Seabed seismic signal denoising enhancement method based on deep residual network

Abstract

The invention discloses a deep residual network-based submarine seismic signal denoising enhancement method. The method comprises the following steps of decomposing an original signal into a plurality of sub-band signals through a wavelet packet; performing feature extraction on each sub-band signal by using a deep residual convolutional network in combination with a frequency domain attention mechanism; network hyper-parameters are optimized by using an improved marine predator algorithm, and each sub-band feature map is fused; enhancing the signal through short-time Fourier transform and a dynamic gating threshold; periodically, adaptively and finely adjusting network parameters; and adjusting the network according to the signal-to-noise ratio gain, the weak seismic event detection success rate and the positioning error index feedback. According to the invention, the signal-to-noise ratio of the seismic signal and the event detection precision in the complex marine environment are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of submarine seismic signal processing technology, and in particular to a method for denoising and enhancing submarine seismic signals based on deep residual networks. Background Technology

[0002] In recent years, with the increasing demand for marine resource development and marine disaster monitoring, seafloor seismic monitoring technology has become an important research direction in the field of marine geophysics. Seafloor seismic monitoring technology mainly acquires seismic signal data through arrays of seismic sensors deployed on the seabed. Effective denoising and enhancement of these seismic signals is crucial for accurately extracting weak seismic events and improving the accuracy of seismic event location. Currently, conventional methods for denoising and enhancing seafloor seismic signals mainly include wavelet transform, empirical mode decomposition, and adaptive filtering methods based on traditional signal processing techniques, as well as denoising and enhancement methods based on deep learning algorithms that have gradually emerged in recent years.

[0003] Traditional signal processing techniques often employ fixed filter parameters or frequency domain feature extraction methods, such as wavelet thresholding for denoising and empirical mode decomposition for signal reconstruction. However, due to the complexity of the marine environment and the significant differences in the types and spectral distributions of various noise sources, traditional methods cannot adaptively handle diverse noise characteristics, leading to the easy submersion of weak seismic events and difficulty in effectively improving the signal-to-noise ratio and detection accuracy of seismic signals. In recent years, deep learning-based denoising and enhancement techniques have achieved some success by automatically learning seismic signal features through data-driven approaches. However, most existing deep learning models suffer from problems such as complex structures, strong subjectivity in hyperparameter selection, insufficient network generalization performance, and difficulty in quickly adapting to changes in the marine environment, especially in handling multi-scale and multi-frequency band signal features. Furthermore, although a number of network parameter optimization methods based on swarm intelligence algorithms have emerged in recent years, such as particle swarm optimization, gray wolf algorithms, and the standard marine predator algorithm, these algorithms have certain contradictions in population initialization, search step size setting, and algorithm convergence speed and accuracy, making it difficult to efficiently and accurately find the optimal hyperparameter combination of the network, thus resulting in limited improvement in seismic signal denoising and enhancement performance.

[0004] Therefore, how to provide a method for denoising and enhancing submarine seismic signals based on deep residual networks is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for denoising and enhancing submarine seismic signals based on deep residual networks. Addressing the problems of low signal-to-noise ratio, difficulty in accurately extracting microseismic signals, and low positioning accuracy in submarine seismic monitoring, this invention proposes a method employing wavelet packet decomposition, multi-scale residual convolution feature extraction, improved marine predator algorithm hyperparameter optimization, frequency domain dynamic gating reconstruction, adaptive online fine-tuning, and feedback optimization. This invention possesses the technical advantages of significantly improving the signal-to-noise ratio and the accuracy of weak seismic event detection.

[0006] A method for denoising and enhancing submarine seismic signals based on deep residual networks according to an embodiment of the present invention includes:

[0007] The original submarine seismic waveform signal is decomposed into multiple sub-band signals of different frequencies using wavelet packets;

[0008] For each sub-band signal, a deep residual convolutional network composed of convolutional kernels of different sizes is used to extract features in parallel, and a frequency domain attention mechanism is used to improve the feature weights of the effective seismic signals to obtain feature maps of each sub-band.

[0009] We initialized the population of the improved marine predator algorithm using chaotic mapping, updated the population position through Levy flight adaptive step size and back learning strategy, and optimized the convolutional kernel size, network skip connection weight coefficient and dynamic gating threshold.

[0010] Based on the optimized combination of network hyperparameters, the feature maps of each sub-band are fused and the preliminary denoised time-domain signal is obtained through inverse wavelet packet transform.

[0011] The initially denoised time-domain signal is reconstructed in the frequency domain through short-time Fourier transform, and the dynamic gating threshold is optimized in real time using the improved marine predator algorithm to enhance the effective seismic signal.

[0012] Based on real-time acquired seabed seismic signal fragments, the parameters of the depth residual convolutional network are periodically adjusted using an improved marine predator algorithm to dynamically respond to changes in the marine environment.

[0013] The enhanced seismic signal is evaluated based on the signal-to-noise ratio gain, the success rate of weak seismic event detection, and the positioning error index, and the parameters of the depth residual convolutional network are adjusted accordingly.

[0014] Optionally, the step of decomposing the original seafloor seismic waveform signal into multiple sub-band signals of different frequencies using wavelet packets specifically involves:

[0015] The original submarine seismic waveform signal is divided into several signal segments of equal length, and the Shannon entropy value of each signal segment is calculated separately.

[0016] Based on the calculated Shannon entropy value of each signal segment, the signal segment with the highest entropy value is selected as the reference segment. The Daubechies wavelet basis function is used to perform the first binary recursive wavelet packet decomposition to obtain multiple initial frequency band sub-signals.

[0017] Calculate the Shannon entropy value of the initial frequency band sub-signals respectively, and determine whether each frequency band sub-signal should continue to be decomposed into wavelet packet at the next scale. Stop the decomposition when the Shannon entropy value of all frequency band sub-signals is lower than or equal to the preset entropy threshold.

[0018] For the remaining signal segments with lower entropy values, wavelet packet decomposition is performed synchronously using the same scale and frequency band division method to obtain multiple sub-frequency band signals of different frequencies;

[0019] All sub-band signals are labeled with their corresponding frequency ranges and arranged in ascending order of frequency.

[0020] In the permuted and combined sub-band signals, the Shannon entropy value of each sub-band signal is calculated again, and the sub-band signals in a specific frequency range whose entropy value exceeds the set threshold are subjected to additional-scale local recursive wavelet packet decomposition to obtain the final sub-band signals with lower entropy values.

[0021] Optionally, for each sub-frequency band signal, parallel feature extraction is performed using deep residual convolutional networks composed of convolutional kernels of different sizes, and a frequency domain attention mechanism is used to improve the feature weights of the effective seismic signals to obtain feature maps for each sub-frequency band, specifically:

[0022] For each sub-band signal, the sub-band signal is transformed to the frequency domain through a frequency domain transformation operation to obtain the corresponding initial frequency domain characteristics;

[0023] For the initial frequency domain features of each sub-band signal, the amplitude features are calculated respectively, and global average pooling and global max pooling operations are performed based on the amplitude features to calculate the frequency domain attention weight value corresponding to each sub-band signal;

[0024] Based on the distribution characteristics of the calculated frequency domain attention weight values, the convolution kernel size corresponding to each sub-band signal is determined, and the determined convolution kernel size is used to construct the deep residual convolutional network corresponding to each sub-band signal.

[0025] Each sub-band signal is passed through a depth residual convolutional network with the corresponding kernel size and subjected to at least two convolution operations to extract temporal features.

[0026] The temporal features extracted by the convolution operation are multiplied element-wise with the previously obtained frequency domain attention weights to obtain the temporal enhanced features after processing by the frequency domain attention mechanism.

[0027] The time-domain enhanced features processed by the frequency-domain attention mechanism are then returned to the frequency domain through an inverse frequency-domain transformation operation to confirm and calibrate the frequency-domain feature weights.

[0028] The frequency domain features, after being confirmed and calibrated, are returned to the time domain through an inverse frequency domain transformation operation, and after at least one residual connection structure fusion process, the final sub-band feature map is obtained.

[0029] Optionally, the construction of the deep residual convolutional network corresponding to each sub-frequency band signal specifically involves:

[0030] For each sub-band signal, a dual-branch parallel convolution path composed of convolution kernels of different sizes is established. The first convolution path uses a smaller convolution kernel and performs three consecutive convolution operations to obtain a high-frequency detail feature map. The second convolution path uses a larger convolution kernel and performs three consecutive convolution operations to obtain a low-frequency contour feature map.

[0031] The output feature maps of the first convolutional path and the second convolutional path are subjected to independent batch normalization and linear rectified function activation operations to obtain the first activation feature map and the second activation feature map.

[0032] The first and second activation feature maps are multiplied and fused element-wise across paths to obtain an interactive fused feature map containing cross-frequency band interaction information.

[0033] Frequency domain attention enhancement is performed on the interactive fusion feature map, wherein the frequency domain attention is calculated by multiplying the frequency weight value calculated by global average pooling element by the interactive fusion feature map to highlight the key frequency features of the signal.

[0034] The interactive fusion feature map enhanced by frequency domain attention is added element by element to the first activated feature map, and then the first residual fusion is completed by the activation operation of the linear rectified function.

[0035] The feature map after the first residual fusion is added element by element to the second activated feature map, and then the second residual fusion is completed by the activation operation of the linear rectified function.

[0036] A global adaptive feature scale calibration operation is performed on the feature map after the second residual fusion to obtain the final sub-band feature map with frequency adaptive capability.

[0037] Optionally, the step of initializing the improved marine predator algorithm population using chaotic mapping, updating the population position through Lévy flight adaptive stride and back-learning strategy, and optimizing convolutional kernel size, network skip connection weight coefficients, and dynamic gating threshold are specifically as follows:

[0038] Using three sets of Logistic chaotic mappings with different initial values, multiple initial population subsets of the improved marine predator algorithm are generated, and these multiple initial population subsets are combined to form a composite initial population with a more even distribution and diversity.

[0039] In each iteration of the improved marine predator algorithm, the fitness function value of each individual in the current population is calculated, and the step size scaling factor of Levi's flight is dynamically adjusted according to the fitness function value. Individuals with poor fitness function values ​​are updated with larger step sizes, while individuals with good fitness function values ​​are updated with smaller step sizes.

[0040] After completing the Levy flight step size update position, the dominant individuals in the first half are determined according to the fitness function value, and the reverse learning strategy is only executed on the remaining non-dominant individuals in the second half.

[0041] Calculate the change in the position of individuals in the population in three consecutive iterations, and determine whether the early convergence threshold condition of the continuous decreasing trend of position change is met. If it is met, terminate the iteration of the improved marine predator algorithm; otherwise, continue to the next iteration.

[0042] The optimal individual position in the improved marine predator algorithm population after the termination iteration is assigned to the kernel size, skip connection weight coefficient, and dynamic gating threshold in the deep residual convolutional network, respectively, and an integer approximation operation is performed on the kernel size parameter.

[0043] Optionally, the improved marine predator algorithm specifically includes:

[0044] Multiple distinct initial conditions are set, and each initial condition is iterated independently using a Logistic chaotic mapping to generate multiple initial population subsets with significant differences in positional distribution.

[0045] The initial population subsets are merged to form a composite initial population with more diverse location distribution and a wider search space coverage;

[0046] The fitness function value of each individual in the composite initial population is calculated, and after sorting according to the fitness function value, the Lévy flight step size of the individual is determined by a piecewise nonlinear step size adjustment strategy. Individuals with higher fitness function values ​​are assigned smaller step sizes, individuals with middle fitness function values ​​are assigned medium step sizes, and individuals with lower fitness function values ​​are assigned larger step sizes.

[0047] The positions of all individuals in the population are updated according to the dynamically adjusted Levi flight stride. After the position update is completed, only the non-dominant individuals at the bottom of the ranking are subjected to reverse learning to obtain new reverse positions that are significantly different from the positions of the non-dominant individuals.

[0048] The fitness function values ​​of the newly generated reverse positions and the corresponding original non-dominant individual positions are calculated separately, and the fitness is compared one by one. Only the positions with higher fitness function values ​​are retained for the next generation of population iteration.

[0049] The algorithm continuously records the position changes of individuals in the population during the most recent iterations. When the position changes of most individuals show an exponential decreasing trend for several consecutive iterations and the rate of decrease exceeds a preset exponential threshold, the algorithm iteration is immediately terminated.

[0050] Optionally, the step of fusing the feature maps of each sub-band based on the optimized network hyperparameter combination and obtaining the preliminary denoised time-domain signal through inverse wavelet packet transform is as follows:

[0051] Based on the optimized convolutional kernel size, a feature scale unification operation is performed on each sub-band feature map to adjust all sub-band feature maps to the same time-frequency scale.

[0052] Based on the network skip connection weight coefficients, the feature maps of each sub-band after scaling are weighted and fused element by element to obtain a unified fused feature map.

[0053] The amplitude features are calculated point by point in the fused feature map, and the amplitude features are subjected to global threshold gating. The feature values ​​corresponding to the positions where the amplitude features are greater than the dynamic gating threshold are retained, and the feature values ​​at other positions are set to zero, thus forming a sparse fused feature map.

[0054] Feature smoothing filtering is applied to the sparse fused feature map to eliminate discontinuities in the sparse fused feature map;

[0055] The smoothed sparse fused feature map is subjected to inverse wavelet packet transform based on the scale structure of the initial wavelet packet decomposition, and then reconstructed to the complete time domain signal.

[0056] An amplitude normalization operation is performed on the time-domain signal reconstructed by inverse wavelet packet transform to obtain a preliminarily denoised time-domain seismic signal.

[0057] Optionally, the step of reconstructing the initially denoised time-domain signal into the frequency domain through short-time Fourier transform, and then using an improved marine predator algorithm to optimize the dynamic gating threshold in real time to enhance the effective seismic signal, specifically involves:

[0058] The pre-denoised time-domain seismic signal is divided into multiple continuous signal segments according to the set window length;

[0059] Perform a short-time Fourier transform operation on each signal segment to transform each signal segment to the frequency domain and obtain the time spectrum corresponding to each segment;

[0060] Based on the real-time optimization results of the improved marine predator algorithm, the real-time dynamic gating threshold corresponding to each time spectrum is determined;

[0061] The amplitude of each frequency position in each time spectrum is compared point by point with the real-time dynamic gating threshold. Frequency domain features of positions with amplitudes greater than the dynamic gating threshold are retained, while frequency domain features of positions with amplitudes less than or equal to the dynamic gating threshold are suppressed.

[0062] For each suppressed time spectrum, a frequency domain reconstruction filtering operation is performed, and the time spectrum of each segment is converted back to the time domain by inverse short-time Fourier transform;

[0063] All signal segments converted back to the time domain are spliced ​​and fused sequentially according to their original time order to obtain a complete time-domain seismic signal with effective seismic signal enhancement after dynamic gating threshold optimization.

[0064] Optionally, the parameters of the depth residual convolutional network, based on real-time acquired seafloor seismic signal segments, are periodically adjusted using an improved marine predator algorithm to dynamically respond to changes in the marine environment. Specifically:

[0065] The system acquires multiple submarine seismic signal segments within the current time period in real time and performs unified data preprocessing on the acquired signal segments.

[0066] The current deep residual convolutional network is used to perform forward processing on the preprocessed real-time acquired signal segments to obtain the network output results for each signal segment.

[0067] Based on the error between the network output of each signal segment and the corresponding actual acquired signal segment, the overall fitness function value of the current depth residual convolutional network is calculated.

[0068] Using the current network parameters as the initial position, the ocean predator algorithm is improved to periodically perform local adaptive search updates on the kernel size, skip connection weight coefficients, and dynamic gating threshold of the deep residual convolutional network.

[0069] After each periodic fine-tuning operation, the latest optimized parameters obtained from the improved ocean predator algorithm are updated into the deep residual convolutional network;

[0070] The updated deep residual convolutional network is used to process the submarine seismic signal segments acquired in real time in the next cycle.

[0071] Optionally, the step of evaluating the enhanced seismic signal based on the signal-to-noise ratio gain, the success rate of weak seismic event detection, and the positioning error index, and then adjusting the parameters of the depth residual convolutional network accordingly, specifically involves:

[0072] The power ratio of the enhanced seismic signal to the noise is calculated for both the original and enhanced seismic signals, and the current signal-to-noise ratio gain is obtained by calculating the ratio.

[0073] Perform weak seismic event detection on the enhanced seismic signal, count the number of successfully detected weak seismic events, and calculate the current weak seismic event detection success rate;

[0074] Based on the waveform characteristics of the detected events in the enhanced seismic signal, the specific location coordinates of the seismic events are determined, and the current location error index is calculated.

[0075] The current signal-to-noise ratio gain, the success rate of detecting weak seismic events, and the positioning error index are compared with their respective preset reference thresholds to obtain the error between each index and the corresponding reference threshold.

[0076] The overall evaluation function value is calculated by weighting the errors between each indicator and the reference threshold, and the overall evaluation function value is used as the fitness function value for network parameter feedback adjustment.

[0077] Using the current network parameters as the initial population position for improving the marine predator algorithm, the kernel size, skip connection weight coefficients, and dynamic gating threshold of the deep residual convolutional network are updated iteratively based on the overall evaluation function value.

[0078] The deep residual convolutional network was reconstructed using the updated network parameters and then used to enhance the seismic signal for the next cycle.

[0079] The beneficial effects of this invention are:

[0080] (1) This invention significantly improves the comprehensiveness and precision of feature extraction of seabed seismic signals in complex marine environments by combining wavelet packet decomposition with deep residual networks, effectively enhancing the signal-to-noise ratio and weak signal detection capability, and breaking through the limitations of traditional single-scale feature extraction.

[0081] (2) This invention significantly optimizes the global and local convergence accuracy of network hyperparameter search by adopting multi-initial-condition chaotic mapping, piecewise nonlinear Lévy flight step size adjustment and reverse learning strategy in the improved marine predator algorithm. It effectively solves the problem that existing swarm intelligence algorithms cannot balance population diversity, convergence speed and optimization accuracy. It breaks through the bottleneck of insufficient parameter optimization ability of traditional algorithms and significantly enhances the robustness and generalization performance of deep residual convolutional networks.

[0082] (3) This invention significantly improves the enhancement effect of effective seismic signals by combining short-time Fourier transform with a real-time optimized dynamic gating threshold reconstruction strategy, effectively reduces the risk of misjudgment and missed judgment caused by multi-source noise interference, and shows better adaptability in complex noise environments.

[0083] (4) This invention effectively solves the problem that existing technologies cannot respond to changes in the marine environment in real time by periodically and in real-time adaptive fine-tuning of network parameters and feedback optimization system. It significantly improves the environmental adaptability and long-term stability of deep residual convolutional networks, and significantly improves the detection success rate and positioning accuracy of weak earthquake events. It also shows better application effects in the scenario of real-time monitoring of seabed earthquakes. Attached Figure Description

[0084] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0085] Figure 1 This is a flowchart illustrating the overall technical process of a deep residual network-based method for denoising and enhancing submarine seismic signals proposed in this invention.

[0086] Figure 2 This is a schematic diagram of the deep residual convolutional network structure for a deep residual network-based method for denoising and enhancing submarine seismic signals proposed in this invention.

[0087] Figure 3 This is a flowchart of the improved marine predator algorithm optimization for a deep residual network-based method for denoising and enhancing submarine seismic signals proposed in this invention. Detailed Implementation

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0089] refer to Figures 1-3 A method for denoising and enhancing submarine seismic signals based on deep residual networks, characterized by comprising:

[0090] The original submarine seismic waveform signal is decomposed into multiple sub-band signals of different frequencies using wavelet packets;

[0091] For each sub-band signal, a deep residual convolutional network composed of convolutional kernels of different sizes is used to extract features in parallel, and a frequency domain attention mechanism is used to improve the feature weights of the effective seismic signals to obtain feature maps of each sub-band.

[0092] We initialized the population of the improved marine predator algorithm using chaotic mapping, updated the population position through Levy flight adaptive step size and back learning strategy, and optimized the convolutional kernel size, network skip connection weight coefficient and dynamic gating threshold.

[0093] Based on the optimized combination of network hyperparameters, the feature maps of each sub-band are fused and the preliminary denoised time-domain signal is obtained through inverse wavelet packet transform.

[0094] The initially denoised time-domain signal is reconstructed in the frequency domain through short-time Fourier transform, and the dynamic gating threshold is optimized in real time using the improved marine predator algorithm to enhance the effective seismic signal.

[0095] Based on real-time acquired seabed seismic signal fragments, the parameters of the depth residual convolutional network are periodically adjusted using an improved marine predator algorithm to dynamically respond to changes in the marine environment.

[0096] The enhanced seismic signal is evaluated based on the signal-to-noise ratio gain, the success rate of weak seismic event detection, and the positioning error index, and the parameters of the depth residual convolutional network are adjusted accordingly.

[0097] By combining wavelet packet decomposition with deep residual convolutional networks, more refined multi-scale feature extraction of seafloor seismic signals can be achieved, significantly improving the signal-to-noise ratio and the detection capability of weak seismic events. The improved marine predator algorithm enhances the efficiency and accuracy of network hyperparameter search by initializing the population through chaotic mapping, adopting an adaptive step size for Lévy flight, and employing a back-learning strategy, effectively addressing the shortcomings of existing algorithms that are prone to getting trapped in local optima. Real-time optimized dynamic gating thresholds enable adaptive enhancement of effective signals. Periodic network parameter fine-tuning and feedback optimization strategies enhance the network's dynamic adaptability to changes in the marine environment.

[0098] In this embodiment, the process of decomposing the original submarine seismic waveform signal into multiple sub-band signals of different frequencies using wavelet packets specifically involves:

[0099] The original submarine seismic waveform signal is divided into several signal segments of equal length, and the Shannon entropy value of each signal segment is calculated separately.

[0100] Based on the calculated Shannon entropy value of each signal segment, the signal segment with the highest entropy value is selected as the reference segment. The Daubechies wavelet basis function is used to perform the first binary recursive wavelet packet decomposition to obtain multiple initial frequency band sub-signals.

[0101] Calculate the Shannon entropy value of the initial frequency band sub-signals respectively, and determine whether each frequency band sub-signal should continue to be decomposed into wavelet packet at the next scale. Stop the decomposition when the Shannon entropy value of all frequency band sub-signals is lower than or equal to the preset entropy threshold.

[0102] For the remaining signal segments with lower entropy values, wavelet packet decomposition is performed synchronously using the same scale and frequency band division method to obtain multiple sub-frequency band signals of different frequencies;

[0103] All sub-band signals are labeled with their corresponding frequency ranges and arranged in ascending order of frequency.

[0104] In the permuted and combined sub-band signals, the Shannon entropy value of each sub-band signal is calculated again, and the sub-band signals in a specific frequency range whose entropy value exceeds the set threshold are subjected to additional-scale local recursive wavelet packet decomposition to obtain the final sub-band signals with lower entropy values.

[0105] By selecting a benchmark segment based on Shannon entropy values ​​for binary recursive wavelet packet decomposition, the adaptability and rationality of the decomposition scale are ensured. A unified scale and frequency band division method are adopted for signal segments with different entropy values, which significantly improves the standardization and consistency of signal processing. Additional local recursive decomposition is performed on high-entropy frequency bands, which effectively enhances the fineness of frequency band division, reduces the entropy level of sub-frequency band signals, and thus improves the effectiveness of subsequent feature extraction and the overall accuracy of seismic signal denoising.

[0106] In this embodiment, for each sub-frequency band signal, parallel feature extraction is performed using deep residual convolutional networks composed of convolutional kernels of different sizes, and a frequency domain attention mechanism is used to improve the feature weights of the effective seismic signals to obtain feature maps for each sub-frequency band, specifically as follows:

[0107] For each sub-band signal, the sub-band signal is transformed to the frequency domain through a frequency domain transformation operation to obtain the corresponding initial frequency domain characteristics;

[0108] For the initial frequency domain features of each sub-band signal, the amplitude features are calculated respectively, and global average pooling and global max pooling operations are performed based on the amplitude features to calculate the frequency domain attention weight value corresponding to each sub-band signal;

[0109] Based on the distribution characteristics of the calculated frequency domain attention weight values, the convolution kernel size corresponding to each sub-band signal is determined, and the determined convolution kernel size is used to construct the deep residual convolutional network corresponding to each sub-band signal.

[0110] Each sub-band signal is passed through a depth residual convolutional network with the corresponding kernel size and subjected to at least two convolution operations to extract temporal features.

[0111] The temporal features extracted by the convolution operation are multiplied element-wise with the previously obtained frequency domain attention weights to obtain the temporal enhanced features after processing by the frequency domain attention mechanism.

[0112] The time-domain enhanced features processed by the frequency-domain attention mechanism are then returned to the frequency domain through an inverse frequency-domain transformation operation to confirm and calibrate the frequency-domain feature weights.

[0113] The frequency domain features, after being confirmed and calibrated, are returned to the time domain through an inverse frequency domain transformation operation, and after at least one residual connection structure fusion process, the final sub-band feature map is obtained.

[0114] By determining the kernel size based on frequency domain attention weights, a dual-branch deep residual convolutional network structure was constructed, achieving full extraction of features from different frequency bands. The combination of cross-path feature fusion and frequency domain attention mechanism enabled effective interaction and enhancement of high-frequency details and low-frequency contour features, improving the prominence of effective signal features. The application of residual fusion and adaptive feature scale calibration further enhanced the frequency adaptability of sub-band feature maps, improved the network's robustness to multi-source noise, and significantly improved the denoising and enhancement effect of seismic signals.

[0115] In this embodiment, the construction of the deep residual convolutional network corresponding to each sub-frequency band signal specifically involves:

[0116] For each sub-band signal, a dual-branch parallel convolution path composed of convolution kernels of different sizes is established. The first convolution path uses a smaller convolution kernel and performs three consecutive convolution operations to obtain a high-frequency detail feature map. The second convolution path uses a larger convolution kernel and performs three consecutive convolution operations to obtain a low-frequency contour feature map.

[0117] The output feature maps of the first convolutional path and the second convolutional path are subjected to independent batch normalization and linear rectified function activation operations to obtain the first activation feature map and the second activation feature map.

[0118] The first and second activation feature maps are multiplied and fused element-wise across paths to obtain an interactive fused feature map containing cross-frequency band interaction information.

[0119] Frequency domain attention enhancement is performed on the interactive fusion feature map, wherein the frequency domain attention is calculated by multiplying the frequency weight value calculated by global average pooling element by the interactive fusion feature map to highlight the key frequency features of the signal.

[0120] The interactive fusion feature map enhanced by frequency domain attention is added element by element to the first activated feature map, and then the first residual fusion is completed by the activation operation of the linear rectified function.

[0121] The feature map after the first residual fusion is added element by element to the second activated feature map, and then the second residual fusion is completed by the activation operation of the linear rectified function.

[0122] A global adaptive feature scale calibration operation is performed on the feature map after the second residual fusion to obtain the final sub-band feature map with frequency adaptive capability.

[0123] By constructing a dual-branch parallel convolution path and extracting high-frequency details and low-frequency contour features respectively, the diversity and complementarity of sub-band feature extraction are effectively improved. The application of cross-path element-wise fusion operation and frequency domain attention enhancement significantly enhances the degree of interaction and fusion of cross-frequency band information within the feature map and increases the weight of key frequency features. The implementation of two residual fusion operations and global adaptive feature scale calibration measures further enhances the stability and adaptability of the sub-band feature map and significantly improves the denoising accuracy and robustness of submarine seismic signals.

[0124] In this embodiment, the process of initializing the improved marine predator algorithm population using chaotic mapping, updating the population position through Lévy flight adaptive stride and back-learning strategies, and optimizing convolutional kernel size, network skip connection weight coefficients, and dynamic gating thresholds are as follows:

[0125] Using three sets of Logistic chaotic mappings with different initial values, multiple initial population subsets of the improved marine predator algorithm are generated, and these multiple initial population subsets are combined to form a composite initial population with a more even distribution and diversity.

[0126] In each iteration of the improved marine predator algorithm, the fitness function value of each individual in the current population is calculated, and the step size scaling factor of Levi's flight is dynamically adjusted according to the fitness function value. Individuals with poor fitness function values ​​are updated with larger step sizes, while individuals with good fitness function values ​​are updated with smaller step sizes.

[0127] After completing the Levy flight step size update position, the dominant individuals in the first half are determined according to the fitness function value, and the reverse learning strategy is only executed on the remaining non-dominant individuals in the second half.

[0128] Calculate the change in the position of individuals in the population in three consecutive iterations, and determine whether the early convergence threshold condition of the continuous decreasing trend of position change is met. If it is met, terminate the iteration of the improved marine predator algorithm; otherwise, continue to the next iteration.

[0129] The optimal individual position in the improved marine predator algorithm population after the termination iteration is assigned to the kernel size, skip connection weight coefficient, and dynamic gating threshold in the deep residual convolutional network, respectively, and an integer approximation operation is performed on the kernel size parameter.

[0130] Multiple population subsets are generated by using three sets of Logistic chaotic mappings with different initial values ​​and then merged into a composite initial population, which effectively improves the comprehensiveness of the algorithm's search space. The combination of dynamically adjusting the Levy flight step size based on the fitness function value and the selective back-learning strategy achieves a balance between local fine-grained search of dominant individuals and rapid convergence of non-dominant individuals, significantly improving the overall optimization efficiency of the population. The early convergence judgment mechanism based on position change monitoring and the integer approximation operation of the convolution kernel size further improve the efficiency and practicality of the network parameter optimization process.

[0131] In this embodiment, the improved marine predator algorithm is specifically as follows:

[0132] Multiple distinct initial conditions are set, and each initial condition is iterated independently using a Logistic chaotic mapping to generate multiple initial population subsets with significant differences in positional distribution.

[0133] The initial population subsets are merged to form a composite initial population with more diverse location distribution and a wider search space coverage;

[0134] The fitness function value of each individual in the composite initial population is calculated, and after sorting according to the fitness function value, the Lévy flight step size of the individual is determined by a piecewise nonlinear step size adjustment strategy. Individuals with higher fitness function values ​​are assigned smaller step sizes, individuals with middle fitness function values ​​are assigned medium step sizes, and individuals with lower fitness function values ​​are assigned larger step sizes.

[0135] The positions of all individuals in the population are updated according to the dynamically adjusted Levi flight stride. After the position update is completed, only the non-dominant individuals at the bottom of the ranking are subjected to reverse learning to obtain new reverse positions that are significantly different from the positions of the non-dominant individuals.

[0136] The fitness function values ​​of the newly generated reverse positions and the corresponding original non-dominant individual positions are calculated separately, and the fitness is compared one by one. Only the positions with higher fitness function values ​​are retained for the next generation of population iteration.

[0137] The algorithm continuously records the position changes of individuals in the population during the most recent iterations. When the position changes of most individuals show an exponential decreasing trend for several consecutive iterations and the rate of decrease exceeds a preset exponential threshold, the algorithm iteration is immediately terminated.

[0138] The population is initialized using three sets of Logistic chaotic mappings with different initial values, forming a diverse composite initial population, which significantly improves the coverage of the search space. The combination of a piecewise nonlinear step size adjustment strategy and a selective back-learning strategy ensures accurate optimization of dominant individuals and rapid convergence of non-dominant individuals during the population update process, significantly enhancing the global search efficiency of the algorithm. The setting of position change monitoring and early convergence mechanism enables the algorithm to effectively avoid unnecessary iterations, significantly improving the convergence accuracy and speed of network hyperparameter optimization.

[0139] In this embodiment, the step of fusing the feature maps of each sub-band based on the optimized network hyperparameter combination and obtaining the initially denoised time-domain signal through inverse wavelet packet transform is as follows:

[0140] Based on the optimized convolutional kernel size, a feature scale unification operation is performed on each sub-band feature map to adjust all sub-band feature maps to the same time-frequency scale.

[0141] Based on the network skip connection weight coefficients, the feature maps of each sub-band after scaling are weighted and fused element by element to obtain a unified fused feature map.

[0142] The amplitude features are calculated point by point in the fused feature map, and the amplitude features are subjected to global threshold gating. The feature values ​​corresponding to the positions where the amplitude features are greater than the dynamic gating threshold are retained, and the feature values ​​at other positions are set to zero, thus forming a sparse fused feature map.

[0143] Feature smoothing filtering is applied to the sparse fused feature map to eliminate discontinuities in the sparse fused feature map;

[0144] The smoothed sparse fused feature map is subjected to inverse wavelet packet transform based on the scale structure of the initial wavelet packet decomposition, and then reconstructed to the complete time domain signal.

[0145] An amplitude normalization operation is performed on the time-domain signal reconstructed by inverse wavelet packet transform to obtain a preliminarily denoised time-domain seismic signal.

[0146] By iterating independently using multiple Logistic chaotic mappings with different initial conditions and then merging them into a composite population, the initial search space of the algorithm is effectively broadened, and the diversity of the initial population is improved. The combination of piecewise nonlinear step size adjustment of fitness function values ​​and selective back-learning mechanism enables local precise search of dominant individuals and rapid updating of non-dominant individuals during the algorithm iteration process. Continuous monitoring of changes in the position of individuals in the population and setting an early convergence mechanism effectively improve the iteration efficiency of the algorithm, reduce unnecessary iteration processes, and significantly enhance the convergence speed and stability of the network parameter optimization process.

[0147] In this embodiment, the step of converting the initially denoised time-domain signal to the frequency domain through short-time Fourier transform and then using an improved marine predator algorithm to optimize the dynamic gating threshold in real time to enhance the effective seismic signal specifically involves:

[0148] The pre-denoised time-domain seismic signal is divided into multiple continuous signal segments according to the set window length;

[0149] Perform a short-time Fourier transform operation on each signal segment to transform each signal segment to the frequency domain and obtain the time spectrum corresponding to each segment;

[0150] Based on the real-time optimization results of the improved marine predator algorithm, the real-time dynamic gating threshold corresponding to each time spectrum is determined;

[0151] The amplitude of each frequency position in each time spectrum is compared point by point with the real-time dynamic gating threshold. Frequency domain features of positions with amplitudes greater than the dynamic gating threshold are retained, while frequency domain features of positions with amplitudes less than or equal to the dynamic gating threshold are suppressed.

[0152] For each suppressed time spectrum, a frequency domain reconstruction filtering operation is performed, and the time spectrum of each segment is converted back to the time domain by inverse short-time Fourier transform;

[0153] All signal segments converted back to the time domain are spliced ​​and fused sequentially according to their original time order to obtain a complete time-domain seismic signal with effective seismic signal enhancement after dynamic gating threshold optimization.

[0154] By using a dynamically optimized threshold in real time to adaptively and selectively enhance the frequency domain features of the time spectrum after the short-time Fourier transform, the prominence of key frequency components in the seismic signal is effectively improved, while background noise and unrelated frequency components are significantly suppressed. The combination of frequency domain reconstruction filtering and inverse short-time Fourier transform further improves the smoothness and continuity of the signal, reduces the discontinuity and error of the time domain reconstructed signal, and effectively improves the stability of the dynamically gated optimized seismic signal under complex noise conditions.

[0155] In this embodiment, the improved marine predator algorithm is used periodically to adaptively fine-tune the parameters of the depth residual convolutional network based on real-time acquired seafloor seismic signal segments, dynamically responding to changes in the marine environment. Specifically:

[0156] The system acquires multiple submarine seismic signal segments within the current time period in real time and performs unified data preprocessing on the acquired signal segments.

[0157] The current deep residual convolutional network is used to perform forward processing on the preprocessed real-time acquired signal segments to obtain the network output results for each signal segment.

[0158] Based on the error between the network output of each signal segment and the corresponding actual acquired signal segment, the overall fitness function value of the current depth residual convolutional network is calculated.

[0159] Using the current network parameters as the initial position, the ocean predator algorithm is improved to periodically perform local adaptive search updates on the kernel size, skip connection weight coefficients, and dynamic gating threshold of the deep residual convolutional network.

[0160] After each periodic fine-tuning operation, the latest optimized parameters obtained from the improved ocean predator algorithm are updated into the deep residual convolutional network;

[0161] The updated deep residual convolutional network is used to process the submarine seismic signal segments acquired in real time in the next cycle.

[0162] By periodically calculating the fitness function value and locally adaptively optimizing the network parameters through real-time acquisition of seabed seismic signal segments, dynamic fine-tuning of the deep residual convolutional network parameters was achieved, effectively improving the network's adaptive response capability to changes in the actual marine environment. Periodically updating the latest optimized parameters to the network significantly enhanced the network's generalization performance and long-term stability, improved the denoising accuracy and reliability of actual acquired seismic signals, and effectively reduced the impact of changes in the marine environment on signal processing results.

[0163] In this embodiment, the step of evaluating the enhanced seismic signal based on the signal-to-noise ratio gain, the success rate of weak seismic event detection, and the positioning error index, and then adjusting the parameters of the depth residual convolutional network accordingly, specifically involves:

[0164] The power ratio of the enhanced seismic signal to the noise is calculated for both the original and enhanced seismic signals, and the current signal-to-noise ratio gain is obtained by calculating the ratio.

[0165] Perform weak seismic event detection on the enhanced seismic signal, count the number of successfully detected weak seismic events, and calculate the current weak seismic event detection success rate;

[0166] Based on the waveform characteristics of the detected events in the enhanced seismic signal, the specific location coordinates of the seismic events are determined, and the current location error index is calculated.

[0167] The current signal-to-noise ratio gain, the success rate of detecting weak seismic events, and the positioning error index are compared with their respective preset reference thresholds to obtain the error between each index and the corresponding reference threshold.

[0168] The overall evaluation function value is calculated by weighting the errors between each indicator and the reference threshold, and the overall evaluation function value is used as the fitness function value for network parameter feedback adjustment.

[0169] Using the current network parameters as the initial population position for improving the marine predator algorithm, the kernel size, skip connection weight coefficients, and dynamic gating threshold of the deep residual convolutional network are updated iteratively based on the overall evaluation function value.

[0170] The deep residual convolutional network was reconstructed using the updated network parameters and then used to enhance the seismic signal for the next cycle.

[0171] By calculating the signal-to-noise ratio gain, the success rate of weak earthquake event detection, and the location error index of the enhanced seismic signal, an objective quantitative evaluation of the seismic signal enhancement effect was clearly achieved. The weighted fusion calculation of the index error, used as the fitness function value for network parameter feedback adjustment, can accurately guide the optimization and adjustment of the hyperparameters of the deep residual convolutional network, effectively overcoming the blindness of the parameter adjustment process. By dynamically constructing the network using the updated optimized parameters, the adaptability and accuracy of seismic signal processing in the next cycle are significantly improved.

[0172] Example 1:

[0173] To verify the feasibility of this invention in practice, the method was applied to a long-term operating seabed seismic monitoring system in a certain sea area. Specific verification experiments were conducted to address the problems of low signal-to-noise ratio, high missed detection rate of microseismic events, and large location errors in the seismic signals acquired by the current monitoring system. In this practical scenario, traditional seabed seismic signal denoising and enhancement techniques mainly employ wavelet threshold denoising with fixed parameters and empirical mode decomposition methods. However, these methods lack adaptability and produce poor denoising results when faced with multi-source interference noise and abrupt changes in the marine environment. This leads to difficulties in accurately detecting and locating weak seismic events, with a missed detection rate exceeding 35% and a location error exceeding 500 meters.

[0174] In practice, the continuously acquired raw submarine seismic waveform signals are first divided into several signal segments of equal length. The Shannon entropy value of each signal segment is calculated, and the signal segment with the highest entropy value is used as the reference segment. Binary recursive wavelet packet decomposition is then performed using the Daubechies wavelet basis function. Further decomposition is then determined based on the entropy value of each sub-band signal. The frequency ranges of each sub-band signal are rearranged and combined, and localized refinement is performed to ultimately obtain multiple sub-band signals with lower entropy values.

[0175] Frequency domain transformation and amplitude features are performed for each sub-band signal. The kernel size of the deep residual convolutional network is determined by the frequency domain attention weights. A dual-branch convolutional path is constructed for parallel convolution operations to extract high-frequency detail feature maps and low-frequency contour feature maps. Key frequency features are then highlighted through cross-path element-wise multiplication fusion and a frequency domain attention mechanism. After two residual fusions and feature scale calibrations, the final sub-band feature maps are obtained.

[0176] During the network hyperparameter optimization phase, three sets of Logistic chaotic mappings with different initial values ​​are used to generate multiple initial population subsets and merge them into a composite initial population. The Levy fly-through adaptive stride and back-learning strategy are used to update the positions, and the convolutional kernel size, network skip connection weights, and dynamic gating threshold are iteratively optimized. After 15 iterations, the algorithm terminates after reaching the convergence threshold condition of continuously decreasing position changes, and the optimal hyperparameters are assigned to the convolutional network.

[0177] The optimized network parameters are combined and fused with the feature maps of each sub-band, and the inverse wavelet packet transform is used to obtain the initially denoised time-domain signal. The time-domain signal is then transformed to the frequency domain using a short-time Fourier transform, and the dynamic gating threshold is optimized in real time by improving the ocean predator algorithm to perform frequency domain feature selection and noise suppression. Finally, an inverse transform is performed to obtain the enhanced complete time-domain signal.

[0178] Based on real-time acquired seismic signal segments, the improved marine predator algorithm is used periodically to locally and adaptively fine-tune the network parameters. The network parameters are adjusted according to the feedback of signal-to-noise ratio gain, success rate of weak seismic event detection, and positioning error index, so as to continuously improve the signal enhancement processing effect.

[0179] During the experiment, the continuously acquired seismic signals were processed, and the specific data results are shown in Table 1:

[0180] Table 1 Comparison of Signal Enhancement Processing Effects Before and After Enhancement

[0181]

[0182] As shown in the table above, after processing the five different signal samples, the signal-to-noise ratio of all samples was significantly improved by more than 4dB, the number of detected weak seismic events increased dramatically, and the event false negative rate decreased from approximately 35% before processing to less than 10% after processing. Specifically, for example, sample 2 detected only 14 events without processing, but after processing with this method, 21 events were detected, increasing the true detection rate from 63.6% to 95.5%, significantly reducing false negatives. Regarding the accuracy of seismic event location, the location errors of each sample were between 480 meters and 610 meters before processing, while after processing with the method of this invention, the error range was significantly reduced to 160 meters to 215 meters. Among them, the location error of sample 5 decreased from 490 meters to 160 meters, a reduction of more than 67%.

[0183] The above data clearly demonstrates that the technical solution of this invention achieves a significant performance improvement in denoising and enhancement of submarine seismic signals. The signal-to-noise ratio, weak event detection rate, and positioning error are all significantly better than traditional methods, effectively solving the problem that existing technologies cannot simultaneously achieve both signal enhancement accuracy and adaptability. In practical implementation, the optimized deep residual convolutional network combined with the improved marine predator algorithm exhibits high efficiency, accuracy, and robustness, enabling this technical solution to effectively cope with multi-source noise interference and rapid changes in the complex marine environment, demonstrating good practical application value and promising prospects for widespread adoption.

[0184] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for denoising and enhancing submarine seismic signals based on deep residual networks, characterized in that, include: The original submarine seismic waveform signal is decomposed into multiple sub-band signals of different frequencies using wavelet packets; For each sub-band signal, a deep residual convolutional network composed of convolutional kernels of different sizes is used to extract features in parallel, and a frequency domain attention mechanism is used to improve the feature weights of the effective seismic signals to obtain feature maps of each sub-band. We initialized the population of the improved marine predator algorithm using chaotic mapping, updated the population position through Levy flight adaptive step size and back learning strategy, and optimized the convolutional kernel size, network skip connection weight coefficient and dynamic gating threshold. Based on the optimized combination of network hyperparameters, the feature maps of each sub-band are fused and the preliminary denoised time-domain signal is obtained through inverse wavelet packet transform. The initially denoised time-domain signal is reconstructed in the frequency domain through short-time Fourier transform, and the dynamic gating threshold is optimized in real time using the improved marine predator algorithm to enhance the effective seismic signal. Based on real-time acquired seabed seismic signal fragments, the parameters of the depth residual convolutional network are periodically adjusted using an improved marine predator algorithm to dynamically respond to changes in the marine environment. The enhanced seismic signal is evaluated based on the signal-to-noise ratio gain, the success rate of weak seismic event detection, and the positioning error index, and the parameters of the depth residual convolutional network are adjusted accordingly.

2. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, The process of decomposing the original seafloor seismic waveform signal into multiple sub-band signals of different frequencies using wavelet packets is specifically as follows: The original submarine seismic waveform signal is divided into several signal segments of equal length, and the Shannon entropy value of each signal segment is calculated separately. Based on the calculated Shannon entropy value of each signal segment, the signal segment with the highest entropy value is selected as the reference segment. The Daubechies wavelet basis function is used to perform the first binary recursive wavelet packet decomposition to obtain multiple initial frequency band sub-signals. Calculate the Shannon entropy value of the initial frequency band sub-signals respectively, and determine whether each frequency band sub-signal should continue to be decomposed into wavelet packet at the next scale. Stop the decomposition when the Shannon entropy value of all frequency band sub-signals is lower than or equal to the preset entropy threshold. For the remaining signal segments with lower entropy values, wavelet packet decomposition is performed synchronously using the same scale and frequency band division method to obtain multiple sub-frequency band signals of different frequencies; All sub-band signals are labeled with their corresponding frequency ranges and arranged in ascending order of frequency. In the permuted and combined sub-band signals, the Shannon entropy value of each sub-band signal is calculated again, and the sub-band signals in a specific frequency range whose entropy value exceeds the set threshold are subjected to additional-scale local recursive wavelet packet decomposition to obtain the final sub-band signals with lower entropy values.

3. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, For each sub-frequency band signal, parallel feature extraction is performed using deep residual convolutional networks composed of convolutional kernels of different sizes. A frequency domain attention mechanism is then used to increase the feature weights of the effective seismic signals, resulting in feature maps for each sub-frequency band. Specifically: For each sub-band signal, the sub-band signal is transformed to the frequency domain through a frequency domain transformation operation to obtain the corresponding initial frequency domain characteristics; For each sub-band signal, the amplitude features are calculated based on the initial frequency domain features. Global average pooling and global max pooling operations are then performed based on the amplitude features to calculate the frequency domain attention weight value corresponding to each sub-band signal. Based on the distribution characteristics of the calculated frequency domain attention weight values, the convolution kernel size corresponding to each sub-band signal is determined, and the determined convolution kernel size is used to construct the deep residual convolutional network corresponding to each sub-band signal. Each sub-band signal is passed through a depth residual convolutional network with the corresponding kernel size and subjected to at least two convolution operations to extract temporal features. The temporal features extracted by the convolution operation are multiplied element-wise with the previously obtained frequency domain attention weights to obtain the temporal enhanced features after processing by the frequency domain attention mechanism. The time-domain enhanced features processed by the frequency-domain attention mechanism are then returned to the frequency domain through an inverse frequency-domain transformation operation to confirm and calibrate the frequency-domain feature weights. The frequency domain features, after being confirmed and calibrated, are returned to the time domain through an inverse frequency domain transformation operation, and after at least one residual connection structure fusion process, the final sub-band feature map is obtained.

4. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 3, characterized in that, The construction of the deep residual convolutional network corresponding to each sub-frequency band signal is specifically as follows: For each sub-band signal, a dual-branch parallel convolution path composed of convolution kernels of different sizes is established. The first convolution path uses a smaller convolution kernel and performs three consecutive convolution operations to obtain a high-frequency detail feature map. The second convolution path uses a larger convolution kernel and performs three consecutive convolution operations to obtain a low-frequency contour feature map. The output feature maps of the first convolutional path and the second convolutional path are subjected to independent batch normalization and linear rectified function activation operations to obtain the first activation feature map and the second activation feature map. The first and second activation feature maps are multiplied and fused element-wise across paths to obtain an interactive fused feature map containing cross-frequency band interaction information. Frequency domain attention enhancement is performed on the interactive fusion feature map, wherein the frequency domain attention is multiplied element-wise by the interactive fusion feature map through the frequency weight value calculated by global average pooling to highlight the key frequency features of the signal. The interactive fusion feature map enhanced by frequency domain attention is added element by element to the first activated feature map, and then the first residual fusion is completed by the activation operation of the linear rectified function. The feature map after the first residual fusion is added element by element to the second activated feature map, and then the second residual fusion is completed by the activation operation of the linear rectified function. A global adaptive feature scale calibration operation is performed on the feature map after the second residual fusion to obtain the final sub-band feature map with frequency adaptive capability.

5. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, The improved marine predator algorithm population is initialized using chaotic mapping, and the population position is updated through Lévy flight adaptive stride and back-learning strategy. Specifically, the convolutional kernel size, network skip connection weight coefficients, and dynamic gating threshold are optimized as follows: Using three sets of Logistic chaotic mappings with different initial values, multiple initial population subsets of the improved marine predator algorithm are generated, and these multiple initial population subsets are combined to form a composite initial population with a more even distribution and diversity. In each iteration of the improved marine predator algorithm, the fitness function value of each individual in the current population is calculated, and the step size scaling factor of Levi's flight is dynamically adjusted according to the fitness function value. Individuals with poor fitness function values ​​are updated with larger step sizes, while individuals with good fitness function values ​​are updated with smaller step sizes. After completing the Levy flight step size update position, the dominant individuals in the first half are determined according to the fitness function value, and the reverse learning strategy is only executed on the remaining non-dominant individuals in the second half. Calculate the change in the position of individuals in the population in three consecutive iterations, and determine whether the early convergence threshold condition of the continuous decreasing trend of position change is met. If it is met, terminate the iteration of the improved marine predator algorithm; otherwise, continue to the next iteration. The optimal individual position in the improved marine predator algorithm population after the termination iteration is assigned to the kernel size, skip connection weight coefficient, and dynamic gating threshold in the deep residual convolutional network, respectively, and an integer approximation operation is performed on the kernel size parameter.

6. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 5, characterized in that, The improved marine predator algorithm is as follows: Multiple distinct initial conditions are set, and each initial condition is iterated independently using a Logistic chaotic mapping to generate multiple initial population subsets with significant differences in positional distribution. The initial population subsets are merged to form a composite initial population with more diverse location distribution and a wider search space coverage; The fitness function value of each individual in the composite initial population is calculated, and after sorting according to the fitness function value, the Lévy flight step size of the individual is determined by a piecewise nonlinear step size adjustment strategy. Individuals with higher fitness function values ​​are assigned smaller step sizes, individuals with middle fitness function values ​​are assigned medium step sizes, and individuals with lower fitness function values ​​are assigned larger step sizes. The positions of all individuals in the population are updated according to the dynamically adjusted Levi flight stride. After the position update is completed, only the non-dominant individuals at the bottom of the ranking are subjected to reverse learning to obtain new reverse positions that are significantly different from the positions of the non-dominant individuals. The fitness function values ​​of the newly generated reverse positions and the corresponding original non-dominant individual positions are calculated separately, and the fitness is compared one by one. Only the positions with higher fitness function values ​​are retained for the next generation of population iteration. The algorithm continuously records the position changes of individuals in the population during the most recent iterations. When the position changes of most individuals show an exponential decreasing trend for several consecutive iterations and the rate of decrease exceeds a preset exponential threshold, the algorithm iteration is immediately terminated.

7. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, The process involves fusing the feature maps of each sub-band based on the optimized network hyperparameter combination, and obtaining the preliminary denoised time-domain signal through inverse wavelet packet transform. Specifically: Based on the optimized convolutional kernel size, a feature scale unification operation is performed on each sub-band feature map to adjust all sub-band feature maps to the same time-frequency scale. Based on the network skip connection weight coefficients, the feature maps of each sub-band after scaling are weighted and fused element by element to obtain a unified fused feature map. The amplitude features are calculated point by point in the fused feature map, and the amplitude features are subjected to global threshold gating. The feature values ​​corresponding to the positions where the amplitude features are greater than the dynamic gating threshold are retained, and the feature values ​​at other positions are set to zero, thus forming a sparse fused feature map. Feature smoothing filtering is applied to the sparse fused feature map to eliminate discontinuities in the sparse fused feature map; The smoothed sparse fused feature map is subjected to inverse wavelet packet transform based on the scale structure of the initial wavelet packet decomposition, and then reconstructed to the complete time domain signal. An amplitude normalization operation is performed on the time-domain signal reconstructed by inverse wavelet packet transform to obtain a preliminarily denoised time-domain seismic signal.

8. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, The process involves reconstructing the frequency domain from the initially denoised time-domain signal using a short-time Fourier transform, and then optimizing the dynamic gating threshold in real time using an improved marine predator algorithm to enhance the effective seismic signal. Specifically: The pre-denoised time-domain seismic signal is divided into multiple continuous signal segments according to the set window length; Perform a short-time Fourier transform operation on each signal segment to transform each signal segment to the frequency domain and obtain the time spectrum corresponding to each segment; Based on the real-time optimization results of the improved marine predator algorithm, the real-time dynamic gating threshold corresponding to each time spectrum is determined; The amplitude of each frequency position in each time spectrum is compared point by point with the real-time dynamic gating threshold. Frequency domain features of positions with amplitudes greater than the dynamic gating threshold are retained, while frequency domain features of positions with amplitudes less than or equal to the dynamic gating threshold are suppressed. For each suppressed time spectrum, a frequency domain reconstruction filtering operation is performed, and the time spectrum of each segment is converted back to the time domain by inverse short-time Fourier transform; All signal segments converted back to the time domain are spliced ​​and fused sequentially according to their original time order to obtain a complete time-domain seismic signal with effective seismic signal enhancement after dynamic gating threshold optimization.

9. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, The aforementioned seafloor seismic signal segments acquired in real time are periodically fine-tuned using an improved marine predator algorithm to adaptively fine-tune the parameters of the depth residual convolutional network, dynamically responding to changes in the marine environment. Specifically: The system acquires multiple submarine seismic signal segments within the current time period in real time and performs unified data preprocessing on the acquired signal segments. The current deep residual convolutional network is used to perform forward processing on the preprocessed real-time acquired signal segments to obtain the network output results for each signal segment. Based on the error between the network output of each signal segment and the corresponding actual acquired signal segment, the overall fitness function value of the current depth residual convolutional network is calculated. Using the current network parameters as the initial position, the ocean predator algorithm is improved to periodically perform local adaptive search updates on the kernel size, skip connection weight coefficients, and dynamic gating threshold of the deep residual convolutional network. After each periodic fine-tuning operation, the latest optimized parameters obtained from the improved ocean predator algorithm are updated into the deep residual convolutional network; The updated deep residual convolutional network is used to process the submarine seismic signal segments acquired in real time in the next cycle.

10. The method for denoising and enhancing submarine seismic signals based on deep residual networks according to claim 1, characterized in that, The process involves evaluating the enhanced seismic signal based on signal-to-noise ratio gain, success rate of weak seismic event detection, and location error index, and then adjusting the parameters of the depth residual convolutional network accordingly. The power ratio of the enhanced seismic signal to the noise is calculated for both the original and enhanced seismic signals, and the current signal-to-noise ratio gain is obtained by calculating the ratio. Perform weak seismic event detection on the enhanced seismic signal, count the number of successfully detected weak seismic events, and calculate the current weak seismic event detection success rate; Based on the waveform characteristics of the detected events in the enhanced seismic signal, the specific location coordinates of the seismic events are determined, and the current location error index is calculated. The current signal-to-noise ratio gain, the success rate of detecting weak seismic events, and the positioning error index are compared with their respective preset reference thresholds to obtain the error between each index and the corresponding reference threshold. The overall evaluation function value is calculated by weighting the errors between each indicator and the reference threshold, and the overall evaluation function value is used as the fitness function value for network parameter feedback adjustment. Using the current network parameters as the initial population position for improving the marine predator algorithm, the kernel size, skip connection weight coefficients, and dynamic gating threshold of the deep residual convolutional network are updated iteratively based on the overall evaluation function value. The deep residual convolutional network was reconstructed using the updated network parameters and then used to enhance the seismic signal for the next cycle.

Images