Seismic exploration data analysis method and system based on large model

Abstract

The application discloses a kind of based on big model's seismic exploration data analysis method and system, it is related to seismic data analysis technical field, including following steps: S1, obtain multi-source data and carry out pre-processing;S2, carry out sliding window segmentation, combination continuous registration fragment, construct training sample set;S3, to window fragment carries out Patch coding, and cross-lane registration relationship is embedded fragment representation, obtains seismic analysis big model by back propagation training;S4, to the seismic state sequence to be analyzed carries out multilayer coding, carries out state recursion and corrects subsequent recursive state;S5, constructs conditional transition probability, response state likelihood and conditional transition probability are formed path probability in conjunction with, global decoding is carried out, obtains exploration data analysis result;S6, obtains feedback data, and updates seismic analysis big model.The application can realize cross-lane correlation analysis and continuous response identification of seismic exploration data, improve the accuracy of seismic data analysis.

Description

Technical Field

[0001] This invention relates to the field of seismic data analysis technology, and in particular to a method and system for seismic exploration data analysis based on a large model. Background Technology

[0002] Seismic exploration is a crucial technique for acquiring information about underground geological structures and strata. It typically involves generating seismic waves from a seismic source and using geophones deployed along the survey line to collect waveform data at different locations. With the development of high-density, multi-channel, and broadband acquisition technologies, the amount of data generated by seismic exploration is constantly increasing, and the spatiotemporal relationships between the data are becoming more complex, placing higher demands on subsequent analysis and processing.

[0003] Current seismic data analysis primarily relies on methods such as filtering, spectral analysis, cross-correlation, and attribute extraction to process features like amplitude, frequency, phase, and travel time, and combines these with human experience to complete reflection response tracking and stratigraphic analysis. While these methods are sensitive to local waveform changes, they are easily affected by noise, waveform distortion, and local anomalies when dealing with large-scale continuous seismic data, making it difficult to maintain stable continuous response relationships across seismic traces.

[0004] In recent years, deep learning and Transformer models have been increasingly used for earthquake data identification and analysis. By learning complex waveform features through training data, they have reduced the reliance on manual feature design to some extent. However, most existing models focus on single-channel or local multi-channel data, and do not adequately utilize long sequences and the time migration and continuous propagation relationships between adjacent seismic traces. When the same seismic response shifts in different seismic traces, the model is prone to problems such as discontinuous correlations or inconsistent local judgments.

[0005] Large-scale time series models provide a new technical foundation for long-sequence analysis of earthquake data. However, general time series models are mainly designed for general time series. Their internal representation and state propagation mechanisms are not designed for cross-channel registration and continuous propagation characteristics of earthquake data. The model output results are also prone to remain at the level of local judgment, lacking overall analysis and correction of continuous response relationships.

[0006] Therefore, how to provide a method and system for seismic exploration data analysis based on large models is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a method and system for analyzing seismic exploration data based on a large-scale model. This invention fully utilizes the learning ability of a time-series large-scale model to learn complex seismic waveforms and cross-track correlations, enabling continuous analysis and overall judgment of large-scale seismic exploration data. It has the advantages of high automation, good analysis continuity, and strong result stability.

[0008] A seismic exploration data analysis method based on a large model according to an embodiment of the present invention includes the following steps: S1. Acquire multi-source data from seismic exploration and preprocess it to form a seismic state sequence; S2. Perform sliding window segmentation on the earthquake state sequence, use time-delay cross-correlation to register adjacent seismic trace window segments, combine continuous registered segments, and construct a training sample set; S3. Based on the training sample set, using the MOMENT class pre-trained time series basic model as the backbone, the window segments are Patch encoded, and the cross-track registration relationship is embedded into the segment representation. The encoded state is recursively corrected, and the large earthquake analysis model is obtained through backpropagation training. S4. Using a large seismic analysis model, multi-layer coding is performed on the seismic state sequence to be analyzed. Based on the coded state, a state transfer operator is fitted, and the state is recursively extrapolated along the seismic trace direction. The subsequent recursively extrapolated state is then corrected based on the extrapolated state and the corresponding propagation deviation. S5. Calculate the response state likelihood of the window segment based on the corrected recursive state, and construct the conditional transition probability between adjacent seismic traces based on the propagation deviation. Combine the response state likelihood and the conditional transition probability to form the path probability. Use the Viterbi algorithm for global decoding to obtain the exploration data analysis results. S6. Obtain the feedback data corresponding to the exploration data analysis results, and update the large seismic analysis model using incremental gradient descent.

[0009] Optionally, the preprocessing specifically includes: Seismic waveform data is acquired by seismic detectors deployed along the exploration lines, and the acquisition location and sampling time of each seismic trace are recorded synchronously by the seismic data acquisition system. The correlation is established according to the seismic trace identifier and sampling time. Seismic waveform data from each seismic trace are resampled to a uniform sampling interval, and the sampling times are time-aligned. The sliding median is used to identify and replace outlier sampling points, and interpolation is used to fill in consecutive missing intervals using adjacent valid sampling points; Bandpass filtering was performed on the seismic waveform data after anomaly processing and missing data completion, and the amplitude scale of each seismic trace was standardized. The processed seismic waveform data are organized according to the seismic trace arrangement order and sampling time order to form a seismic state sequence.

[0010] Optionally, S2 specifically includes: S21. According to the arrangement order of seismic traces, the seismic state sequence corresponding to each seismic trace is divided into equal-length sliding windows, with adjacent windows maintaining partial overlap, to obtain window segments of each seismic trace. S22. Change the time offset point by point for the window segment corresponding to the time position in the adjacent seismic trace, calculate the cross-correlation coefficient under each time offset, and take the time offset corresponding to the maximum cross-correlation coefficient as the response time delay. S23. Move the corresponding window segment according to the response time delay, perform time registration on the adjacent seismic trace window segments, and record the response time delay corresponding to each registered segment in sequence along the seismic trace arrangement direction to form a response time delay sequence. S24. Perform median filtering on the response time delay sequence to remove isolated time delay points, and continuously connect the remaining registration segments along the seismic trace arrangement direction. Divide the segments into combinations at the points where the connection is interrupted to construct a training sample set.

[0011] Optionally, S3 specifically includes: S31. Divide the window segment in the training sample set into equal-length Patches, use one-dimensional convolution along the sampling direction to extract the local waveform changes in each Pat, and combine the arrangement position of the Pat in the window segment to form the Pat encoding state. S32. Based on the cross-trace registration relationship, associate the corresponding patches in adjacent seismic traces, encode the response time delay and seismic trace interval as cross-trace constraint quantities, and adjust the coupling strength between the encoded states of the corresponding patches based on the cross-trace constraint quantities. S33. In the coding layers of the MOMENT class pre-trained temporal base model, calculate the propagation residual between the mutually registered Patch coding states, and adjust the propagation residual by gating according to the coupling strength. Then, add the adjusted propagation residual to the subsequent coding state and complete the recursive correction of the coding state layer by layer. S34. Based on the recursive correction of the coding state, calculate the propagation consistency loss between the mutual registration patches, propagate the propagation consistency loss backward along each coding layer, and iteratively update the model parameters to obtain the large earthquake analysis model.

[0012] Optionally, S33 specifically includes: S331. According to the cross-channel registration relationship, obtain the coding state of the mutually registered patches in the current coding layer one by one, calculate the state difference between the corresponding coding states, and form the propagation residual. S332. Jointly gate the propagation residual and the coupling strength, adjust the gate coefficient corresponding to the propagation residual according to the coupling strength, and correct the gate coefficient using the change range of the propagation residual. S333. The amplitude of the propagation residual is adjusted based on the modified gating coefficient, and the adjusted propagation residual is superimposed on the coding state of the Patch corresponding to the next seismic trace to form the modified coding state. S334. Pass the corrected coding state to the next coding layer, and recalculate the propagation residual based on the coding state of the mutually registered patches in the next coding layer. Repeat the gating adjustment and state superposition until the recursive correction of each coding layer is completed.

[0013] Optionally, S4 specifically includes: S41. Using a large-scale seismic analysis model, multi-layer coding is performed on the seismic state sequence to be analyzed. According to the arrangement order of seismic traces, the window segment coding states of adjacent seismic traces are combined accordingly to form continuous coding state pairs. S42. Based on continuous coded state pairs, the least squares method with regular constraints is used to solve the state transition coefficients from the previous seismic trace coded state to the next seismic trace coded state. The state transition coefficients are repeatedly corrected according to the transition residuals of each coded state pair to form a state transfer operator. S43. Apply the state transfer operator to the coding state of the current seismic trace to obtain the recursive state of the next seismic trace. Differ the recursive state from the coding state corresponding to the next seismic trace to form the current propagation deviation. S44. Based on the current propagation deviation, adjust the deviation feedback ratio by combining the direction and magnitude of the previous propagation deviation change, feed the adjusted propagation deviation back to the current recursive state, and use the corrected recursive state as the starting state for the next seismic trace. Repeat the state transition and deviation feedback along the seismic trace direction.

[0014] Optionally, S42 specifically includes: S421. Establish the state transition relationship according to the arrangement order of consecutive coded state pairs, take the previous coded state as the starting point of the transition, take the corresponding subsequent coded state as the fitting target, and initialize the state transition coefficients and recursive gain. S422. Introduce consecutive encoded state pairs in sequence, generate corresponding transition states using the current state transition coefficients, update the recursive gain according to the degree of deviation between the transition state and the fitted target, and use the updated recursive gain to correct the state transition coefficients. S423. In the continuous correction process of state transition coefficients, regular constraints are introduced to shrink the transition coefficients whose change amplitude continues to increase, and the subsequent correction weights are dynamically adjusted according to the transition residuals of each continuous coded state pair. S424. Repeatedly update the gain, correct the state transition coefficients, and perform regular shrinkage along the continuous encoding state. When the state transition coefficients remain stable for multiple consecutive rounds, the fitting is terminated, and the current state transition coefficients are used to form the state transfer operator.

[0015] Optionally, S5 specifically includes: S51. Response state scoring is performed on the corrected recursive state, the scores of each response state are probability-normalized, and the probability distribution is corrected by combining the state changes on the continuous seismic traces to obtain the response state likelihood. S52. Configure state transition penalties based on the magnitude and direction of propagation deviation, reduce penalties for state transitions with continuous propagation deviation, increase penalties for state transitions with abrupt propagation deviation, and convert the penalty results of each candidate transition into conditional transition probabilities. S53. Using Viterbi forward recursion, the likelihood of the current response state is jointly calculated with the conditional transition probability corresponding to each preceding state. For each response state, the preceding state with the highest path probability is retained, and the position of the corresponding preceding state is recorded synchronously. S54. Starting from the response state with the highest probability at the end path, backtrack step by step according to the recorded preceding state positions, connect the optimal response states corresponding to each seismic trace, form the global optimal response path, and obtain the exploration data analysis results.

[0016] Optionally, S6 specifically includes: S61. Obtain the feedback data corresponding to the exploration data analysis results. According to the seismic trace arrangement order and window position, match the response status in the feedback data with the global decoding results one by one to complete the cross-trace alignment. S62. Compare the aligned response state with the global decoding result, generate state differences according to the direction and magnitude of state change, and associate continuous state differences along the seismic trace direction. S63. Adjust the correction direction and magnitude of the corresponding propagation deviation based on the continuous state difference, and feed the corrected propagation deviation back to the coding state recursion process of the large earthquake analysis model. S64. Calculate the incremental training loss using the corrected propagation bias, update the model parameters by using incremental gradient descent along each coding layer of the large seismic analysis model, and continue to perform seismic state sequence analysis using the updated model.

[0017] According to an embodiment of the present invention, a seismic exploration data analysis system based on a large model includes: The data processing module is used to acquire multi-source data from seismic exploration and preprocess it to form a seismic state sequence; The sample construction module is used to perform sliding window segmentation on the earthquake state sequence, register adjacent seismic trace window segments using time-delay cross-correlation, combine continuous registered segments, and construct a training sample set. The model training module is used to perform patch encoding on window segments based on the training sample set, using the MOMENT class pre-trained time series basic model as the backbone, embedding cross-track registration relationships into segment representations, recursively correcting the encoding state, and obtaining a large earthquake analysis model through backpropagation training. The state recursion module is used to employ a large seismic analysis model to encode the seismic state sequence to be analyzed in multiple layers, fit a state transfer operator based on the encoded state, perform state recursion along the seismic trace direction, and correct subsequent recursion states based on the recursion state and the corresponding propagation deviation. The global decoding module is used to calculate the response state likelihood of a window segment based on the corrected recursive state, and to construct the conditional transition probability between adjacent seismic traces based on the propagation deviation. The response state likelihood and the conditional transition probability are combined to form the path probability. The Viterbi algorithm is used for global decoding to obtain the exploration data analysis results. The feedback update module is used to obtain feedback data corresponding to the exploration data analysis results and update the large seismic analysis model using incremental gradient descent.

[0018] The beneficial effects of this invention are: First, this invention fully utilizes the learning ability of large time-series models for complex seismic waveforms, incorporating the correlation information between adjacent seismic traces into the analysis process. This enables the model to better identify continuously changing seismic responses across traces, reducing the impact of local noise and waveform migration on the analysis results.

[0019] Secondly, by continuously transmitting and correcting the state changes between different seismic traces, this invention enables the analysis results to be no longer limited to a single seismic trace or local segment, which helps to maintain the spatial continuity and stability of the seismic response and improves the overall analytical capability of complex seismic data.

[0020] Finally, this invention combines global path judgment and feedback update mechanism to perform overall correction on local analysis results, and continuously adjusts the model based on subsequent feedback, so that it can adapt to different exploration areas and data changes, thereby improving the accuracy, stability and automation of seismic exploration data analysis. Attached Figure Description

[0021] 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: Figure 1 This is an overall flowchart of a seismic exploration data analysis method based on a large model proposed in this invention; Figure 2 This is a flowchart illustrating the construction and training process of a large seismic analysis model for a large-scale seismic exploration data analysis method proposed in this invention. Figure 3 This is a module structure diagram of a seismic exploration data analysis system based on a large model proposed in this invention. Detailed Implementation

[0022] 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.

[0023] refer to Figures 1-2 A seismic exploration data analysis method based on a large model includes the following steps: S1. Acquire multi-source data from seismic exploration and preprocess it to form a seismic state sequence; S2. Perform sliding window segmentation on the earthquake state sequence, use time-delay cross-correlation to register adjacent seismic trace window segments, combine continuous registered segments, and construct a training sample set; S3. Based on the training sample set, using the MOMENT class pre-trained time series basic model as the backbone, the window segments are Patch encoded, and the cross-track registration relationship is embedded into the segment representation. The encoded state is recursively corrected, and the large earthquake analysis model is obtained through backpropagation training. S4. Using a large seismic analysis model, multi-layer coding is performed on the seismic state sequence to be analyzed. Based on the coded state, a state transfer operator is fitted, and the state is recursively extrapolated along the seismic trace direction. The subsequent recursively extrapolated state is then corrected based on the extrapolated state and the corresponding propagation deviation. S5. Calculate the response state likelihood of the window segment based on the corrected recursive state, and construct the conditional transition probability between adjacent seismic traces based on the propagation deviation. Combine the response state likelihood and the conditional transition probability to form the path probability. Use the Viterbi algorithm for global decoding to obtain the exploration data analysis results. S6. Obtain the feedback data corresponding to the exploration data analysis results, and update the large seismic analysis model using incremental gradient descent.

[0024] In this embodiment, the preprocessing specifically includes: Seismic waveform data is acquired by seismic detectors deployed along the exploration lines, and the acquisition location and sampling time of each seismic trace are recorded synchronously by the seismic data acquisition system. The correlation is established according to the seismic trace identifier and sampling time. Seismic waveform data from each seismic trace are resampled to a uniform sampling interval, and the sampling times are time-aligned. The sliding median is used to identify and replace outlier sampling points, and interpolation is used to fill in consecutive missing intervals using adjacent valid sampling points; Bandpass filtering was performed on the seismic waveform data after anomaly processing and missing data completion, and the amplitude scale of each seismic trace was standardized. The processed seismic waveform data are organized according to the seismic trace arrangement order and sampling time order to form a seismic state sequence; In this embodiment, the seismic waveform data of each seismic trace are uniformly resampled to a 2ms sampling interval and aligned according to the sampling completion time. A sliding median window is formed by seven consecutive sampling points, and abnormal sampling points that deviate from the trend of the median change in the window are replaced with the median of the window. For data segments with no more than 3 consecutive missing sampling points, linear interpolation is used to complete the missing data segments by using the valid sampling data at both ends of the missing interval; for data segments with more than 3 consecutive missing sampling points, invalidation is marked. A bandpass filter of 5Hz to 80Hz was used to preserve the effective frequency band of the earthquake and suppress low-frequency drift and high-frequency interference. Then, the amplitude data was standardized on a per-trace basis and continuously organized according to the seismic trace arrangement order and 2ms sampling interval to form an earthquake state sequence.

[0025] In this embodiment, S2 specifically includes: S21. According to the arrangement order of seismic traces, the seismic state sequence corresponding to each seismic trace is divided into equal-length sliding windows, with adjacent windows maintaining partial overlap, to obtain window segments of each seismic trace. S22. Change the time offset point by point for the window segment corresponding to the time position in the adjacent seismic trace, calculate the cross-correlation coefficient under each time offset, and take the time offset corresponding to the maximum cross-correlation coefficient as the response time delay. S23. Move the corresponding window segment according to the response time delay, perform time registration on the adjacent seismic trace window segments, and record the response time delay corresponding to each registered segment in sequence along the seismic trace arrangement direction to form a response time delay sequence. S24. Perform median filtering on the response time delay sequence to remove isolated time delay points, and continuously connect the remaining registration segments along the seismic trace arrangement direction. Divide the segments into combinations at the points where the connection is interrupted to construct a training sample set.

[0026] In this embodiment, S3 specifically includes: S31. Divide the window segment in the training sample set into equal-length Patches, use one-dimensional convolution along the sampling direction to extract the local waveform changes in each Pat, and combine the arrangement position of the Pat in the window segment to form the Pat encoding state. S32. Based on the cross-trace registration relationship, associate the corresponding patches in adjacent seismic traces, encode the response time delay and seismic trace interval as cross-trace constraint quantities, and adjust the coupling strength between the encoded states of the corresponding patches based on the cross-trace constraint quantities. S33. In the coding layers of the MOMENT class pre-trained temporal base model, calculate the propagation residual between the mutually registered Patch coding states, and adjust the propagation residual by gating according to the coupling strength. Then, add the adjusted propagation residual to the subsequent coding state and complete the recursive correction of the coding state layer by layer. S34. Based on the recursive correction of the coding state, calculate the propagation consistency loss between the mutual registration patches, propagate the propagation consistency loss backward along each coding layer, and iteratively update the model parameters to obtain the large earthquake analysis model. In this embodiment, each training sample consists of multiple consecutively registered seismic trace window segments, and each window segment is organized according to the seismic trace arrangement order. Each window segment contains 256 consecutive sampling points. The window segment is divided into a Patch with a length of 16 sampling points. Each window segment forms 16 Patches. A one-dimensional convolution with a kernel length of 3 is used to extract local waveform changes along the sampling direction. The convolution result and the Patch arrangement position are jointly encoded into a 256-dimensional Patch encoding state. The large earthquake analysis model is composed of a Patch coding layer, a cross-channel constraint layer, a multi-layer temporal coding layer and a gated correction layer connected in sequence. The multi-layer temporal coding layer adopts the encoder backbone of the MOMENT-type pre-trained temporal basic model. In this embodiment, 6 temporal coding layers are set. The Patch encoding state first enters the cross-trace constraint layer, associates the corresponding Patch of adjacent seismic traces according to the cross-trace registration relationship recorded in the training samples, and encodes the response time delay and seismic trace interval into the corresponding associated cross-trace constraint quantity. The cross-trace constraint quantity is used to adjust the coupling strength between the registration Patch.

[0027] A gating correction layer is set after the 2nd, 4th and 6th temporal coding layers to form a propagation residual for the coding state of the mutually registered patches. The propagation residual and the corresponding coupling strength are used together in the gating calculation. The retention ratio is increased for the propagation residual with stable coupling relationship and decreased for the propagation residual with weakened coupling relationship. The propagation residual after gating is fed back to the coding state of the corresponding patch of the next seismic trace, and then the corrected coding state is passed to the next temporal coding layer, so that the cross-trace registration relationship continues to participate in the state recursion in the multi-layer coding process.

[0028] During the training phase, MOMENT-class pre-trained parameters are used as the initial parameters of the temporal coding layer. The first two temporal coding layers are frozen, and the last four temporal coding layers, cross-channel constraint layer, and gating correction layer are updated. The propagation consistency loss is constructed by recursively correcting the state differences of the mutual registration patches. The AdamW optimizer is used for backpropagation. The training batch is set to 32 training samples and the initial learning rate is set to 0.0001. After each round of training, the propagation consistency loss is calculated on the validation samples. The training ends when the propagation consistency loss decreases by less than 1% for 5 consecutive rounds, and the large earthquake analysis model is obtained. After training, the large seismic analysis model receives a sequence of window segments organized according to the seismic trace order and outputs the final-level patch encoding state corresponding to each window segment. The final-level patch encoding state serves as the encoding basis for subsequent state transfer operator fitting and cross-trace state recursion.

[0029] In this embodiment, S33 specifically includes: S331. According to the cross-channel registration relationship, obtain the coding state of the mutually registered patches in the current coding layer one by one, calculate the state difference between the corresponding coding states, and form the propagation residual. S332. Jointly gate the propagation residual and the coupling strength, adjust the gate coefficient corresponding to the propagation residual according to the coupling strength, and correct the gate coefficient using the change range of the propagation residual. S333. The amplitude of the propagation residual is adjusted based on the modified gating coefficient, and the adjusted propagation residual is superimposed on the coding state of the Patch corresponding to the next seismic trace to form the modified coding state. S334. Pass the corrected coding state to the next coding layer, and recalculate the propagation residual based on the coding state of the mutually registered patches in the next coding layer. Repeat the gating adjustment and state superposition until the recursive correction of each coding layer is completed.

[0030] In this embodiment, S4 specifically includes: S41. Using a large-scale seismic analysis model, multi-layer coding is performed on the seismic state sequence to be analyzed. According to the arrangement order of seismic traces, the window segment coding states of adjacent seismic traces are combined accordingly to form continuous coding state pairs. S42. Based on continuous coded state pairs, the least squares method with regular constraints is used to solve the state transition coefficients from the previous seismic trace coded state to the next seismic trace coded state. The state transition coefficients are repeatedly corrected according to the transition residuals of each coded state pair to form a state transfer operator. S43. Apply the state transfer operator to the coding state of the current seismic trace to obtain the recursive state of the next seismic trace. Differ the recursive state from the coding state corresponding to the next seismic trace to form the current propagation deviation. S44. Based on the current propagation deviation, adjust the deviation feedback ratio by combining the direction and magnitude of the previous propagation deviation change, feed the adjusted propagation deviation back to the current recursive state, and use the corrected recursive state as the starting state for the next seismic trace. Repeat the state transition and deviation feedback along the seismic trace direction.

[0031] In this embodiment, S42 specifically includes: S421. Establish the state transition relationship according to the arrangement order of consecutive coded state pairs, take the previous coded state as the starting point of the transition, take the corresponding subsequent coded state as the fitting target, and initialize the state transition coefficients and recursive gain. S422. Introduce consecutive encoded state pairs in sequence, generate corresponding transition states using the current state transition coefficients, update the recursive gain according to the degree of deviation between the transition state and the fitted target, and use the updated recursive gain to correct the state transition coefficients. S423. In the continuous correction process of state transition coefficients, regular constraints are introduced to shrink the transition coefficients whose change amplitude continues to increase, and the subsequent correction weights are dynamically adjusted according to the transition residuals of each continuous coded state pair. S424. Repeatedly recursively update the gain, correct the state transition coefficients, and perform regular shrinkage along the continuous coding state. When the state transition coefficients remain stable for multiple consecutive rounds, the fitting is terminated, and the current state transition coefficients are formed into a state transfer operator. In this embodiment, no less than 16 sets of continuous coded states are selected from the continuous seismic traces to fit the state transfer operator. Each set adopts the 256-dimensional final-level coded state output by the large seismic analysis model. The state transition coefficients are initialized to unit transition states, the recursive gain is initialized to 0.1, and continuous coded state pairs are introduced one by one according to the seismic trace arrangement order. The recursive gain is updated and the state transition coefficients are corrected according to the degree of deviation between the transition state and the corresponding subsequent coded state. During the state transition coefficient correction process, the regularization coefficient is set to 0.001 to shrink the continuously increasing state transition coefficients, and the subsequent correction weights are adjusted according to the transition residuals of each consecutive encoded state pair. When the overall change of the state transition coefficients is less than 1% for three consecutive rounds, or when the fitting is terminated after 20 rounds of correction, the current state transition coefficients are used as the state transfer operator.

[0032] In this embodiment, S5 specifically includes: S51. Response state scoring is performed on the corrected recursive state, the scores of each response state are probability-normalized, and the probability distribution is corrected by combining the state changes on the continuous seismic traces to obtain the response state likelihood. S52. Configure state transition penalties based on the magnitude and direction of propagation deviation, reduce penalties for state transitions with continuous propagation deviation, increase penalties for state transitions with abrupt propagation deviation, and convert the penalty results of each candidate transition into conditional transition probabilities. S53. Using Viterbi forward recursion, the likelihood of the current response state is jointly calculated with the conditional transition probability corresponding to each preceding state. For each response state, the preceding state with the highest path probability is retained, and the position of the corresponding preceding state is recorded synchronously. S54. Starting from the response state with the highest probability at the end path, backtrack step by step according to the recorded preceding state positions, connect the optimal response states corresponding to each seismic trace, form the global optimal response path, and obtain the exploration data analysis results.

[0033] In this embodiment, S6 specifically includes: S61. Obtain the feedback data corresponding to the exploration data analysis results. According to the seismic trace arrangement order and window position, match the response status in the feedback data with the global decoding results one by one to complete the cross-trace alignment. S62. Compare the aligned response state with the global decoding result, generate state differences according to the direction and magnitude of state change, and associate continuous state differences along the seismic trace direction. S63. Adjust the correction direction and magnitude of the corresponding propagation deviation based on the continuous state difference, and feed the corrected propagation deviation back to the coding state recursion process of the large earthquake analysis model. S64. Calculate the incremental training loss using the corrected propagation bias, update the model parameters by using incremental gradient descent along each coding layer of the large seismic analysis model, and continue to perform seismic state sequence analysis using the updated model.

[0034] refer to Figure 3 A seismic exploration data analysis system based on a large model, comprising: The data processing module is used to acquire multi-source data from seismic exploration and preprocess it to form a seismic state sequence; The sample construction module is used to perform sliding window segmentation on the earthquake state sequence, register adjacent seismic trace window segments using time-delay cross-correlation, combine continuous registered segments, and construct a training sample set. The model training module is used to perform patch encoding on window segments based on the training sample set, using the MOMENT class pre-trained time series basic model as the backbone, embedding cross-track registration relationships into segment representations, recursively correcting the encoding state, and obtaining a large earthquake analysis model through backpropagation training. The state recursion module is used to employ a large seismic analysis model to encode the seismic state sequence to be analyzed in multiple layers, fit a state transfer operator based on the encoded state, perform state recursion along the seismic trace direction, and correct subsequent recursion states based on the recursion state and the corresponding propagation deviation. The global decoding module is used to calculate the response state likelihood of a window segment based on the corrected recursive state, and to construct the conditional transition probability between adjacent seismic traces based on the propagation deviation. The response state likelihood and the conditional transition probability are combined to form the path probability. The Viterbi algorithm is used for global decoding to obtain the exploration data analysis results. The feedback update module is used to obtain feedback data corresponding to the exploration data analysis results and update the large seismic analysis model using incremental gradient descent.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to the analysis of multichannel seismic data in a certain onshore seismic exploration scenario. The data acquired in this scenario is characterized by a large number of seismic traces, complex waveform variations, and significant local noise interference. The same reflection response exhibits varying degrees of time shift between adjacent seismic traces, and some continuous responses also show local amplitude attenuation, waveform distortion, and short-distance interruptions. When using analysis methods based on a single seismic trace or a fixed neighborhood, it is easy to divide the originally continuous seismic response into multiple independent segments, leading to missed or incorrect assessments in areas of weak response.

[0036] Therefore, this embodiment selects 480 continuous seismic data for analysis and compiles 126 sets of continuous seismic responses that have been manually verified, including 84 sets of normal continuous responses, 27 sets of local weak responses, and 15 sets of responses with obvious noise interference, to verify the analysis effect under different conditions.

[0037] In this scenario, the acquired multi-channel seismic waveforms are first uniformly organized and quality-processed, forming a continuous seismic state sequence according to the order of the seismic channels. Subsequently, windowed analysis is performed on each seismic channel, and waveform segments with corresponding relationships between adjacent seismic channels are found through time-delay cross-correlation. Segments that can maintain continuous changes across channels are combined as training samples.

[0038] The large-scale seismic analysis model uses a MOMENT-type pre-trained time series basic model as its backbone, and incorporates cross-trace registration relationships into segment encoding during the training process, enabling the model to perceive the extension and changes of the same seismic response between adjacent seismic traces while learning local waveform characteristics.

[0039] After the model training is completed, the earthquake data to be analyzed is continuously fed into the large earthquake analysis model to obtain the deep coding state corresponding to each seismic trace. Then, the cross-trace state transfer relationship is fitted based on the state changes between adjacent seismic traces.

[0040] For state shifts occurring during propagation, the recursive results are corrected using the actual encoded state, enabling seismic traces that are locally affected by noise or have weak responses to restore their correlation by utilizing consecutive responses. Subsequently, probabilistic evaluations are performed on different response states, and global path determination is made by comprehensively considering the continuous changes between adjacent seismic traces, in order to reduce the impact of local anomalies of individual seismic traces on the final analysis results.

[0041] Of the 126 reference responses, this invention correctly identified 118, achieving an overall continuous response identification accuracy of 93.7%. Among these, 24 out of 27 local weak responses were correctly identified, achieving a correct identification rate of 88.9%; and 13 out of 15 responses with significant noise interference correctly maintained cross-channel correlation, achieving a correct correlation rate of 86.7%.

[0042] In contrast, when using a conventional time series model that does not incorporate cross-track state transfer and global continuous analysis, 110 groups were correctly identified overall, 19 groups of local weak responses were correctly identified, and 9 groups of noise interference responses were correctly associated. Further statistical analysis revealed that the cross-track response continuity rate of this invention reached 91.3%, the average cross-track time delay deviation was 3.2 ms, the response path interruption rate was 6.1%, and the isolated anomaly misjudgment rate was 4.6%, all demonstrating more stable continuous analysis performance.

[0043] Table 1 Comparison of Seismic Continuous Response Analysis Results

[0044] As can be seen from the data in Table 1, the present invention outperforms conventional time series models in terms of overall identification accuracy, cross-track continuity, and analysis performance under weak response and noise interference conditions. At the same time, it reduces cross-track time delay bias, response path interruption rate, and isolated anomaly misjudgment rate, indicating that the present invention can more stably maintain the correlation between continuous seismic responses and different seismic traces, thereby improving the overall reliability of analysis of complex seismic exploration data.

[0045] 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 analyzing seismic exploration data based on a large model, characterized in that, Includes the following steps: S1. Acquire multi-source seismic exploration data and preprocess it to form a seismic state sequence; S2. Perform sliding window segmentation on the earthquake state sequence, use time-delay cross-correlation to register adjacent seismic trace window segments, combine continuous registered segments, and construct a training sample set; S3. Based on the training sample set, using the MOMENT class pre-trained time series basic model as the backbone, the window segments are Patch encoded, and the cross-track registration relationship is embedded into the segment representation. The encoded state is recursively corrected, and the large earthquake analysis model is obtained through backpropagation training. S4. Using a large seismic analysis model, multi-layer coding is performed on the seismic state sequence to be analyzed. Based on the coded state, a state transfer operator is fitted, and the state is recursively extrapolated along the seismic trace direction. The subsequent recursively extrapolated state is then corrected based on the extrapolated state and the corresponding propagation deviation. S5. Calculate the response state likelihood of the window segment based on the corrected recursive state, and construct the conditional transition probability between adjacent seismic traces based on the propagation deviation. Combine the response state likelihood and the conditional transition probability to form the path probability. Use the Viterbi algorithm for global decoding to obtain the exploration data analysis results. S6. Obtain the feedback data corresponding to the exploration data analysis results, and update the large seismic analysis model using incremental gradient descent.

2. The method for analyzing seismic exploration data based on a large model according to claim 1, characterized in that, The preprocessing specifically includes: Seismic waveform data is acquired by seismic detectors deployed along the exploration lines, and the acquisition location and sampling time of each seismic trace are recorded synchronously by the seismic data acquisition system. The correlation is established according to the seismic trace identifier and sampling time. The seismic waveform data of each seismic trace are resampled to a uniform sampling interval, and the sampling times are time-aligned. The sliding median is used to identify and replace outlier sampling points, and interpolation is used to fill in consecutive missing intervals using adjacent valid sampling points; Bandpass filtering was performed on the seismic waveform data after anomaly processing and missing data completion, and the amplitude scale of each seismic trace was standardized. The processed seismic waveform data are organized according to the seismic trace arrangement order and sampling time order to form a seismic state sequence.

3. The seismic exploration data analysis method based on a large model according to claim 1, characterized in that, S2 specifically includes: S21. According to the arrangement order of seismic traces, the seismic state sequence corresponding to each seismic trace is divided into equal-length sliding windows, with adjacent windows maintaining partial overlap, to obtain window segments of each seismic trace. S22. Change the time offset point by point for the window segment corresponding to the time position in the adjacent seismic trace, calculate the cross-correlation coefficient under each time offset, and take the time offset corresponding to the maximum cross-correlation coefficient as the response time delay. S23. Move the corresponding window segment according to the response time delay, perform time registration on the adjacent seismic trace window segments, and record the response time delay corresponding to each registered segment in sequence along the seismic trace arrangement direction to form a response time delay sequence. S24. Perform median filtering on the response time delay sequence to remove isolated time delay points, and continuously connect the remaining registration segments along the seismic trace arrangement direction. Divide the segments into combinations at the points where the connection is interrupted to construct a training sample set.

4. The seismic exploration data analysis method based on a large model according to claim 1, characterized in that, S3 specifically includes: S31. Divide the window segment in the training sample set into equal-length Patches, use one-dimensional convolution along the sampling direction to extract the local waveform changes in each Pat, and combine the arrangement position of the Pat in the window segment to form the Pat encoding state. S32. Based on the cross-trace registration relationship, associate the corresponding patches in adjacent seismic traces, encode the response time delay and seismic trace interval as cross-trace constraint quantities, and adjust the coupling strength between the encoded states of the corresponding patches based on the cross-trace constraint quantities. S33. In the coding layers of the MOMENT class pre-trained temporal base model, calculate the propagation residual between the mutually registered Patch coding states, and adjust the propagation residual by gating according to the coupling strength. Then, add the adjusted propagation residual to the subsequent coding state and complete the recursive correction of the coding state layer by layer. S34. Based on the recursive correction of the coding state, calculate the propagation consistency loss between the mutual registration patches, propagate the propagation consistency loss backward along each coding layer, and iteratively update the model parameters to obtain the large earthquake analysis model.

5. The seismic exploration data analysis method based on a large model according to claim 4, characterized in that, S33 specifically includes: S331. According to the cross-channel registration relationship, obtain the coding state of the mutually registered patches in the current coding layer one by one, calculate the state difference between the corresponding coding states, and form the propagation residual. S332. Jointly gate the propagation residual and the coupling strength, adjust the gate coefficient corresponding to the propagation residual according to the coupling strength, and correct the gate coefficient using the change range of the propagation residual. S333. The amplitude of the propagation residual is adjusted based on the modified gating coefficient, and the adjusted propagation residual is superimposed on the coding state of the Patch corresponding to the next seismic trace to form the modified coding state. S334. Pass the corrected coding state to the next coding layer, and recalculate the propagation residual based on the coding state of the mutually registered patches in the next coding layer. Repeat the gating adjustment and state superposition until the recursive correction of each coding layer is completed.

6. The seismic exploration data analysis method based on a large model according to claim 1, characterized in that, S4 specifically includes: S41. Using a large-scale seismic analysis model, multi-layer coding is performed on the seismic state sequence to be analyzed. According to the arrangement order of seismic traces, the window segment coding states of adjacent seismic traces are combined accordingly to form continuous coding state pairs. S42. Based on continuous coded state pairs, the least squares method with regular constraints is used to solve the state transition coefficients from the previous seismic trace coded state to the next seismic trace coded state. The state transition coefficients are repeatedly corrected according to the transition residuals of each coded state pair to form a state transfer operator. S43. Apply the state transfer operator to the coding state of the current seismic trace to obtain the recursive state of the next seismic trace. Differ the recursive state from the coding state corresponding to the next seismic trace to form the current propagation deviation. S44. Based on the current propagation deviation, adjust the deviation feedback ratio by combining the direction and magnitude of the previous propagation deviation change, feed the adjusted propagation deviation back to the current recursive state, and use the corrected recursive state as the starting state for the next seismic trace. Repeat the state transition and deviation feedback along the seismic trace direction.

7. The seismic exploration data analysis method based on a large model according to claim 6, characterized in that, S42 specifically includes: S421. Establish the state transition relationship according to the arrangement order of consecutive coded state pairs, take the previous coded state as the starting point of the transition, take the corresponding subsequent coded state as the fitting target, and initialize the state transition coefficients and recursive gain. S422. Introduce consecutive encoded state pairs in sequence, generate corresponding transition states using the current state transition coefficients, update the recursive gain according to the degree of deviation between the transition state and the fitted target, and use the updated recursive gain to correct the state transition coefficients. S423. In the continuous correction process of state transition coefficients, regular constraints are introduced to shrink the transition coefficients whose change amplitude continues to increase, and the subsequent correction weights are dynamically adjusted according to the transition residuals of each continuous coded state pair. S424. Repeatedly update the gain, correct the state transition coefficients, and perform regular shrinkage along the continuous encoding state. When the state transition coefficients remain stable for multiple consecutive rounds, the fitting is terminated, and the current state transition coefficients are used to form the state transfer operator.

8. The seismic exploration data analysis method based on a large model according to claim 1, characterized in that, S5 specifically includes: S51. Response state scoring is performed on the corrected recursive state, the scores of each response state are probability-normalized, and the probability distribution is corrected by combining the state changes on the continuous seismic traces to obtain the response state likelihood. S52. Configure state transition penalties based on the magnitude and direction of propagation deviation, reduce penalties for state transitions with continuous propagation deviation, increase penalties for state transitions with abrupt propagation deviation, and convert the penalty results of each candidate transition into conditional transition probabilities. S53. Using Viterbi forward recursion, the likelihood of the current response state is jointly calculated with the conditional transition probability corresponding to each preceding state. For each response state, the preceding state with the highest path probability is retained, and the position of the corresponding preceding state is recorded synchronously. S54. Starting from the response state with the highest probability at the end path, backtrack step by step according to the recorded preceding state positions, connect the optimal response states corresponding to each seismic trace, form the global optimal response path, and obtain the exploration data analysis results.

9. The seismic exploration data analysis method based on a large model according to claim 1, characterized in that, S6 specifically includes: S61. Obtain the feedback data corresponding to the exploration data analysis results. According to the seismic trace arrangement order and window position, match the response status in the feedback data with the global decoding results one by one to complete the cross-trace alignment. S62. Compare the aligned response state with the global decoding result, generate state differences according to the direction and magnitude of state change, and associate continuous state differences along the seismic trace direction. S63. Adjust the correction direction and magnitude of the corresponding propagation deviation based on the continuous state difference, and feed the corrected propagation deviation back to the coding state recursion process of the large earthquake analysis model. S64. Calculate the incremental training loss using the corrected propagation bias, update the model parameters by using incremental gradient descent along each coding layer of the large seismic analysis model, and continue to perform seismic state sequence analysis using the updated model.

10. A seismic exploration data analysis system based on a large model, comprising executing the seismic exploration data analysis method based on a large model as described in any one of claims 1 to 9, characterized in that, include: The data processing module is used to acquire multi-source data from seismic exploration and preprocess it to form a seismic state sequence; The sample construction module is used to perform sliding window segmentation on the earthquake state sequence, register adjacent seismic trace window segments using time-delay cross-correlation, combine continuous registered segments, and construct a training sample set. The model training module is used to perform patch encoding on window segments based on the training sample set, using the MOMENT class pre-trained time series basic model as the backbone, embedding cross-track registration relationships into segment representations, recursively correcting the encoding state, and obtaining a large earthquake analysis model through backpropagation training. The state recursion module is used to employ a large seismic analysis model to encode the seismic state sequence to be analyzed in multiple layers, fit a state transfer operator based on the encoded state, perform state recursion along the seismic trace direction, and correct subsequent recursion states based on the recursion state and the corresponding propagation deviation. The global decoding module is used to calculate the response state likelihood of a window segment based on the corrected recursive state, and to construct the conditional transition probability between adjacent seismic traces based on the propagation deviation. The response state likelihood and the conditional transition probability are combined to form the path probability. The Viterbi algorithm is used for global decoding to obtain the exploration data analysis results. The feedback update module is used to obtain feedback data corresponding to the exploration data analysis results and update the large seismic analysis model using incremental gradient descent.

Images