Training method and processing method of seismic sensor signal optimization model, model

Abstract

The application provides a training method and processing method of a seismic sensor signal optimization model, and relates to the technical field of seismic signal optimization. The training method comprises: obtaining a first feature alignment loss and a first detail modulation loss by forward conversion of a low-quality reference signal; obtaining a second feature alignment loss and a second detail modulation loss by reverse conversion of a high-quality reference signal; obtaining a first cycle consistency loss by reconstruction of the forward conversion result; obtaining a second cycle consistency loss by reconstruction of the reverse conversion result; obtaining a first adversarial loss by discriminating the forward conversion result, and obtaining a second adversarial loss by discriminating the reverse conversion result; weighting and combining all the losses obtained into a total loss function, and alternately optimizing and training the model based on the total loss function until convergence, to obtain a trained optimization model. Finally, a seismic sensor signal optimization model capable of effectively suppressing noise and maintaining effective wave components is obtained.

Description

Technical Field

[0001] This application relates to the field of seismic signal optimization technology, and more specifically, to a training method, processing method, and model for seismic sensor signal optimization. Background Technology

[0002] To improve the signal quality of MEMS sensors, existing technologies typically employ signal processing or data-driven methods to enhance the acquired signals. However, in sensor signal enhancement tasks that require high-fidelity detail restoration of unpaired data, there are still problems such as insufficient supervisory information and limited detail reconstruction capabilities. There is an urgent need to explore more effective enhancement frameworks and methods.

[0003] First, enhancement methods based on traditional signal processing often rely on certain assumptions about the distribution of signals or noise, and their effectiveness is easily affected in complex real-world environments, lacking sufficient adaptability and robustness. At the same time, while suppressing noise, these methods often inevitably lose subtle features and useful information in the signal, resulting in a deviation between the enhanced waveform and the real high-quality signal.

[0004] Secondly, with the development of machine learning methods, data-driven signal enhancement techniques have shown considerable potential. These methods can suppress noise and restore signals by learning from large amounts of data. However, these methods typically require a large number of strictly paired high- and low-quality signals as training data. In practice, due to limitations in equipment cost, deployment conditions, and synchronous acquisition, such paired samples are often difficult to obtain, thus severely restricting their application in real-world scenarios.

[0005] Furthermore, although existing studies have attempted to use unpaired data for signal enhancement to avoid dependence on paired data, such methods often fail to ensure that features in low-quality signals can be fully and accurately aligned with features in high-quality signals in the absence of direct supervision. This results in significant deficiencies in detail restoration and waveform fidelity of the enhanced signals, making it impossible to fully recover the fine signal structure and dynamic features of high-end sensing devices. Summary of the Invention

[0006] In view of this, the purpose of this application is to provide a training method, processing method, and model for an earthquake sensor signal optimization model, so as to improve the above-mentioned problems existing in the prior art.

[0007] In a first aspect, embodiments of this application provide a training method for an optimized model of a seismic sensor signal. The training method includes: obtaining a first feature alignment loss and a first detail modulation loss by forward transformation of a low-quality reference signal; obtaining a second feature alignment loss and a second detail modulation loss by inverse transformation of a high-quality reference signal; obtaining a first cycle consistency loss by reconstructing the forward transformation result; obtaining a second cycle consistency loss by reconstructing the inverse transformation result; obtaining a first adversarial loss by discriminating the forward transformation result; obtaining a second adversarial loss by discriminating the inverse transformation result; weighting and combining all obtained losses into a total loss function, and performing alternating optimization training on the model based on the total loss function until convergence, thereby obtaining a trained optimized model; wherein, the weighted combination of all losses includes: a weighted combination of the first feature alignment loss, the second feature alignment loss, the first detail modulation loss, the second detail modulation loss, the first cycle consistency loss, the second cycle consistency loss, the first adversarial loss, and the second adversarial loss.

[0008] In the above implementation process, feature alignment loss and detail modulation loss are introduced into the forward and reverse branches, respectively. These not only constrain the semantic consistency of cross-domain feature distribution but also enhance the fidelity of local texture details. Cyclic consistency loss further suppresses mode collapse, ensuring that the signal can still be reconstructed after round-trip mapping, thereby improving the network's robustness to small changes in seismic waveforms. Adversarial loss uses a discriminator to distinguish between true and false transformation results, driving the generated signal to statistically approximate the real data in the target domain, enhancing the naturalness and spectral continuity of the output waveform. Finally, the eight losses are weighted and fused into a unified objective function according to learnable weights. An alternating optimization strategy of generator and discriminator is adopted, so that the network dynamically balances various constraints during training and gradually converges to the optimum, resulting in a seismic sensor signal optimization model that can effectively suppress noise, improve the signal-to-noise ratio, and maintain effective waveform components.

[0009] Optionally, obtaining the first feature alignment loss and the first detail modulation loss by forward conversion of the low-quality reference signal, and obtaining the second feature alignment loss and the second detail modulation loss by reverse conversion of the high-quality reference signal, includes: converting the low-quality reference signal into a high-quality converted signal and calculating the first feature alignment loss and the first detail modulation loss; converting the high-quality reference signal into a low-quality converted signal and calculating the second feature alignment loss and the second detail modulation loss.

[0010] In the above implementation process, by synchronously calculating the feature alignment loss and detail modulation loss through a low-to-high forward conversion, the network can be forced to retain the micro-amplitude waveform structure while improving the signal-to-noise ratio, thus avoiding excessive smoothing. Correspondingly, by synchronously calculating the two losses through a high-to-low reverse conversion, the network learns to maintain phase consistency in the degraded mapping, thereby enhancing the stability of subsequent reverse reconstruction. The two paths share the same set of encoding and decoding parameters, which reduces redundant calculations and ensures strict alignment of the bidirectional latent space. Ultimately, the trained optimized model has consistent and interpretable suppression capabilities for various real noise scenarios at field stations.

[0011] Optionally, the first feature alignment loss and the second feature alignment loss are calculated based on the optimal transmission distance of the feature distribution during the signal conversion process; the first feature alignment loss represents the feature distribution difference between the high-quality converted signal and the high-quality reference signal, and the second feature alignment loss represents the feature distribution difference between the low-quality converted signal and the low-quality reference signal.

[0012] In the above implementation process, the optimal transmission distance is used to measure the difference in feature distribution. This transforms the distribution distance between high-quality converted signals and high-quality reference signals, and between low-quality converted signals and low-quality reference signals, into a differentiable transportation cost, enabling the network to directly minimize the inter-domain distribution distance during the training phase. This approach avoids the pattern omissions and training oscillations that may result from traditional adversarial losses, and improves the stability and convergence consistency of feature alignment. At the same time, the optimal transmission cost is geometrically sensitive to local feature displacements, and can automatically penalize phase shifts and amplitude distortions while maintaining overall distribution matching. This ensures that the output signal maintains a high degree of consistency with the reference signal in terms of spectral structure, energy distribution, and waveform phase, effectively enhancing the optimization model's ability to reconstruct seismic sensor signals in complex noise environments with high fidelity.

[0013] Optionally, the first detail modulation loss and the second detail modulation loss are calculated based on the local fluctuation energy of the output signal; the first detail modulation loss represents the local fluctuation energy of the high-quality converted signal, and the second detail modulation loss represents the local fluctuation energy of the low-quality converted signal.

[0014] In the above implementation process, using local fluctuation energy as the quantification basis for detail modulation loss, the instantaneous amplitude changes of high-quality and low-quality converted signals can be directly mapped to a differentiable energy index. This allows the network to simultaneously suppress spurious oscillations and retain true microseismic textures during parameter updates. This index is monotonically sensitive to high-frequency components of the waveform and can automatically penalize abnormal energy accumulation caused by noise without introducing additional annotations, while rewarding weak effective signal fluctuations consistent with the reference signal. Since energy calculation only relies on differential statistics within the local sampling window, the computational overhead is low and it is robust to changes in the sampling rate. It can maintain overall distribution alignment while constraining detail fidelity point by point, thereby improving the optimization model's ability to reproduce the details of effective earthquake first arrival, wake attenuation, and fracture microfracture characteristics in complex noise environments in the field.

[0015] Optionally, obtaining the first cycle consistency loss by reconstructing the forward conversion result and obtaining the second cycle consistency loss by reconstructing the reverse conversion result includes: reconstructing the high-quality conversion signal into a low-quality reconstruction signal and calculating the first cycle consistency loss; reconstructing the low-quality conversion signal into a high-quality reconstruction signal and calculating the second cycle consistency loss.

[0016] In the above implementation process, by mapping the high-quality converted signal back to the low-quality reconstructed signal and calculating the first cycle consistency loss, the network can be forced to retain the key phase and travel time information of the original low-quality domain while improving the signal-to-noise ratio, thus avoiding the loss of effective waveforms due to excessive denoising. Correspondingly, reconstructing the low-quality converted signal into a high-quality reconstructed signal and calculating the second cycle consistency loss allows the network to maintain its memory of the amplitude-frequency characteristics of the original high-quality domain when simulating the degradation process, preventing the stability of subsequent forward reconstruction from being reduced due to inverse mapping distortion. The two cyclic paths share the same set of encoding and decoding parameters, forming a closed constraint loop without increasing the additional model capacity, suppressing mode collapse and mapping ambiguity, and enabling the trained optimized model to have robust and interpretable bidirectional reconstruction capabilities for asymmetric noise, instrument response differences, and transmission channel attenuation in real seismic records.

[0017] Optionally, the first cyclic consistency loss represents the difference between the low-quality reconstructed signal and the low-quality reference signal; the second cyclic consistency loss represents the difference between the high-quality reconstructed signal and the high-quality reference signal.

[0018] In the above implementation process, the first cyclic consistency loss directly quantifies the point-by-point differences in amplitude, phase, and travel time between the low-quality reconstructed signal and the original low-quality reference signal, forcing the network to completely retain the effective seismic information of the original record in the closed loop of denoising and reconstruction, and preventing first arrival misalignment and loss of microseismic events caused by excessive smoothing; the second cyclic consistency loss constrains the consistency between the high-quality reconstructed signal and the original high-quality reference signal in the same way, ensuring that the network does not introduce additional distortion during the de-qualitying and reconstruction process, and maintains a stable memory of the real high signal-to-noise ratio waveform; the two losses work together to enable the model to form a self-verification mechanism in bidirectional mapping, reduce the distribution shift caused by the accumulation of one-way transformation errors, and thus improve the reversibility and fidelity of the optimization model in the round-trip processing of seismic waveforms in the complex noise environment of field stations.

[0019] Optionally, obtaining the first adversarial loss by judging the forward conversion result and obtaining the second adversarial loss by judging the reverse conversion result includes: judging the difference between the high-quality converted signal and the high-quality reference signal as the first adversarial loss, and judging the difference between the low-quality converted signal and the low-quality reference signal as the second adversarial loss.

[0020] In the above implementation process, the difference between the high-quality converted signal and the high-quality reference signal is input into the discriminator to calculate the first adversarial loss, and the difference between the low-quality converted signal and the low-quality reference signal is simultaneously input into the same discriminator to calculate the second adversarial loss. This allows the discriminator to continuously receive two pairs of samples in a single training phase, driving the generator to simultaneously reduce the statistical distribution gap of the bidirectional conversion results. This method allows the discriminator parameters to be refreshed alternately by samples from both domains, avoiding inter-domain discrimination bias caused by individual updates and reducing the risk of training oscillations and mode collapse. Since the adversarial target only depends on the sample difference rather than additional annotations, the network can automatically enhance its ability to distinguish between noisy background and effective signal edges while maintaining phase consistency, thereby improving the high-fidelity reproduction performance of the optimized model in terms of amplitude distribution, spectral envelope, and instantaneous energy of real earthquake records in the field.

[0021] Optionally, the step of weighting and combining all the obtained losses into a total loss function, and then alternately optimizing and training the model based on the total loss function until convergence, to obtain a trained optimized model, includes: in the iteration, first fixing the model parameters used for signal discrimination, and updating the model parameters used for signal conversion and reconstruction to minimize the total loss function; then fixing the model parameters used for signal conversion and reconstruction, and updating the model parameters used for signal discrimination to minimize the discrimination loss; and executing the iterative process until the model converges.

[0022] In the above implementation process, the eight-term weighted loss of the generator and the binary classification loss of the discriminator are decoupled in each iteration to avoid gradient conflicts and training oscillations caused by simultaneous updates of the two types of parameters. This approach ensures that the generator is not disturbed by the instantaneous gradient direction of the discriminator when updating parameters, ensuring that feature alignment, detail modulation, cycle consistency and adversarial objectives converge simultaneously, reducing the risk of the model falling into local Nash equilibrium. Correspondingly, the discriminator adjusts the decision boundary only based on the difference between real and fake samples in the independent update phase, maintaining continuous sensitivity to the statistical characteristics of the generated samples, thereby improving the distribution matching accuracy and waveform naturalness of the optimized model at the end of training, and finally obtaining an optimized seismic sensor signal model with robust suppression capability and reproducible convergence in complex outdoor noise environments.

[0023] Secondly, embodiments of this application provide a method for processing seismic sensor signals, the method comprising: acquiring a low-quality seismic sensor signal to be processed; inputting the low-quality seismic sensor signal to be processed into an optimized model obtained by any of the training methods described above; performing feature distribution alignment and detail energy modulation operations on the low-quality signal within the optimized model; and outputting a high-quality seismic signal optimized by feature alignment and detail modulation.

[0024] In the above implementation process, the low-quality seismic sensor signals acquired in real time are first input into the optimized model that has been trained. The model learns the shared latent space through the joint constraints of bidirectional adversarial and cyclic consistency, and automatically maps the noisy waveform to the feature distribution of the clean domain to complete the amplitude-frequency consistency alignment. Then, within the same forward propagation, the detail energy modulation operator is called to perform point-by-point correction on the wave energy within the local window, suppressing abnormal oscillations caused by noise and preserving effective microseismic texture. The entire inference process relies on only a single forward calculation, without additional post-processing or manual threshold adjustment, and can output high-quality seismic signals with the same phase travel time, spectral structure and energy envelope as the real high signal-to-noise ratio record, which improves the reliability and computational efficiency of subsequent phase identification, event location and source parameter inversion.

[0025] Thirdly, this application provides an optimization model for seismic sensor signals, the model comprising: a transformation module, a reconstruction module, a calculation module, a discrimination module, and a training module; the transformation module is configured to perform forward transformation on low-quality reference signals and reverse transformation on high-quality reference signals; the calculation module is configured to calculate a first feature alignment loss, a first detail modulation loss, a first cycle consistency loss, a second feature alignment loss, a second detail modulation loss, and a second cycle consistency loss; the reconstruction module is configured to reconstruct the forward transformation results and the reverse transformation results; the discrimination module is configured to discriminate the forward transformation results to obtain a first adversarial loss and discriminate the reverse transformation results to obtain a second adversarial loss; the training module is configured to weight and combine all obtained losses into a total loss function, and perform alternating optimization training on the model based on the total loss function until convergence, thereby obtaining a trained optimized model; wherein, the weighted combination of all losses includes: a weighted combination of the first feature alignment loss, the second feature alignment loss, the first detail modulation loss, the second detail modulation loss, the first cycle consistency loss, the second cycle consistency loss, the first adversarial loss, and the second adversarial loss.

[0026] In the above implementation process, the transformation module, reconstruction module, calculation module, discriminant module, and training module perform end-to-end collaborative updates with the same total loss function as the sole optimization objective. This ensures that the transformation module is subject to the dual constraints of feature alignment loss and detail modulation loss during both forward and reverse mapping, reducing cross-domain distribution shifts and suppressing spurious oscillations. The reconstruction module completes bidirectional cyclic reconstruction under the same network parameters and directly accepts closed-loop verification of the cyclic consistency loss, avoiding parameter redundancy and error accumulation introduced by additional sub-networks. The calculation module calculates the six differentiable losses at once and shares them with the training module, ensuring that the adversarial loss obtained by the discriminant module matches the generation-side loss in real time in terms of numerical scale and gradient direction, thus preventing scale imbalances common in multi-stage training. The training module couples the eight losses into a single scalar objective through unified weighted combination and uses an alternating generation-discrimination optimization strategy to drive the synchronous convergence of all modules. This results in a final optimized model that possesses robust denoising and detail preservation capabilities for low signal-to-noise ratio seismic records in the field without increasing the number of parameters, inference latency, or memory usage.

[0027] Fourthly, embodiments of this application also provide an electronic device, which includes a memory and a processor. The memory stores program instructions, and when the processor reads and runs the program instructions, it executes the steps in any of the above implementation methods.

[0028] Fifthly, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which, when read and executed by a processor, perform the steps in any of the above implementations. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 A first schematic diagram illustrating the training method for the seismic sensor signal optimization model provided in this application embodiment; Figure 2 A second schematic diagram illustrating the training method for the seismic sensor signal optimization model provided in this application embodiment; Figure 3 This is an actual flowchart of the seismic sensor signal optimization model provided in the embodiments of this application; Figure 4 The third flowchart of the training method for the earthquake sensor signal optimization model provided in the embodiments of this application; Figure 5 The fourth flowchart of the training method for the earthquake sensor signal optimization model provided in the embodiments of this application; Figure 6 The fifth flowchart of the training method for the earthquake sensor signal optimization model provided in the embodiments of this application; Figure 7 This is a flowchart of a method for processing seismic sensor signals provided in an embodiment of this application; Figure 8 This is a schematic diagram of the seismic sensor signal optimization model provided in the embodiments of this application; Figure 9 This is a block flowchart illustrating an electronic device provided in an embodiment of this application.

[0031] Icons: 001-Conversion module; 002-Reconstruction module; 003-Calculation module; 004-Discrimination module; 005-Training module; 100-Electronic device; 111-Memory; 112-Memory controller; 113-Processor; 114-Peripheral interface; 115-Input / output unit; 116-Display unit. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of the embodiments of this application.

[0033] In view of this, the purpose of this application is to provide a training method, processing method, and model for an earthquake sensor signal optimization model, so as to improve the above-mentioned problems existing in the prior art.

[0034] Firstly, please refer to Figure 1 , Figure 1 The first flowchart illustrates the training method for the earthquake sensor signal optimization model provided in this application embodiment.

[0035] This application provides a training method for an optimized model of seismic sensor signals, which is applied to a server. The server can be an electronic device with logical computing functions, such as a personal computer (PC), tablet computer, smartphone, or personal digital assistant (PDA).

[0036] The training method includes: obtaining a first feature alignment loss and a first detail modulation loss through forward transformation of a low-quality reference signal; obtaining a second feature alignment loss and a second detail modulation loss through inverse transformation of a high-quality reference signal; obtaining a first cycle consistency loss through reconstruction of the forward transformation result; obtaining a second cycle consistency loss through reconstruction of the inverse transformation result; obtaining a first adversarial loss by discriminating the forward transformation result; obtaining a second adversarial loss by discriminating the inverse transformation result; weighting all the obtained losses into a total loss function, and then performing alternating optimization training on the model based on this total loss function until convergence, thus obtaining the trained optimized model; wherein, all the weighted combinations of losses include: a weighted combination of the first feature alignment loss, the second feature alignment loss, the first detail modulation loss, the second detail modulation loss, the first cycle consistency loss, the second cycle consistency loss, the first adversarial loss, and the second adversarial loss.

[0037] In the above implementation process, a forward transformation is performed on the low-quality reference signal to generate a corresponding enhanced signal. Based on this, the first feature alignment loss and the first detail modulation loss are calculated, enabling the model to learn the feature mapping from low-quality to high-quality signals. This achieves structural alignment and detail enhancement of the signal in the feature space, improving the model's adaptability to low-quality inputs and reconstruction accuracy. An inverse transformation is performed on the high-quality reference signal to reconstruct the low-quality signal, and the second feature alignment loss and the second detail modulation loss are calculated. This enhances the model's understanding of the quality degradation process, resulting in improved consistency in bidirectional transformations. This improves the model's generalization ability and inverse reasoning capability. The results of both forward and inverse transformations are reconstructed, and the first and second cycle consistency losses are calculated. By forcing the signals before and after the transformation to be cyclically reconstructed, the reliability and information integrity of the transformation process are ensured, reducing the risk of mode collapse and enhancing transformation stability. The forward and reverse transformation results are judged separately, generating a first adversarial loss and a second adversarial loss. Adversarial training improves the realism of the generated results, making the output signal distribution closer to the target domain, which improves the visual and perceptual quality of the signal. All the above losses are weighted and combined according to preset weights to form a total loss function. Based on this function, the generator and discriminator are alternately optimized and trained until the model converges. This takes into account multiple objective constraints, resulting in a balance between various performance aspects of the model, and finally obtaining an optimized model after training. The benefit of joint optimization is improved overall performance and robustness.

[0038] Optionally, please refer to Figure 2 , Figure 2 The second flowchart illustrates the training method for the earthquake sensor signal optimization model provided in this application embodiment.

[0039] The method involves obtaining a first feature alignment loss and a first detail modulation loss by forward conversion of a low-quality reference signal; and obtaining a second feature alignment loss and a second detail modulation loss by inverse conversion of a high-quality reference signal, including: converting the low-quality reference signal into a high-quality converted signal and calculating the first feature alignment loss and the first detail modulation loss; converting the high-quality reference signal into a low-quality converted signal and calculating the second feature alignment loss and the second detail modulation loss.

[0040] In the above implementation, a low-quality reference signal is taken as input and converted into a high-quality transformed signal through its forward generation network. The model simultaneously calculates the first feature alignment loss and the first detail modulation loss, aiming to drive the model to learn the mapping from the low-quality domain to the high-quality domain, generating an output signal that is aligned with the target high-quality signal in high-level features and enhanced in detail texture. This directly constrains the structural fidelity and detail richness of the generated signal, laying the foundation for improving the overall output quality. Possible implementations typically involve using a pre-trained network to extract multi-scale features for alignment comparison and combining pixel-level or gradient-level constraints for detail modulation. Symmetrically, a high-quality reference signal is taken as input and converted into a low-quality transformed signal through its inverse generation network, and the second feature alignment loss and the second detail modulation loss are calculated accordingly. As a mirror process of the forward transformation, the model simultaneously learns the degradation pattern from high-quality signal to low-quality signal, generating a signal that conforms to the distribution of the low-quality domain. By constructing a bidirectional transformation path, the model's understanding of the complex relationship between the two domains is enhanced, which helps to improve the rationality and generalization ability of the forward transformation.

[0041] In one embodiment of this application, please refer to Figure 3 , Figure 3 The actual flowchart of the seismic sensor signal optimization model provided in the embodiments of this application is shown.

[0042] Forward transformation and loss calculation utilize a generator G to receive a low-quality reference signal L, and performs forward transformation through its internal one-dimensional convolutional neural network to output a high-quality transformed signal. The system extracts feature representations from the intermediate hidden layer of generator G, calculates the difference between this feature distribution and the feature distribution of the target high-quality signal based on optimal transmission theory, specifically by solving for the optimal transmission plan (e.g., using the Sinkhorn algorithm to approximate the Wasserstein distance) to obtain the minimum transmission cost, and quantifies this difference as the first feature alignment loss. The first detail modulation loss is calculated by applying the Laplacian operator to the feature layer (or the signal itself) output by generator G to calculate its local fluctuation energy (i.e., Laplacian energy), normalizing it, and generating an energy regularization term based on preset modulation parameters (e.g., kurtosis). This regularization term is used as the first detail modulation loss to constrain the detail richness of the generated signal. Generator F receives the high-quality reference signal H, performs an inverse transformation through its internal one-dimensional convolutional neural network, and outputs a low-quality transformed signal. The system extracts feature representations from the intermediate hidden layer of generator F. Based on optimal transmission theory, it calculates the difference between this feature distribution and the feature distribution of the target low-quality signal. By solving for the optimal transmission plan, it obtains the minimum transmission cost and quantifies this difference as the second feature alignment loss. The system calculates the Laplace energy of the feature layer (or the signal itself) output by generator F. After normalization, it combines the energy with modulation parameters to generate an energy regularization term. This regularization term serves as the second detail modulation loss to constrain the degree of detail degradation in the generated signal. Forward and inverse transformations are performed by generators G and F, respectively. During the transformation process, the feature alignment loss based on optimal transmission and the detail modulation loss based on Laplace energy are calculated simultaneously, thus providing specific supervision signals for feature distribution alignment and detail energy modulation for model training.

[0043] Optionally, the first feature alignment loss and the second feature alignment loss are calculated based on the optimal transmission distance of the feature distribution during the signal conversion process; the first feature alignment loss represents the feature distribution difference between the high-quality converted signal and the high-quality reference signal, and the second feature alignment loss represents the feature distribution difference between the low-quality converted signal and the low-quality reference signal.

[0044] In the above implementation process, the distribution distance between the high-quality transformed signal generated by the forward transformation and the real high-quality reference signal in the deep feature space is calculated, resulting in a scalar value that quantifies the difference in feature distribution between the two high-quality signals, namely the first feature alignment loss. By minimizing the optimal transmission distance, the system not only pursues a simple correspondence of feature points but also strives to achieve overall alignment of the entire feature distribution, thereby more robustly and comprehensively ensuring that the generated signal is consistent with the real high-quality target in semantics and structure. Symmetrically, the optimal transmission distance of the feature distribution between the low-quality transformed signal generated by the inverse transformation and the real low-quality reference signal is calculated, resulting in the second feature alignment loss. This quantifies the degree of matching between the generated low-quality signal and the real low-quality domain at the feature level, ensuring that the inverse degradation process of model learning is realistic and consistent with the actual observed distribution. This, in turn, provides more accurate constraints for the forward transformation, enhancing the reliability and consistency of the bidirectional mapping.

[0045] In one embodiment of this application, similarly, in conjunction with Figure 3 During the forward conversion, the generator G not only outputs the final high-quality conversion signal. Simultaneously, a certain intermediate layer (feature extraction layer) within it outputs the corresponding feature representation. At the same time, the system extracts the corresponding feature representation from the input high-quality reference signal. HExtracting the corresponding feature representations yields two sets of feature vectors or feature maps representing the generated signal and the target real signal, respectively. This provides a foundation for subsequent distribution alignment at the abstract semantic level rather than the original signal level. Treating the two extracted features as two probability distributions, a transmission cost function is defined (usually the square of the Euclidean or cosine distance between feature vectors). The Sinkhorn iterative algorithm is used to efficiently solve for the optimal transmission plan between these two distributions, ultimately calculating the approximate optimal transmission distance (i.e., Wasserstein distance), resulting in a scalar value that quantifies the overall difference between the feature distribution of the generated high-quality converted signal and the feature distribution of the real high-quality reference signal. This scalar value is defined as the first feature alignment loss. Its advantage is that it achieves distribution-level alignment, forcing the generator to learn to produce signals that are highly consistent with the target domain in overall statistical properties, rather than merely pursuing point-wise similarity, which enhances the robustness and generalization of the conversion. Symmetrically, during the inverse conversion, the generator F outputs a low-quality converted signal. The system extracts the feature representation of the low-quality input reference signal L, and simultaneously extracts the feature representation of the low-quality input reference signal L. The system then executes the exact same optimal transmission distance calculation process to calculate the optimal transmission distance between the feature distribution of the generated low-quality converted signal and the feature distribution of the real low-quality reference signal; this distance is defined as the second feature alignment loss. The advantage of this is that it ensures that the degradation process of the model learning (from high quality to low quality) also conforms to the distribution law of the real low-quality signal, thus providing a robust feature-level constraint for the consistency of the entire bidirectional loop.

[0046] In one specific embodiment of this application, firstly, the feature representation of the low-quality input extracted by the generator G is calculated. and the feature representation extracted from high-quality input by generator F ( and This represents a hidden layer mapping of the generator. Then, the solution is obtained through optimal transport. and The optimal match is obtained by converting the matching result into a supervision signal. For example, the Euclidean distance or sum of squared differences between the corresponding features of the match is calculated and used as the loss to minimize it. This embodiment defines... ,in Indicates high-quality features Mapping to The corresponding transport operator. Minimize Features that prompt generator G to generate It can closely approximate a high-quality feature, thereby improving the accuracy of cross-domain mapping.

[0047] Optionally, the first detail modulation loss and the second detail modulation loss are calculated based on the local fluctuation energy of the output signal; the first detail modulation loss represents the local fluctuation energy of the high-quality converted signal, and the second detail modulation loss represents the local fluctuation energy of the low-quality converted signal.

[0048] In the above implementation process, for the high-quality converted signal obtained by forward conversion, its local fluctuation energy is calculated using specific operators (such as gradient, high-pass filtering, or wavelet transform), and this energy is used as the first detail modulation loss to obtain a scalar index of the richness of detail and texture complexity of the generated high-quality signal. By explicitly measuring and constraining this energy, the model can be directly incentivized to synthesize or recover appropriate details and textures that conform to the statistical characteristics of natural images or signals during the conversion process, avoiding overly smooth results or the generation of unrealistic noise textures. Symmetrically, for the low-quality converted signal obtained by inverse conversion, the system calculates its local fluctuation energy, which is used as the second detail modulation loss to quantify the level of detail fluctuations contained in the degraded signal generated by the model. This ensures that the model, during inverse conversion, not only simulates the overall quality degradation but also accurately simulates the detail loss or blurring patterns unique to low-quality signals, i.e., the local fluctuation energy should be reduced to a level consistent with the real low-quality domain, thus making the modeling of bidirectional conversion more accurate.

[0049] In one embodiment of this application, the Laplacian operator is applied to the high-quality converted signal (or its designated feature layer) output by generator G to calculate the second-order difference at each local location. Subsequently, the system averages the squares of these difference values ​​to obtain the Laplacian energy of the signal, yielding a scalar value that directly quantifies the richness of internal detail and the degree of textural variation in the generated high-quality signal. The Laplacian energy calculated in the previous step is normalized so that its value falls within a standard range (e.g., between 0 and 1). Simultaneously, the system determines a modulation parameter based on the statistical characteristics of the target high-quality signal or the feature itself (e.g., calculating the kurtosis of its feature distribution). Then, the system generates an energy regularization term using a specific loss function (e.g., a negative logarithmic penalty term based on the normalized energy and the modulation parameter). This step results in a first detail modulation loss. This loss function is designed such that the loss value increases significantly when the detail energy of the generated signal deviates from a “moderate” range determined by the modulation parameter and the characteristics of the original signal. This approach encourages the generator to inject necessary detail energy into smooth signal regions to avoid over-smoothing, while also preventing the introduction of excessive noise or unstable oscillations into regions that are already highly volatile, thus achieving a global balance between detail enhancement and noise suppression.

[0050] Symmetrically, the system performs the exact same local fluctuation energy calculation and adaptive regularization term generation process on the low-quality converted signal output by generator F. This step results in a second detail modulation loss. Its advantage is that it ensures the model accurately modulates away an appropriate amount of detail energy during the simulated signal degradation process, allowing the generated low-quality signal to match the real low-quality domain signal in terms of detail roughness or smoothness, thus making both bidirectional detail modulation physically meaningful and realistic.

[0051] In one specific embodiment of this application, the normalized Laplace energy is calculated for the feature map output by the generator G. Substitute into the formula Obtain the regularized loss value ( (The modulation coefficient). This regularization term is added to the generator's loss function to encourage the generator G to output characteristic patterns with moderate fluctuations, avoiding over-smoothing or oscillations. This regularization term is equivalent to imposing a soft constraint on the spectral flatness of the generator G's output: when the output is too flat and lacks detail (…), it is considered less likely to cause problems. If the output is too low, the increased loss will drive the generator G to introduce more high-frequency details; when the output is too volatile and contains noise spikes ( If the value is too high, the loss will increase, thus suppressing the excessive details introduced by the generator G.

[0052] Optionally, please refer to Figure 4 , Figure 4 The third flowchart is a training method for the earthquake sensor signal optimization model provided in the embodiments of this application.

[0053] The first cyclic consistency loss is obtained by reconstructing the forward conversion result; the second cyclic consistency loss is obtained by reconstructing the reverse conversion result, including: reconstructing the high-quality conversion signal into a low-quality reconstruction signal and calculating the first cyclic consistency loss; reconstructing the low-quality conversion signal into a high-quality reconstruction signal and calculating the second cyclic consistency loss.

[0054] In the above implementation process, the model inputs the high-quality transformed signal obtained from the forward transformation back into the inverse transformation network to reconstruct a low-quality reconstructed signal. Subsequently, the system calculates the difference between this reconstructed signal and the original low-quality reference signal, using this difference as the first cyclic consistency loss. This effectively constrains the forward transformation process from arbitrarily altering or losing the core content of the original signal, as the model must ensure that this content can be accurately recovered in reverse, greatly enhancing the reliability and credibility of the transformation. Symmetrically, the model inputs the low-quality transformed signal obtained from the inverse transformation back into the forward transformation network to reconstruct a high-quality reconstructed signal, calculating the difference between this reconstructed signal and the original high-quality reference signal as the second cyclic consistency loss. This ensures that the inverse transformation is controllable and has clear semantics, rather than random information destruction, while further consolidating the consistency of the bidirectional mapping, forming a closed constraint loop together with the first cyclic consistency loss.

[0055] In one embodiment of this application, generator F receives a high-quality conversion signal obtained by forward conversion from generator G. As input, an inverse transformation operation is performed to reconstruct it into a low-quality reconstructed signal. Then, this reconstructed signal is calculated. The difference between the original low-quality reference signal L and the original signal, typically measured using the L1 norm or mean squared error (MSE) in the signal space or feature space, ensures that the forward conversion (quality improvement) process does not arbitrarily distort or discard the essential content of the original signal. This is because the model must guarantee that this content can be accurately recovered inversely, thus greatly enhancing the reliability and content preservation capability of the conversion process. Symmetrically, generator G receives the low-quality converted signal obtained by the inverse conversion of generator F. As input, a forward conversion operation is performed to reconstruct it into a high-quality reconstructed signal. Calculate high-quality reconstructed signals The difference between the original high-quality reference signal H and the original signal is also measured using L1 loss and other methods. On the one hand, this verifies the rationality and controllability of the inverse transformation (degradation simulation); on the other hand, it, together with the first loss term, forms a closed-loop constraint, ensuring the overall consistency and stability of the bidirectional mapping from two directions, effectively preventing training collapse.

[0056] Optionally, the first cyclic consistency loss represents the difference between the low-quality reconstructed signal and the low-quality reference signal; the second cyclic consistency loss represents the difference between the high-quality reconstructed signal and the high-quality reference signal.

[0057] In the above implementation process, the first cycle consistency loss is explicitly defined and calculated as the difference between the low-quality reconstructed signal and the original low-quality reference signal. This imposes a hard constraint on the forward transformation (i.e., the quality improvement process), requiring that the process must be a conditional enhancement, meaning that it cannot lose or distort the underlying information necessary for reconstructing the original input in the inverse transformation. This effectively prevents arbitrary patterns in the generated results that are unrelated to the input content. Symmetrically, the system explicitly defines and calculates the second cycle consistency loss as the difference between the high-quality reconstructed signal and the original high-quality reference signal. On the one hand, it ensures that the inverse transformation, i.e., the simulated degradation process, is controllable and physically meaningful, rather than an irreversible destruction. On the other hand, it works synergistically with the first loss to form a closed, bidirectional content consistency constraint loop, greatly improving the stability of the entire model training and the reliability of the output results.

[0058] In one embodiment of this application, generator F receives a high-quality conversion signal output by generator G. It then performs an inverse transformation to reconstruct a low-quality reconstructed signal. The reconstructed low-quality signal is then compared point-by-point or feature-level with the original low-quality reference signal input to generator G, and the difference between the two is calculated, typically using the L1 norm. This difference is explicitly quantified as the first cycle consistency loss. A strict content conservation constraint is set for the forward conversion process (low-quality to high-quality), requiring that the boosted signal must retain all core information that can be accurately traced back to the original input, thereby effectively preventing the generator from generating arbitrary illusions or distortions unrelated to the source content, ensuring the high reliability of the enhancement process. Symmetrically, generator G receives the low-quality converted signal output by generator F. It is then positively transformed and reconstructed into a high-quality reconstructed signal. The high-quality reconstructed signal obtained from the reconstruction The difference is calculated by comparing the signal with the original high-quality reference signal H input to the generator F, often using the L1 norm. This difference is explicitly quantified as the second cyclic consistency loss. First, it verifies the rationality and reversibility of the reverse transformation (simulated degradation) process, ensuring that the degradation simulation is not random destruction. Second, together with the first cyclic consistency loss, it forms a closed-loop, bidirectional content consistency verification loop, greatly improving the stability of the entire model training and the fidelity of the learned mapping.

[0059] In one specific embodiment of this application, the cycle consistency loss The constraint ensures that the original sample can be reconstructed after bidirectional mapping, thus guaranteeing the low-quality transformed signal output by generator F. and the high-quality conversion signal output by generator G .use Norm calculation of cyclic reconstruction error ,in, The low-quality converted signal output by generator F , High-quality conversion signal output by generator G This ensures that the overall waveform shape of the signal is maintained during the round-trip mapping.

[0060] Optionally, please refer to Figure 5 , Figure 5 The fourth flowchart is a training method for the earthquake sensor signal optimization model provided in the embodiments of this application.

[0061] The first adversarial loss is obtained by judging the forward conversion result, and the second adversarial loss is obtained by judging the reverse conversion result, including: judging the difference between the high-quality converted signal and the high-quality reference signal as the first adversarial loss, and judging the difference between the low-quality converted signal and the low-quality reference signal as the second adversarial loss.

[0062] In the above implementation, the discriminator is responsible for distinguishing the high-quality transformed signal generated by the generator from the real high-quality reference signal. The result of this adversarial process is a first adversarial loss, which quantifies the difficulty of the generated high-quality signal being misclassified as real by the discriminator, or in other words, the difference between it and the distribution of real high-quality data. Its advantage lies in the fact that it does not rely on a predefined, fixed similarity metric, but rather, through adversarial competition, incentivizes the generator to produce outputs that are statistically indistinguishable from the set of real high-quality signals, thereby greatly improving the overall realism and naturalness of the generated results. Symmetrically, another discriminator (or one with the same shared parameters) is responsible for distinguishing the low-quality transformed signal generated by the generator from the real low-quality reference signal, and calculates a second adversarial loss accordingly. This ensures that when the model simulates the signal degradation process (inverse transformation), its output also conforms to the distribution characteristics of real low-quality data. On the one hand, it makes the results of the inverse transformation more realistic, thus providing a more reliable signal for the cycle consistency constraint in the forward transformation; on the other hand, by forcing the generator to deeply understand the data characteristics of the low-quality domain, it indirectly improves the targeting and effectiveness of the forward transformation (quality enhancement).

[0063] In one embodiment of this application, the discriminator It is responsible for receiving and evaluating two types of signals, one of which is the high-quality converted signal generated by the generator G. Another type is a high-quality reference signal H from a real dataset. (Discriminator) The system determines whether the input signal is genuine or not, and its output is a scalar value or a set of local discriminant values ​​representing the probability that the signal is "real". Based on the discriminator's output, the system calculates the first adversarial loss using an adversarial objective function (such as least squares loss or standard cross-entropy loss). This loss quantifies the generated signal. The difficulty for the discriminator to detect is determined by the degree of difference between the output and the real high-quality data distribution. A training objective for the generator G is provided based on the realism of the data distribution rather than a fixed metric. By minimizing this adversarial loss, the generator is driven to produce outputs that are statistically indistinguishable from real high-quality signals, thereby significantly improving the overall naturalness, realism, and perceptual quality of the augmented results. Symmetrically, the discriminator... It is responsible for receiving and evaluating two types of signals, one of which is the low-quality converted signal generated by generator F. Another type is a low-quality reference signal L from a real dataset, which ensures that the output of the inverse transformation (degradation simulation) conforms to the observed distribution of the real low-quality signal, making the simulated degradation process more realistic. Secondly, a high-quality inverse transformation is crucial for the cyclic consistency constraint of the forward transformation. By providing a more realistic pseudo-low-quality signal, it indirectly constrains and improves the performance of the forward generator G.

[0064] In a specific case, including and Two adversarial terms, the former constraining the generated With true high quality In the discriminator It appears indistinguishable, the latter constraining With real low quality exist There is no distinction between them. To improve training stability, an LSGAN (least squares GAN) loss form is used, whereby the discriminator minimizes... Generator Minimization (The same applies to low-quality products).

[0065] Optionally, please refer to Figure 6 , Figure 6 The fifth flowchart illustrates the training method for the earthquake sensor signal optimization model provided in this application embodiment.

[0066] All the obtained losses are weighted and combined into a total loss function, and the model is alternately optimized and trained based on this total loss function until convergence, resulting in a trained optimized model. This includes: in the iteration, first fixing the model parameters used for signal discrimination, and updating the model parameters used for signal conversion and reconstruction to minimize the total loss function; then fixing the model parameters used for signal conversion and reconstruction, and updating the model parameters used for signal discrimination to minimize the discrimination loss; and executing the iterative process until the model converges.

[0067] In the above implementation process, eight independent losses (i.e., first feature alignment loss, second feature alignment loss, first detail modulation loss, second detail modulation loss, first cycle consistency loss, second cycle consistency loss, first adversarial loss, and second adversarial loss) calculated from four dimensions—feature alignment, detail modulation, cycle consistency, and adversarial discrimination—are assigned pre-defined weight coefficients based on their respective contributions and importance. These losses are then combined into a unified total loss function through weighted summation, constructing a comprehensive and differentiable optimization objective that balances the multiple requirements of content fidelity, detail quality, information consistency, and distribution realism in signal conversion tasks. Through a concise mathematical framework, different and potentially competing optimization objectives are systematically coordinated, enabling the model to learn towards a balanced and comprehensive optimal solution. By weighted combination of the total loss function and rigorous alternating optimization training, this implementation transforms the complex multi-task learning problem into a stable and controllable end-to-end optimization process. It not only ensures that all the technical goals of the design (alignment, detail, consistency, realism) are effectively reflected in the final model, but also improves the overall performance ceiling of the model through the game mechanism of adversarial training. It is the key guarantee and driving engine for the successful operation of the whole method.

[0068] In one embodiment of this application, the weighted sum of the various losses constitutes the final objective function:

[0069] in, , , These are weight hyperparameters used to balance the impact of each loss term on training. Furthermore, and These are the losses from the first confrontation and the losses from the second confrontation, respectively. It consists of the first cycle consistency loss and the second cycle consistency loss; It consists of the first feature alignment loss and the second feature alignment loss; It consists of the first detail modulation loss and the second detail modulation loss.

[0070] In this embodiment, both cycle consistency loss and adversarial loss are important for preserving the basic waveform and style transfer. Taking larger values ​​ensures the waveform is not distorted; OTS and MLE are newly introduced terms, and their weights can be adjusted based on the validation set results to achieve the best balance. During training, minimizing... Train generators G and F while maximizing the adversarial loss of the discriminator during training. , Until all losses are mitigated.

[0071] Secondly, please refer to Figure 7 , Figure 7 This is a flowchart illustrating a method for processing seismic sensor signals provided in an embodiment of this application.

[0072] This application provides a method for processing seismic sensor signals, the method comprising: acquiring a low-quality seismic sensor signal to be processed; inputting the low-quality seismic sensor signal to be processed into an optimized model obtained by any of the above training methods; performing feature distribution alignment and detail energy modulation operations on the low-quality signal within the optimized model; and outputting a high-quality seismic signal optimized by feature alignment and detail modulation.

[0073] In the above implementation process, a feature distribution alignment mechanism is introduced, which effectively reduces the signal distribution offset caused by sensor differences or environmental interference. At the same time, by using detail energy modulation operation, the effective components in the signal are enhanced and noise is suppressed, thereby improving the overall usability and analytical accuracy of the seismic signal. End-to-end optimization from low-quality raw signal to high-quality output signal is achieved, providing a reliable foundation for subsequent seismic data analysis and interpretation.

[0074] Thirdly, please refer to Figure 8 , Figure 8 This is a schematic diagram of the seismic sensor signal optimization model provided in the embodiments of this application.

[0075] This application provides an optimization model for seismic sensor signals, comprising: a transformation module 001, a reconstruction module 002, a calculation module 003, a discrimination module 004, and a training module 005; the transformation module 001 is configured to perform a forward transformation on low-quality reference signals and a reverse transformation on high-quality reference signals; the calculation module 003 is configured to calculate a first feature alignment loss, a first detail modulation loss, a first cycle consistency loss, a second feature alignment loss, a second detail modulation loss, and a second cycle consistency loss; the reconstruction module 002 is configured to reconstruct the forward transformation results and the reverse transformation results. The discrimination module 004 is configured to discriminate the forward conversion result to obtain the first adversarial loss, and to discriminate the reverse conversion result to obtain the second adversarial loss; the training module 005 is configured to weight and combine all the obtained losses into a total loss function, and to perform alternating optimization training on the model based on the total loss function until convergence, thereby obtaining the trained optimized model; wherein, the weighted combination of all losses includes: the weighted combination of the first feature alignment loss, the second feature alignment loss, the first detail modulation loss, the second detail modulation loss, the first cycle consistency loss, the second cycle consistency loss, the first adversarial loss, and the second adversarial loss.

[0076] In the above implementation process, by introducing feature alignment and detail modulation losses, the semantic features of the signal are preserved and the detail energy is enhanced during the conversion process; cyclic consistency loss ensures the bidirectional reliability of the conversion path; and adversarial loss further drives the generated signal to converge with the target high-quality signal in terms of distribution. The modules work together, and under the framework of multi-task joint optimization, they jointly realize the high-fidelity and high-robust conversion of seismic sensor signals from low quality to high quality.

[0077] Fourthly, embodiments of this application also provide an electronic device, which includes a memory and a processor. The memory stores program instructions, and when the processor reads and runs the program instructions, it executes the steps in any of the above implementation methods.

[0078] Fifthly, embodiments of this application also provide a computer-readable storage medium storing computer program instructions, which are read and executed by a processor to perform the steps in any of the above implementations.

[0079] Optionally, please refer to Figure 9 , Figure 9 This is a block flowchart illustrating an electronic device according to an embodiment of this application. The electronic device 100 may include a memory 111, a memory controller 112, a processor 113, a peripheral interface 114, an input / output unit 115, and a display unit 116. Those skilled in the art will understand that... Figure 8 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device 100. For example, the electronic device 100 may also include components that are more... Figure 8 The more or fewer components shown, or having the same Figure 8 The different configurations shown.

[0080] The aforementioned memory 111, memory controller 112, processor 113, peripheral interface 114, input / output unit 115, and display unit 116 are electrically connected directly or indirectly to each other to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The aforementioned processor 113 is used to execute executable modules stored in the memory.

[0081] The memory 111 can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory 111 stores programs, and the processor 113 executes these programs upon receiving execution instructions. The methods executed by the electronic device 100 as defined in any embodiment of this application can be applied to the processor 113, or implemented by the processor 113.

[0082] The aforementioned processor 113 may be an integrated circuit chip with signal processing capabilities. The processor 113 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a digital signal processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor.

[0083] The peripheral interface 114 described above couples various input / output devices to the processor 113 and the memory 111. In some embodiments, the peripheral interface 114, the processor 113, and the memory controller 112 can be implemented on a single chip. In other instances, they can be implemented on separate chips.

[0084] The input / output unit 115 described above is used to provide user input data. The input / output unit 115 may be, but is not limited to, a mouse and keyboard, etc.

[0085] The aforementioned display unit 116 provides an interactive interface (e.g., a user interface) between the electronic device 100 and the user, or displays image data for the user's reference. In this embodiment, the display unit can be a liquid crystal display (LCD) or a touch display. If it is a touch display, it can be a capacitive touchscreen or a resistive touchscreen that supports single-point and multi-point touch operations. Supporting single-point and multi-point touch operations means that the touch display can sense touch operations generated simultaneously from one or more locations on the touch display and pass the sensed touch operations to the processor for calculation and processing.

[0086] In summary, this application provides a training method, processing method, and model for an optimized seismic sensor signal, relating to the field of seismic signal optimization technology. The training method includes: obtaining a first feature alignment loss and a first detail modulation loss through forward transformation of a low-quality reference signal; obtaining a second feature alignment loss and a second detail modulation loss through inverse transformation of a high-quality reference signal; obtaining a first cycle consistency loss through reconstruction of the forward transformation result; obtaining a second cycle consistency loss through reconstruction of the inverse transformation result; obtaining a first adversarial loss by discriminating the forward transformation result; obtaining a second adversarial loss by discriminating the inverse transformation result; weighting and combining all obtained losses into a total loss function, and then alternately optimizing the model based on this total loss function until convergence, resulting in a trained optimized model. By constructing a multi-level, bidirectional cyclic training framework integrating feature alignment, detail modulation, cycle consistency, and adversarial discrimination mechanisms, compared to traditional methods, it has comprehensive advantages in improving seismic signal quality, including systematicity, efficiency, and high fidelity.

[0087] In the several embodiments provided in this application, it should be understood that the disclosed device can also be implemented in other ways. The device embodiments described above are merely illustrative; for example, the block diagrams in the accompanying drawings illustrate the possible architecture, functions, and operations of the device according to various embodiments of this application. In this regard, each block in the block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram, and combinations of block diagrams, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0088] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0089] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0090] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. 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.

[0091] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

[0092] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A training method for an optimization model of seismic sensor signals, characterized in that, The training method includes: The first feature alignment loss and the first detail modulation loss are obtained by forward transformation of the low-quality reference signal; the second feature alignment loss and the second detail modulation loss are obtained by inverse transformation of the high-quality reference signal. The first cycle consistency loss is obtained by reconstructing the forward transformation result; the second cycle consistency loss is obtained by reconstructing the reverse transformation result; the first adversarial loss is obtained by discriminating the forward transformation result; and the second adversarial loss is obtained by discriminating the reverse transformation result. All the obtained losses are weighted and combined into a total loss function, and the seismic sensor signal optimization model is alternately optimized and trained based on the total loss function until convergence, thus obtaining the trained optimization model. The weighted combination of all losses includes: a weighted combination of the first feature alignment loss, the second feature alignment loss, the first detail modulation loss, the second detail modulation loss, the first cycle consistency loss, the second cycle consistency loss, the first adversarial loss, and the second adversarial loss.

2. The method according to claim 1, characterized in that, The process of obtaining a first feature alignment loss and a first detail modulation loss through a forward transformation of a low-quality reference signal, and obtaining a second feature alignment loss and a second detail modulation loss through an inverse transformation of a high-quality reference signal, includes: The low-quality reference signal is converted into a high-quality converted signal, and the first feature alignment loss and the first detail modulation loss are calculated; the high-quality reference signal is converted into a low-quality converted signal, and the second feature alignment loss and the second detail modulation loss are calculated.

3. The method according to claim 2, characterized in that, in, The first feature alignment loss and the second feature alignment loss are calculated based on the optimal transmission distance of the feature distribution during the signal conversion process; the first feature alignment loss represents the feature distribution difference between the high-quality converted signal and the high-quality reference signal, and the second feature alignment loss represents the feature distribution difference between the low-quality converted signal and the low-quality reference signal.

4. The method according to claim 2, characterized in that, in, The first detail modulation loss and the second detail modulation loss are calculated based on the local fluctuation energy of the output signal; the first detail modulation loss represents the local fluctuation energy of the high-quality converted signal, and the second detail modulation loss represents the local fluctuation energy of the low-quality converted signal.

5. The method according to claim 2, characterized in that, The process of obtaining a first cycle consistency loss by reconstructing the result of the forward transformation and a second cycle consistency loss by reconstructing the result of the reverse transformation includes: The high-quality converted signal is reconstructed into a low-quality reconstructed signal, and the first cycle consistency loss is calculated; the low-quality converted signal is reconstructed into a high-quality reconstructed signal, and the second cycle consistency loss is calculated.

6. The method according to claim 5, characterized in that, in, The first cyclic consistency loss represents the difference between the low-quality reconstructed signal and the low-quality reference signal; the second cyclic consistency loss represents the difference between the high-quality reconstructed signal and the high-quality reference signal.

7. The method according to claim 2, characterized in that, The process of obtaining a first adversarial loss by discriminating the forward conversion result and a second adversarial loss by discriminating the reverse conversion result includes: The difference between the high-quality converted signal and the high-quality reference signal is determined as a first adversarial loss, and the difference between the low-quality converted signal and the low-quality reference signal is determined as a second adversarial loss.

8. The method according to claim 1, characterized in that, The process of weighting and combining all obtained losses into a total loss function, and then alternately optimizing and training the model based on this total loss function until convergence, to obtain the trained optimized model, includes: In the iteration, the model parameters used for signal discrimination are first fixed, and the model parameters used for signal conversion and reconstruction are updated to minimize the total loss function; then the model parameters used for signal conversion and reconstruction are fixed, and the model parameters used for signal discrimination are updated to minimize the discrimination loss; the iterative process is executed until the model converges.

9. A method for processing seismic sensor signals, characterized in that, The method includes: Acquire low-quality seismic sensor signals to be processed; The low-quality seismic sensor signal to be processed is input into the optimized model obtained by the training method as described in any one of claims 1 to 8; Within the optimization model, feature distribution alignment and detail energy modulation operations are performed on the low-quality signal; Outputs high-quality seismic signals optimized by feature alignment and detail modulation.

10. A seismic sensor signal optimization model, characterized in that, The model includes: a transformation module, a reconstruction module, a calculation module, a discrimination module, and a training module; The conversion module is configured to perform forward conversion on low-quality reference signals and reverse conversion on high-quality reference signals; The calculation module is configured to calculate a first feature alignment loss, a first detail modulation loss, a first cycle consistency loss, a second feature alignment loss, a second detail modulation loss, and a second cycle consistency loss; The reconstruction module is configured to reconstruct the forward conversion result and the reverse conversion result; The discrimination module is configured to discriminate the forward conversion result to obtain a first adversarial loss, and to discriminate the reverse conversion result to obtain a second adversarial loss; The training module is configured to weight and combine all the obtained losses into a total loss function, and then perform alternating optimization training on the model based on the total loss function until convergence, thereby obtaining the optimized model after training. The weighted combination of all losses includes: a weighted combination of the first feature alignment loss, the second feature alignment loss, the first detail modulation loss, the second detail modulation loss, the first cycle consistency loss, the second cycle consistency loss, the first adversarial loss, and the second adversarial loss.

11. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing program instructions, and the processor executing the program instructions to perform the steps of the method according to any one of claims 1 to 8 or 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions that, when executed by a processor, perform the steps of the method according to any one of claims 1 to 8 or 9.

Images