligence to improve the completeness and accuracy of seismic data while effectively controlling costs has become an urgent problem to be solved in current seismic exploration research.

[0004] For a long time, resolution reconstruction and random noise suppression of seismic data have been key research directions in geophysical exploration. To improve the clarity and signal-to-noise ratio of seismic profile images, numerous traditional methods have been proposed and widely applied in practical production and research. Traditional resolution enhancement methods mainly include deconvolution techniques, high-resolution spectral estimation, super-resolution interpolation, and frequency spreading. These methods typically rely on wavefield propagation theory or the assumption that the reflection coefficients have a certain spectral structure. For random noise suppression, common methods include statistical feature-based filtering techniques (such as median filtering and Wiener filtering), frequency domain denoising methods (such as bandwidth truncation), and spatial or temporal smoothing methods (such as FX prediction filtering and KL transform). In addition, sparse representation, low-rank decomposition, and adaptive transforms have also been applied to joint reconstruction and denoising problems. Although these classical methods perform well under specific conditions, they usually rely on prior knowledge of subsurface medium properties or waveform characteristics and are prone to failure in the face of high-noise backgrounds or complex subsurface structures. Especially in scenarios with high random noise intensity or large signal strength contrast, traditional methods struggle to simultaneously achieve noise suppression and preservation of effective signals. Furthermore, the sensitivity of parameter selection and the consumption of computational resources also limit the application of these methods in large-scale seismic data processing.

[0005] In recent years, with the development of deep learning technology, data-driven methods for seismic data resolution enhancement and noise suppression have gradually attracted attention. Initially, researchers used convolutional neural networks (CNNs) to build end-to-end models, attempting to learn the mapping relationship between the original seismic profile and high-resolution, low-noise images. Network structures such as U-Net, DnCNN, Residual-CNN, and GAN have been widely used in this task, achieving better performance than traditional methods. However, due to the inherent limitation of the local receptive field of CNNs, they have shortcomings in capturing long-distance dependencies, resulting in limited effectiveness in waveform detail reconstruction and complex interference noise suppression. To further improve modeling capabilities, researchers have begun to introduce the Transformer structure into seismic data augmentation tasks in recent years. Through its self-attention mechanism, the Transformer can capture local and global information simultaneously during feature extraction, demonstrating stronger modeling capabilities in waveform detail restoration and high-frequency information preservation. For example, the multi-scale Transformer network proposed by Wang et al. (2023) effectively suppressed background noise while improving data resolution. However, the Transformer architecture's high dependence on computing resources, especially the significant memory overhead and computational complexity it brings when processing long-term recorded or high-sampling-rate seismic data, has become a major obstacle to its widespread application in engineering practice.

[0006] Given the increasing urgency of balancing model complexity with practical application needs, researchers are exploring more innovative and promising model architectures. Generative Adversarial Networks (GANs), due to their outstanding performance in image enhancement and generation tasks, are being further applied to resolution reconstruction and noise suppression of seismic data, particularly demonstrating unique advantages in maintaining waveform realism and improving structural consistency. Meanwhile, with the development of quantum computing theory, Quantum Convolutional Neural Networks (QCNNs), as a deep learning architecture that integrates the advantages of quantum computing, are gradually being applied to complex signal processing tasks. They possess potential advantages in modeling nonlinear features, compressing model size, and improving generalization performance.

[0007] Based on this, if we can combine the strong reconstruction capabilities of GANs with the advantages of quantum convolution in feature representation to develop a high-resolution reconstruction and random noise suppression method for seismic data, this would be a new attempt and could provide a new path for building an efficient, accurate, and scalable seismic data augmentation framework while maintaining high resolution and low noise performance. Summary of the Invention

[0008] The purpose of this invention is to provide a high-resolution reconstruction and random noise suppression method for seismic data based on quantum generative adversarial networks, so as to solve the problem of improving the modeling ability and expression accuracy of complex seismic waveform features.

[0009] This invention is achieved through the following technical solution:

[0010] A method for high-resolution reconstruction and random noise suppression of seismic data includes the following steps:

[0011] S1. Construct training, validation, and test datasets for high-resolution reconstruction and random noise suppression methods for seismic data;

[0012] S2. Build a high-resolution reconstruction and random noise suppression model for seismic data based on quantum generative adversarial networks to improve the clarity and signal-to-noise ratio of seismic profile images;

[0013] S3. Design a loss function to evaluate the reconstruction quality and denoising effect of the high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial networks.

[0014] S4. Model Training Process: The high-resolution reconstruction and random noise suppression model for seismic data based on quantum generative adversarial networks constructed in step S2 is trained. This training process learns the data mapping mechanism between low-resolution noisy seismic data and high-resolution noise-free seismic data, and continuously optimizes the loss function set in step S3 based on the conversion result. After completing the predetermined number of training rounds, the model parameters gradually converge, and the optimal model parameters are obtained. The training task at this stage is then completed.

[0015] S5. Model Performance Evaluation: Test the quantum generative adversarial model trained in step S4 for high-resolution reconstruction and random noise suppression of seismic data. If the performance evaluation index obtained during the test meets the preset threshold requirements, the trained model can be identified as the optimal high-resolution reconstruction and random noise suppression model for seismic data. If the test index fails to meet the set standard, return to step S4 and retrain the model by adjusting the training parameters to further improve the model performance.

[0016] S6. Model Deployment and Application: Input the acquired real seismic profile data into the high-resolution reconstruction and random noise suppression model of the seismic data that has been determined to be optimal. After model processing, the output is the seismic profile data after high-resolution enhancement and noise suppression.

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

[0018] S11. Reflection Coefficient Model Construction: By constructing a large number of random reflection coefficient models, velocity parameters are set for various geological structures, including horizontal layers, stratum dip angles, fold structures, and fault zones, to simulate velocity variation characteristics in the actual geological environment.

[0019] S12. Generation of synthetic seismic data: The constructed reflection coefficient model is convolved with the Ricker wavelet of different frequencies. High-resolution noise-free seismic images are generated by convolving the high-frequency wavelet with the reflection coefficient model. At the same time, low-resolution noisy images are generated by convolving the low-frequency wavelet and adding random noise, and then by spatiotemporal thinning downsampling technology, thus forming a pair of synthetic seismic data.

[0020] S13. Diversified Data Expansion: Modify the stratigraphic structure by controlling random parameters, repeat steps S11-S12, and generate 1000 pairs of three-dimensional seismic bodies, covering diversified data with high-frequency wavelet peak frequencies of 30-50Hz, low-frequency wavelets of 5-20Hz, and signal-to-noise ratios of 5-15.

[0021] S14. Dataset partitioning: The synthesized 3D seismic body is extracted into 20,000 pairs of 2D seismic slices, which are then divided into training set (80%), validation set (10%), and test set (10%) in an 8:1:1 ratio to ensure the balance of data distribution.

[0022] Furthermore, in step S12, each pair of seismic data slices has a spatial resolution of 256×512 pixels and 512×1024 pixels, respectively.

[0023] Further, in step S2, a quantum generative adversarial network structure consisting of a classical generator and a quantum-enhanced dual-path discriminator is constructed to achieve high-resolution reconstruction and random noise suppression of seismic data. During the training phase, a low-resolution seismic image containing noise is fed into the generator as input, and the generator outputs a synthesized high-resolution denoised seismic image. This image is then passed as input to the quantum-enhanced dual-path discriminator, which performs discriminative analysis on the image through stepwise feature extraction to determine whether it is a real high-resolution noise-free image or an image synthesized by the generator.

[0024] Furthermore, the classic generator consists of a shallow feature extraction module, five residual learning modules, a feature fusion module, and an output reconstruction module. The input seismic profile image first passes through the shallow feature extraction module to extract basic features and expand the dimensions. Subsequently, the profile image features are sequentially passed to the five residual learning modules to achieve deep feature extraction. The calculated deep feature map and shallow feature map are then fused element-wise in the spatial dimension through the feature fusion module to integrate the information. Finally, the output reconstruction module generates a reconstructed high-resolution noise-free image.

[0025] Furthermore, the shallow feature extraction module includes a 9×9 convolutional layer and a PReLU activation function; the input image is first fed into the 9×9 convolutional layer to extract shallow features, and then the PReLU activation function is used to enhance the model's ability to express features; the residual learning module includes two 3×3 convolutional layers, two batch normalization layers, and a PReLU activation function; the input feature map is sequentially passed through the first 3×3 convolutional layer, the batch normalization layer, and the PReLU activation function to extract deep features; subsequently, the feature information is further fused through the second 3×3 convolutional layer and the batch normalization layer; finally, the input feature map of this module is residually connected with the feature map processed above; the feature fusion module includes a 3×3 convolutional layer and a batch normalization layer. The input feature map is first fed into a 3×3 convolutional layer for further feature fusion. Then, the feature distribution is standardized by a batch normalization layer. Finally, it is added element-wise with the feature map obtained by the shallow feature extraction module in the spatial dimension to achieve feature fusion. The output reconstruction module includes a 3×3 convolutional layer, a pixel rearrangement layer, a PReLU activation function, a 9×9 convolutional layer, and a Tanh activation function. The input feature map is first passed through a 3×3 convolutional layer and a pixel rearrangement layer to improve the image spatial resolution. Then, the high-dimensional feature map is restored to the seismic profile image space by a 9×9 convolutional layer. Finally, the output value is compressed to the range [0,1] by the Tanh activation function to meet the output standardization requirements.

[0026] Furthermore, the quantum-enhanced dual-path discriminator includes a classical feature extraction module, a quantum feature extraction module, an enhanced dual-path fusion module, and a classification and discrimination module. In the classical feature extraction module, the input image first passes through a 3×3 convolutional layer and a LeakyReLU activation function, and then through seven convolutional blocks consisting of 3×3 convolutional layers, batch normalization layers, and LeakyReLU activation functions. The resulting feature map is then input to the quantum feature extraction module, which processes the features using quantum convolution operations and outputs the dimension. The quantum feature map is consistent with the original space; the enhanced dual-path fusion module splices and fuses the classical feature map and the quantum feature map in the channel dimension; finally, it is sent to the classification and discrimination module, which includes an adaptive average pooling layer and two 1×1 convolutional layers; the feature map is first input to the adaptive average pooling layer to compress the spatial dimension features to a size of 1×1; then it is passed through two 1×1 convolutional layers in sequence to realize the dimensionality increase and dimensionality reduction operations of the features respectively; finally, the discrimination result is output to determine whether the input image is a real high-resolution noise-free image or an image synthesized by the generator.

[0027] Furthermore, the quantum convolution consists of four parts: image block segmentation, quantum state encoding, quantum variational circuit evolution, and measurement reconstruction. The input feature map is first divided into several non-overlapping image blocks of size 2×2. The pixel values ​​in each image block are mapped to angle parameters according to a normalization method to generate the corresponding quantum state encoding. Subsequently, the quantum state evolves through a predefined quantum variational circuit, which includes several quantum gate operations, such as RX, RY, RZ rotation gates and CNOT gates, to perform nonlinear transformations on the quantum features of the image blocks. Different image blocks use the same quantum variational circuit structure, which is simulated in PyTorch through the qiskit interface. After the evolution is completed, a measurement operation is performed on each qubit, and the expected value corresponding to the Pauli-Z basis is extracted as the output feature after quantum processing. The measurement results are reconstructed into a feature map block with the same position as the original image block. All processed blocks are stitched back to the spatial position of the original image to form the final quantum feature map.

[0028] Furthermore, in step S3, the loss function includes the generator loss. G And quantum-enhanced dual-path discriminator loss QDP Generator Loss G Loss due to basic reconstruction R Perceived loss F and combat loss A It consists of three parts:

[0029] Loss G =λ1Loss R +λ2Loss F +λ3Loss A (1)

[0030] Where λ1, λ2, and λ3 are adjustable hyperparameters used to control the weights of each loss term, and Loss... R This represents pixel-level reconstruction error, specifically defined as:

[0031]

[0032] in, This indicates that the generator network responds to noisy input X. i The reconstruction results, Y i For the corresponding true high-resolution image, Z = t 2 ·W·H is the total number of pixels in the image, t is the magnification factor, and W and H are the width and height of the image;

[0033] Perceived Loss FRepresenting the feature differences in the deep semantic space, using the feature map ψ extracted from the l-th layer of the deep residual network. (l) (·) is represented as:

[0034]

[0035] Among them, C l H l W l These are the number of channels, height, and width of the feature map in layer l, respectively.

[0036] Combat Loss A The algorithm learns the difference between the generated images and the real data in the overall distribution, where N represents the number of generated images, and its output is defined as:

[0037]

[0038] Quantum-enhanced dual-path discriminator loss QDP The mathematical expression for measuring the separability of generated data and real data in the quantum joint feature space is:

[0039]

[0040] in, f represents the sample generated by the j-th generator. C (·) is the feature map extracted from the path of a classic convolutional network, f Q (·) indicates the quantum path. This represents a dual-path fusion discriminator that accepts classical and quantum features as input, where M represents the number of real images.

[0041] Furthermore, in step S5, the model testing specifically includes:

[0042] S51. Input the pairs of low-resolution noisy seismic profiles and corresponding high-resolution noise-free seismic profiles from the test set into a pre-trained seismic data reconstruction and noise suppression model based on a quantum generative adversarial network for testing and verification.

[0043] S52. If the evaluation index (peak signal-to-noise ratio and structural similarity) obtained in the test results does not meet the preset threshold requirements, return to step S4 and re-optimize and train the parameters of the high-resolution reconstruction and random noise suppression method of seismic data based on quantum generative adversarial network.

[0044] S53. When the evaluation metrics (peak signal-to-noise ratio and structural similarity) obtained during the testing phase reach the preset threshold standard, the model training can be terminated, and the trained high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial networks can be applied to actual seismic profile data processing tasks.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. This invention combines quantum convolution and GAN network to design the model structure, so that while maintaining the high-resolution reconstruction and noise suppression capabilities of seismic data, the model has the advantages of stronger feature expression, more accurate waveform detail recovery, and greater adaptability to complex structures compared to traditional CNN or pure GAN architectures.

[0047] 2. This invention designs a quantum-enhanced dual-path discriminator. By employing a feature extraction module constructed collaboratively by classical convolution and quantum convolution, it fully mines the local and global feature information in seismic profiles. At the same time, it introduces a dimensionality-upgrading and fusion strategy to achieve deep interaction between classical and quantum features in the channel dimension, effectively enhancing the model's ability to identify and reconstruct complex waveform structures.

[0048] 3. This invention uses a GAN network generator to model the mapping relationship between noise space and seismic profile data distribution, thereby achieving high-fidelity generation and restoration of weak reflection signals and detailed structures in seismic profiles, effectively improving the resolution and noise suppression performance of seismic data.

[0049] 4. This invention employs a quantum convolution mechanism to divide the input feature map into multiple image blocks, and extracts nonlinear features through quantum state encoding, quantum variational circuit evolution, and measurement operations. Finally, the processed image blocks are reconstructed into a complete feature map, thereby effectively enhancing local perception and global modeling capabilities. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A schematic diagram of the high-resolution reconstruction and random noise suppression network structure for seismic data in this invention;

[0052] Figure 2 Flowchart of the steps in the high-resolution reconstruction and random noise suppression method for seismic data of this invention;

[0053] Figure 3 The test set includes low-resolution and noisy seismic profiles, corresponding high-resolution and denoised seismic profiles obtained using ordinary GAN networks, and seismic profiles reconstructed from seismic data using quantum generative adversarial networks with high resolution and random noise suppression.

[0054] Figure 4 Seismic profiles with high-resolution reconstruction and random noise suppression based on ordinary GAN network, and the average amplitude of all traces of seismic profiles with high-resolution reconstruction and random noise suppression based on quantum generative adversarial network and corresponding low-resolution and noisy seismic profiles in the test set are compared.

[0055] Figure 5 Comparison curves of single-channel time-domain waveforms of seismic profiles after high-resolution reconstruction and random noise suppression based on ordinary GAN networks, seismic profiles after high-resolution reconstruction and random noise suppression based on quantum generative adversarial networks, and corresponding low-resolution and noisy seismic profiles in the test set. Detailed Implementation

[0056] The present invention will be further described below with reference to embodiments:

[0057] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0058] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0059] This invention was developed and debugged in the Python environment using the Anaconda platform, running on Ubuntu 22.04LTS, with an NVIDIA GeForce RTX 4090 graphics card and PyTorch 2.5.1 as the deep learning framework. The core of this paper is to propose a high-resolution reconstruction and random noise suppression method for seismic data based on quantum generative adversarial networks (GANs). Due to limitations in exploration environment and economic costs, acquired seismic data often suffers from sparse sampling, missing data, and noise contamination, severely impacting subsequent interpretation and imaging accuracy. Existing deep neural network-based reconstruction and denoising methods often face bottlenecks such as insufficient feature extraction capabilities and poor generalization when dealing with complex geological structures or non-Gaussian noise backgrounds. Therefore, this invention proposes a high-resolution reconstruction and random noise suppression method for seismic data based on quantum generative adversarial networks (GANs). By designing a quantum-enhanced dual-path discriminator to enhance nonlinear feature representation capabilities, more accurate seismic data detail recovery and noise suppression are achieved. During the model training phase, paired low-resolution noisy and high-resolution noise-free seismic profile datasets were constructed to train and evaluate the proposed network structure. When the network's performance metrics (signal-to-noise ratio SNR, structural similarity SSIM, etc.) on the test set reach a preset threshold, the network can be determined as the current optimal model and can be used for resolution enhancement and noise suppression tasks in actual seismic data.

[0060] like Figure 2 As shown, the high-resolution reconstruction and random noise suppression method for seismic data of the present invention includes the following steps:

[0061] S1. Construct training, validation, and test datasets for high-resolution reconstruction and random noise suppression methods for seismic data;

[0062] First, by constructing a large number of random reflection coefficient models, and by setting velocity parameters for various geological structures such as horizontal layers, stratum dip angles, fold structures, and fault zones, the velocity variation characteristics in the actual geological environment are simulated.

[0063] Secondly, the constructed reflection coefficient model is convolved with the Ricker wavelet of different frequencies. High-resolution noise-free seismic images are generated by convolving the high-frequency wavelet with the reflection coefficient model. At the same time, low-resolution noisy images are generated by convolving the low-frequency wavelet and adding random noise, and then by spatiotemporal thinning downsampling technology. This forms a pair of synthetic seismic data, with each pair of seismic data slices having spatial resolutions of 256×512 pixels and 512×1024 pixels, respectively.

[0064] Then, by modifying the stratigraphic structure through random parameter control, the process of steps S11-S12 is repeated to generate 1000 pairs of three-dimensional seismic bodies, covering diverse data with high-frequency wavelet peak frequencies of 30-50Hz, low-frequency wavelets of 5-20Hz, and signal-to-noise ratios of 5-15.

[0065] Finally, the synthesized 3D seismic volume was extracted into 20,000 pairs of 2D seismic slices, which were divided into training set (80%), validation set (10%) and test set (10%) in an 8:1:1 ratio to ensure the balance of data distribution.

[0066] Through the above data preparation stages, this invention constructs a paired dataset consisting of low-resolution noisy seismic profile slices and corresponding high-resolution noise-free slices. The training set contains 16,000 data pairs, while the validation and test sets each contain 2,000 slice data pairs. Each pair of seismic data slices has a spatial resolution of 256×512 pixels and 512×1024 pixels, respectively.

[0067] S2. Build a high-resolution reconstruction and random noise suppression model for seismic data based on quantum generative adversarial networks to improve the clarity and signal-to-noise ratio of seismic profile images;

[0068] The designed quantum generative adversarial network (GAN) architecture consists of a classical generator and a quantum-enhanced dual-path discriminator, aiming to achieve high-resolution reconstruction and random noise suppression of seismic data. During the training phase, a noisy, low-resolution seismic image is fed into the generator, which outputs a synthesized high-resolution, denoised seismic image. This image is then fed into the quantum-enhanced dual-path discriminator, which performs discriminative analysis through step-by-step feature extraction to determine whether it is a genuine high-resolution, noise-free image or an image synthesized by the generator.

[0069] The aforementioned classic generator consists of a shallow feature extraction module, five residual learning modules, a feature fusion module, and an output reconstruction module. The input seismic profile image first passes through the shallow feature extraction module to extract basic features and expand its dimensions. Subsequently, the profile image features are sequentially passed to the five residual learning modules to achieve deep feature extraction. The calculated deep feature map and shallow feature map are then fused element-wise in the spatial dimension by the feature fusion module to integrate the information. Finally, the output reconstruction module generates a reconstructed high-resolution noise-free image.

[0070] The shallow feature extraction module includes a 9×9 convolutional layer and a PReLU activation function. The input image is first fed into the 9×9 convolutional layer to extract shallow features, and then the PReLU activation function enhances the model's ability to express features. The residual learning module includes two 3×3 convolutional layers, two batch normalization layers, and a PReLU activation function. The input feature map is sequentially passed through the first 3×3 convolutional layer, the batch normalization layer, and the PReLU activation function to extract deep features; then, the feature information is further fused through the second 3×3 convolutional layer and the batch normalization layer. Finally, the input feature map of this module is residually connected to the processed feature map. The feature fusion module includes a 3×3 convolutional layer and a batch normalization layer. The input feature map is first fed into the 3×3 convolutional layer for further feature fusion, then the feature distribution is standardized by the batch normalization layer, and finally, it is element-wise added to the feature map obtained by the shallow feature extraction module in the spatial dimension to achieve feature fusion. The output reconstruction module sequentially includes a 3×3 convolutional layer, a pixel rearrangement layer, a PReLU activation function, a 9×9 convolutional layer, and a Tanh activation function. The input feature map is first passed through a 3×3 convolutional layer and a pixel rearrangement layer to improve the image spatial resolution; then, a 9×9 convolutional layer is used to restore the high-dimensional feature map to the seismic profile image space; finally, the Tanh activation function compresses the output value to the range [0,1] to meet the output standardization requirements.

[0071] The aforementioned quantum-enhanced dual-path discriminator includes a classical feature extraction module, a quantum feature extraction module, an enhanced dual-path fusion module, and a classification and discrimination module. In the classical feature extraction module, the input image is sequentially passed through four 3×3 convolutional layers, and the resulting feature map is then input to the quantum feature extraction module. The quantum feature extraction module processes the features using quantum convolution operations and outputs a quantum feature map with the same dimensionality as the original space. The enhanced dual-path fusion module concatenates and fuses the classical and quantum feature maps along the channel dimension, and finally sends them to the classification and discrimination module. This module includes an adaptive average pooling layer and two 1×1 convolutional layers. The feature map is first input to the adaptive average pooling layer, compressing the spatial dimension features to a 1×1 size; then it sequentially passes through two 1×1 convolutional layers, respectively implementing dimensionality upscaling and dimensionality reduction operations; finally, it outputs a discrimination result to determine whether the input image is a true high-resolution noise-free image or an image synthesized by the generator.

[0072] The aforementioned quantum convolution consists of four parts: image patch segmentation, quantum state encoding, quantum variational circuit evolution, and measurement reconstruction. The input feature map is first divided into several 2×2 non-overlapping image patches. The pixel values ​​in each patch are normalized and mapped to angle parameters to generate the corresponding quantum state encoding. Subsequently, the quantum states evolve through a predefined quantum variational circuit, which includes four quantum gate operations: RX, RY, RZ rotation gates, and a CNOT gate. Different image patches use the same quantum circuit structure. After evolution, a measurement operation is performed on each qubit, and the expected value corresponding to the Pauli-Z basis is extracted as the output feature after quantum processing. The measurement results are reconstructed into a feature patch with the same position as the original image patch. All processed patches are stitched back to their original spatial positions to form the final quantum feature map.

[0073] S3. Design a loss function to evaluate the reconstruction quality and denoising effect of the high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial networks.

[0074] The loss function includes the generator loss. G And quantum-enhanced dual-path discriminator loss QDP Generator Loss G Loss due to basic reconstruction R Perceived loss F and combat loss A It consists of three parts:

[0075] Loss G =λ1Loss R +λ2Loss F +λ3Loss A (1)

[0076] Where λ1, λ2, and λ3 are adjustable hyperparameters used to control the weights of each loss term, and Loss... R This represents pixel-level reconstruction error, specifically defined as:

[0077]

[0078] in, This indicates that the generator network responds to noisy input X. i The reconstruction results, Y i For the corresponding true high-resolution image, Z = t 2 ·W·H represents the total number of pixels in the image, t represents the magnification factor, and W and H represent the width and height of the image.

[0079] Perceived Loss FRepresenting the feature differences in the deep semantic space, using the feature map ψ extracted from the l-th layer of the deep residual network. (l) (·) is represented as:

[0080]

[0081] Among them, C l H l W l These are the number of channels, height, and width of the feature map in layer l, respectively.

[0082] Combat Loss A The algorithm learns the difference between the generated images and the real data in the overall distribution, where N represents the number of generated images, and its output is defined as:

[0083]

[0084] Quantum-enhanced dual-path discriminator loss QDP The mathematical expression for measuring the separability of generated data and real data in the quantum joint feature space is:

[0085]

[0086] in, f represents the sample generated by the j-th generator. C (·) is the feature map extracted from the path of a classic convolutional network, f Q (·) indicates the quantum path. This represents a dual-path fusion discriminator that accepts classical and quantum features as input, where M represents the number of real images.

[0087] S4. Model Training Process: The high-resolution reconstruction and random noise suppression model for seismic data based on quantum generative adversarial networks constructed in step S2 is trained. This training process learns the data mapping mechanism from low-resolution noisy seismic data to high-resolution noise-free seismic data, and continuously optimizes the loss function set in step S3 based on the transformation result. After completing the predetermined number of training rounds, the model parameters gradually converge, and finally, the optimal model parameters are obtained, thus completing this stage of the training task.

[0088] S5. Model Performance Evaluation: The quantum generative adversarial network model trained in step S4 for high-resolution reconstruction and random noise suppression of seismic data is tested. If the performance evaluation index obtained during the test meets the preset threshold requirements, the trained network model can be identified as the optimal model for high-resolution reconstruction and random noise suppression of seismic data. If the test index fails to meet the set standard, return to step S4 and retrain the model by adjusting the training parameters to further improve the model performance.

[0089] The model testing in step S5 specifically includes:

[0090] S51. Input the pairs of low-resolution noisy seismic profiles and corresponding high-resolution noise-free seismic profiles from the test set into a pre-trained seismic data reconstruction and noise suppression model based on a quantum generative adversarial network for testing and verification.

[0091] S52. If the evaluation index (peak signal-to-noise ratio and structural similarity) obtained in the test results does not meet the preset threshold requirements, return to step S4 and re-optimize and train the parameters of the high-resolution reconstruction and random noise suppression method of seismic data based on quantum generative adversarial network.

[0092] S53. When the evaluation metrics (peak signal-to-noise ratio and structural similarity) obtained during the testing phase reach the preset threshold standard, the model training can be terminated, and the trained high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial networks can be applied to actual seismic profile data processing tasks.

[0093] S6. Model Deployment and Application: Input the acquired real seismic profile data into the high-resolution reconstruction and random noise suppression model of the seismic data that has been determined to be optimal. After model processing, the output is the seismic profile data after high-resolution enhancement and noise suppression.

[0094] Example 1

[0095] This embodiment provides an application of a quantum generative adversarial model for high-resolution reconstruction and random noise suppression of low-resolution noisy seismic profiles, as detailed below:

[0096] In this embodiment, low-resolution and noisy data from the test set are input into a pre-trained high-resolution seismic data reconstruction and random noise suppression model based on quantum generative adversarial methods. The output results are as follows: Figure 3 As shown, the output seismic profile not only possesses a clearer bedding structure and finer waveform details, but also exhibits significantly suppressed background noise, resulting in a marked improvement in the overall signal-to-noise ratio. Figure 4As shown, the high-resolution reconstructed and random noise-suppressed seismic profiles of this invention exhibit a more significant amplitude increase compared to seismic profiles generated by ordinary GAN networks and the original seismic profiles in the test set, effectively recovering high-frequency components. Figure 5 As shown, the seismic profile reconstructed and suppressed by the present invention has a single-channel amplitude that is closer to the original seismic profile in the test set than the seismic profile generated by a conventional GAN ​​network, proving that the present method has effective high-resolution reconstruction and random noise suppression performance.

[0097] This invention employs a classical generator to extract and reconstruct features from low-resolution, noisy seismic data, generating high-resolution, denoised seismic images. A quantum-enhanced dual-path discriminator is used to learn and discriminate between real seismic data and the generated results, continuously optimizing the generator's reconstruction capabilities. The quantum-enhanced dual-path structure designed in this invention not only improves the model's ability to model complex seismic signals, resulting in more detailed and structurally clearer reconstructed images, but also enhances the network's robustness to random noise, achieving more effective noise suppression. Simultaneously, the residual learning mechanism in this invention allows for more thorough feature interaction between layers, facilitating the transmission and fusion of seismic features at different levels, thereby improving the overall reconstruction quality. This invention can achieve high-resolution reconstruction of seismic data while significantly improving the suppression of random noise, providing a more accurate and robust data foundation for subsequent structural fault analysis and refined geological modeling.

[0098] Note that the above description is merely a preferred embodiment of the present invention and the technical principles employed. Those skilled in the art will understand that the present invention is not limited to the specific embodiments described herein, and various obvious changes, readjustments, and substitutions can be made without departing from the scope of protection of the present invention. Therefore, although the present invention has been described in detail through the above embodiments, the present invention is not limited to the above embodiments, and may include many other equivalent embodiments without departing from the concept of the present invention, the scope of which is determined by the scope of the appended claims.

Claims

1. A method for high-resolution reconstruction and random noise suppression of seismic data, characterized in that, Includes the following steps: S1. Construct training, validation, and test datasets for high-resolution reconstruction and random noise suppression methods for seismic data; S2. Build a high-resolution reconstruction and random noise suppression model for seismic data based on quantum generative adversarial networks to improve the clarity and signal-to-noise ratio of seismic profile images; S3. Design a loss function to evaluate the reconstruction quality and denoising effect of the high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial networks. S4. Model training process: Train the high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial network constructed in step S2; The training process learns the data mapping mechanism between low-resolution noisy seismic data and high-resolution noiseless seismic data, and continuously optimizes the loss function set in step S3 based on the transformation result; after completing the predetermined number of training rounds, the model parameters gradually converge, and finally the optimal model parameters are obtained. S5. Model Performance Evaluation: Test the quantum generative adversarial model trained in step S4 for high-resolution reconstruction and random noise suppression of seismic data. If the performance evaluation index obtained during the test meets the preset threshold requirements, the trained model can be identified as the optimal high-resolution reconstruction and random noise suppression model for seismic data; if the test index fails to meet the set standard, return to step S4, and retrain the model by adjusting the training parameters to further improve the model performance. S6. Model Deployment and Application: Input the acquired real seismic profile data into the high-resolution reconstruction and random noise suppression model of the seismic data that has been determined to be optimal. After model processing, the output is the seismic profile data after high-resolution enhancement and noise suppression.

2. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Reflection Coefficient Model Construction: By constructing a large number of random reflection coefficient models, velocity parameters are set for various geological structures, including horizontal layers, stratum dip angles, fold structures, and fault zones, to simulate velocity variation characteristics in the actual geological environment. S12. Generation of synthetic seismic data: The constructed reflection coefficient model is convolved with the Ricker wavelet of different frequencies. High-resolution noise-free seismic images are generated by convolving the high-frequency wavelet with the reflection coefficient model. At the same time, low-resolution noisy images are generated by convolving the low-frequency wavelet and adding random noise, and then by spatiotemporal thinning downsampling technology, thus forming a pair of synthetic seismic data. S13. Diversified Data Expansion: Modify the stratigraphic structure by controlling random parameters, repeat steps S11-S12, and generate 1000 pairs of three-dimensional seismic bodies, covering diversified data with high-frequency wavelet peak frequencies of 30-50Hz, low-frequency wavelets of 5-20Hz, and signal-to-noise ratios of 5-15. S14. Dataset partitioning: The synthesized 3D seismic body is extracted into 20,000 pairs of 2D seismic slices, which are then divided into training, validation, and test sets in an 8:1:1 ratio to ensure the balance of data distribution.

3. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 2, characterized in that, Step S12: Each pair of seismic data slices has a spatial resolution of 256×512 pixels and 512×1024 pixels, respectively.

4. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 1, characterized in that: Step S2 specifically involves constructing a quantum generative adversarial network structure consisting of a classical generator and a quantum-enhanced dual-path discriminator to achieve high-resolution reconstruction and random noise suppression of seismic data. During the training phase, a low-resolution seismic image containing noise is fed into the generator as input, and the generator outputs a synthesized high-resolution denoised seismic image. This image is then passed as input to the quantum-enhanced dual-path discriminator, which performs discriminative analysis on the image through stepwise feature extraction to determine whether it is a real high-resolution noise-free image or an image synthesized by the generator.

5. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 4, characterized in that: The classic generator consists of a shallow feature extraction module, five residual learning modules, a feature fusion module, and an output reconstruction module. The input seismic profile image first passes through the shallow feature extraction module to extract basic features and expand the dimensions. Subsequently, the profile image features are passed sequentially to the five residual learning modules to achieve deep feature extraction. The calculated deep feature map and shallow feature map are then fused element-wise in the spatial dimension by the feature fusion module to integrate the information. Finally, the output reconstruction module generates a reconstructed high-resolution noise-free image.

6. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 5, characterized in that: The shallow feature extraction module includes a 9×9 convolutional layer and a PReLU activation function. The input image is first fed into the 9×9 convolutional layer to extract shallow features, and then the PReLU activation function is used to enhance the model's ability to express features. The residual learning module includes two 3×3 convolutional layers, two batch normalization layers, and a PReLU activation function. The input feature map is sequentially passed through the first 3×3 convolutional layer, the batch normalization layer, and the PReLU activation function to extract deep features. Subsequently, the feature information is further fused through the second 3×3 convolutional layer and the batch normalization layer. Finally, the input feature map of this module is compared with the processed feature map. The residual connection is now established; the feature fusion module includes a 3×3 convolutional layer and a batch normalization layer; the input feature map is first fed into the 3×3 convolutional layer for further feature fusion, then the feature distribution is standardized by the batch normalization layer, and finally, it is added element-wise in the spatial dimension to the feature map obtained by the shallow feature extraction module to achieve feature fusion; the output reconstruction module includes a 3×3 convolutional layer, a pixel rearrangement layer, a PReLU activation function, a 9×9 convolutional layer, and a Tanh activation function; the input feature map is first passed through the 3×3 convolutional layer and the pixel rearrangement layer to improve the image spatial resolution; Subsequently, the high-dimensional feature maps are restored to the seismic profile image space through a 9×9 convolutional layer; finally, the output values ​​are compressed to the range of [0,1] by the Tanh activation function to meet the standardization requirements of the output.

7. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 4, characterized in that: The quantum-enhanced dual-path discriminator includes a classical feature extraction module, a quantum feature extraction module, an enhanced dual-path fusion module, and a classification and discrimination module. In the classical feature extraction module, the input image first passes through a 3×3 convolutional layer and a LeakyReLU activation function, and then through seven convolutional blocks consisting of 3×3 convolutional layers, batch normalization layers, and LeakyReLU activation functions. The resulting feature map is then input to the quantum feature extraction module. The quantum feature extraction module processes the features using quantum convolution operations and outputs a quantum feature map with the same dimension as the original space. The enhanced dual-path fusion module concatenates and fuses classical and quantum feature maps along the channel dimension. Finally, the merged feature maps are fed into a classification and discrimination module, which includes an adaptive average pooling layer and two 1×1 convolutional layers. The feature maps are first input to the adaptive average pooling layer, compressing the spatial dimension features to a 1×1 size. They then pass through two 1×1 convolutional layers sequentially, performing dimensionality upscaling and downscaling operations respectively. The final output is a discrimination result used to determine whether the input image is a true high-resolution, noise-free image or an image synthesized by the generator.

8. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 7, characterized in that: The quantum convolution consists of four parts: image block segmentation, quantum state encoding, quantum variational circuit evolution, and measurement reconstruction. The input feature map is first divided into several 2×2 non-overlapping image blocks. Pixel values ​​in each block are normalized and mapped to angle parameters to generate corresponding quantum state codes. Subsequently, the quantum states evolve through a predefined quantum variational circuit, which includes four quantum gate operations: RX, RY, RZ rotation gates, and CNOT gates. Different image blocks use the same quantum circuit structure. After evolution, a measurement operation is performed on each qubit, and the expected value corresponding to the Pauli-Z basis is extracted as the output feature after quantum processing. The measurement results are reconstructed into a feature map block with the same position as the original image block. All processed blocks are stitched back to their original spatial positions to form the final quantum feature map.

9. The method for high-resolution reconstruction and random noise suppression of seismic data according to claim 1, characterized in that: In step S3, the loss function includes the generator loss. G And quantum-enhanced dual-path discriminator loss QDP Generator Loss G Loss due to basic reconstruction R Perceived loss F and combat loss A Three parts composition: Loss G =λ1Loss R +λ2Loss F +λ3Loss A (1) Where λ1, λ2, and λ3 are adjustable hyperparameters used to control the weights of each loss term, and Loss... R This represents pixel-level reconstruction error, specifically defined as: in, This indicates that the generator network responds to noisy input X. i The reconstruction results, Y i For the corresponding true high-resolution image, Z = t 2 ·W·H is the total number of pixels in the image, t is the magnification factor, and W and H are the width and height of the image; Perceived Loss F Representing the feature differences in the deep semantic space, using the feature map ψ extracted from the l-th layer of the deep residual network. (l) (·) is represented as: Among them, C l H l W l These are the number of channels, height, and width of the feature map in layer l, respectively. Combat Loss A The algorithm learns the difference between the generated images and the real data in the overall distribution, where N represents the number of generated images, and its output is defined as: Quantum-enhanced dual-path discriminator loss QDP The mathematical expression for measuring the separability of generated data and real data in the quantum joint feature space is: in, f represents the sample generated by the j-th generator. C (·) is the feature map extracted from the path of a classic convolutional network, f Q (·) indicates the quantum path. This represents a dual-path fusion discriminator that accepts classical and quantum features as input, where M represents the number of real images.

10. A method for high-resolution reconstruction and random noise suppression of seismic data according to claim 1, characterized in that, The model testing in step S5 specifically includes: S51. Input the pairs of low-resolution noisy seismic profiles and corresponding high-resolution noise-free seismic profiles from the test set into a pre-trained seismic data reconstruction and noise suppression model based on a quantum generative adversarial network for testing and verification. S52. If the evaluation index obtained in the test results does not meet the preset threshold requirements, return to step S4 and re-optimize and train the parameters of the high-resolution reconstruction and random noise suppression method of seismic data based on quantum generative adversarial network. S53. When the evaluation index obtained in the testing phase reaches the preset threshold standard, the model training can be terminated, and the trained high-resolution reconstruction and random noise suppression model of seismic data based on quantum generative adversarial network can be applied to the actual seismic profile data processing task.

Images