A seismic data multi-model adaptive selection and scheduling method, device, equipment and medium

Abstract

The present application relates to the field of seismic data processing, and particularly relates to a seismic data multi-model adaptive selection and scheduling method and device, equipment and medium. The method comprises the following steps: acquiring seismic data flow to be processed and dividing the seismic data flow into multiple data segments; performing seismic facies identification on each data segment to determine the seismic facies category thereof; according to the identified seismic facies category, screening a matched candidate model subset from a preset model library; inputting the multi-dimensional features of the data segment into learning score units corresponding to each model in the candidate model subset to obtain the scores of the models; acquiring current available computing resource information and preset processing constraint conditions, and combining the scores of the models to determine the model set and execution mode called by the current data segment; according to the multi-dimensional features of the data segment and the execution mode, allocating computing resources, and processing the data segment through each model in the model set. The geological structure information and resource constraints are fused in the model scheduling decision, thereby improving the adaptability of seismic data processing.

Description

Technical Field

[0001] This application relates to the field of seismic data technology, and in particular to a method, apparatus, equipment and medium for adaptive selection and scheduling of multiple seismic data models. Background Technology

[0002] In seismic data processing, such as denoising, deconvolution, migration, and inversion, different computational models are often selected based on the data characteristics. A single model is insufficient to adapt to the complex, non-stationary, and nonlinear variations within seismic data, thus requiring the selection and scheduling of multiple models. Current techniques typically extract statistical features from seismic data first, then select models based on preset rules or classifiers. However, the core characteristic of seismic data is the geophysical response of the subsurface medium's reflection structure; different geological structures exhibit significant differences in reflection morphology, continuity, and amplitude variation patterns. Existing methods often focus on signal statistical attributes, lacking direct utilization of subsurface reflection structure morphology. This makes it difficult to selectively screen models based on geological type in the early stages, leading to mismatched models potentially entering subsequent stages, resulting in wasted resources or inaccurate results. Furthermore, when determining the specific set of models to be invoked and the execution method, existing methods typically rely solely on model scores for sorting or threshold selection, failing to adequately consider the system's real-time available computing resources and processing time constraints, and rarely considering the mutual influence of results between models, resulting in a lack of reasonable basis for the selection of execution modes.

[0003] Secondly, some existing methods allow users to preset scheduling parameters, such as the maximum number of concurrent models and scoring thresholds. However, these parameters are often globally fixed or fine-tuned based only on simple statistics. In reality, different seismic facies categories reflect significant differences in subsurface structural complexity, leading to varying requirements for model scheduling. Simple structures can tolerate more lenient scheduling strategies, while complex structures require more conservative and reliable execution methods. Existing technologies lack a mechanism to dynamically adjust scheduling parameters based on seismic facies categories, resulting in a lack of adaptability of scheduling strategies to different geological environments. Summary of the Invention

[0004] To address one or more problems in the prior art, the main objective of this application is to provide a method, apparatus, device, and medium for adaptive selection and scheduling of multiple seismic data models.

[0005] To achieve the aforementioned objectives, this application proposes a multi-model adaptive selection and scheduling method for seismic data, the method comprising: Acquire the seismic data stream to be processed and divide the seismic data stream into multiple data segments; Based on the segmentation results, seismic facies identification is performed on each data segment to determine the seismic facies category to which each data segment belongs; Based on the identified seismic facies category, a subset of candidate models matching the seismic facies category is selected from a pre-defined model library; The multidimensional features of the current data segment are input into the learning and scoring units corresponding to each model in the candidate model subset, and the scoring results are output through the learning and scoring units. Obtain information on currently available computing resources and preset processing constraints. Based on the scores of each model, currently available computing resources, and preset processing constraints, determine the set of models to be called and the execution mode for the current data segment. Based on the multidimensional characteristics of the current data segment and the determined execution mode, computing resources are allocated to each model in the model set, and the current data segment is processed by each model in the model set.

[0006] This application also provides a seismic data multi-model adaptive selection and scheduling device, including: The acquisition module is used to acquire the seismic data stream to be processed and divide the seismic data stream into multiple data segments; The identification module is used to identify seismic facies for each data segment based on the segmentation results, and determine the seismic facies category to which each data segment belongs; The filtering module is used to filter out a subset of candidate models that match the identified seismic facies category from a preset model library. The input module is used to input the multidimensional features of the current data segment into the learning and scoring unit corresponding to each model in the candidate model subset, and output the scoring result through the learning and scoring unit; The determination module is used to obtain information on currently available computing resources and preset processing constraints. Based on the scores of each model, currently available computing resources, and preset processing constraints, it determines the set of models to be called and the execution mode for the current data segment. The allocation module is used to allocate computing resources to each model in the model set according to the multidimensional characteristics of the current data segment and the determined execution mode, and to process the current data segment through each model in the model set.

[0007] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0008] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0009] The seismic data multi-model adaptive selection and scheduling method, apparatus, device, and medium of this application introduce seismic facies identification, allowing the system to first understand the geological structure type of the current data segment, and then filter out matching candidate models accordingly, eliminating a large number of unreasonable models from the source. Subsequently, a refined matching degree evaluation is performed through a learnable scoring unit, and combined with real-time available computing resources and preset processing constraints, a greedy resource simulation method is used to determine the final set of models to be called and the execution mode. Finally, processing is performed according to the allocated resources. This invention provides geological priors for seismic facies identification, data-driven adaptability through the learnable scoring unit, and engineering feasibility through resource constraints. The combined effect of these three elements allows model selection and scheduling to no longer rely on fixed rules or coarse statistical thresholds, but to adaptively and dynamically adjust with changes in geological structure and system load. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating an embodiment of the seismic data multi-model adaptive selection and scheduling method of this application; Figure 2 This is a flowchart illustrating an embodiment of the seismic data multi-model adaptive selection and scheduling method of this application; Figure 3 This is a schematic block diagram of a seismic data multi-model adaptive selection and scheduling device according to an embodiment of this application; Figure 4 This is a schematic block diagram of the structure of a computer device according to an embodiment of this application.

[0011] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0013] Reference Figure 1 This application provides a method for adaptive selection and scheduling of multiple seismic data models, the method comprising: S1. Obtain the seismic data stream to be processed and divide the seismic data stream into multiple data segments; S2. Based on the division results, perform seismic facies identification on each data segment to determine the seismic facies category to which each data segment belongs; S3. Based on the identified seismic facies category, select a subset of candidate models that match the seismic facies category from a preset model library; S4. Input the multidimensional features of the current data segment into the learning and scoring unit corresponding to each model in the candidate model subset, and output the scoring result through the learning and scoring unit. S5. Obtain information on currently available computing resources and preset processing constraints. Based on the scores of each model, currently available computing resources, and preset processing constraints, determine the set of models to be called and the execution mode for the current data segment. S6. Based on the multidimensional characteristics of the current data segment and the determined execution mode, allocate computing resources to each model in the model set, and process the current data segment through each model in the model set.

[0014] As described in steps S1-S3 above, in actual seismic data processing, the raw data is often a continuous seismic trace record, possibly corresponding to a subsurface profile several kilometers long. Processing all the data at once is not only computationally burdensome but also difficult to adapt to the lateral changes in subsurface structures. Therefore, this scheme first acquires the seismic data stream to be processed and then divides it into several shorter data segments according to time or spatial windows. The length of each data segment can be set according to actual needs, for example, corresponding to a depth range of 200 to 500 meters underground, or a recording duration of 0.5 to 2 seconds. The purpose of the division is to transform non-stationary seismic signals into multiple locally stationary or weakly varying segments, so that subsequent processing can be finely scheduled according to the characteristics of each segment. This divide-and-conquer strategy avoids the coarseness problems caused by global uniform processing and lays the foundation for using different model combinations for different geological regions. Step two, seismic facies, refers to the set of characteristics such as reflection structure morphology, phase axis continuity, and amplitude variation patterns reflected by subsurface sedimentary bodies or structures on seismic profiles. Common seismic facies categories include parallel continuous facies, mound-like progradation facies, chaotic infill facies, and lenticular facies. In this scheme, for each pre-divided data segment, seismic facies identification is performed first. The identification process typically includes two steps: first, extracting quantitative attributes that characterize the reflection morphology from the data segment, such as facies axis dip angle, continuity measure, and waveform complexity; second, inputting these attributes into a pre-trained classifier, which outputs the probability that the data segment belongs to each seismic facies category, and finally selecting the category with the highest probability as the identification result. For data segments with indistinct features and low probabilities, they can be marked as transitional facies categories. Through seismic facies identification, this scheme can clearly identify the geological structure type corresponding to the current data segment. Elevating the processing from the pure signal level to the geological structure level makes subsequent model selection no longer blind but has a clear geological orientation. Step 3: There are various models in the field of seismic data processing, such as deconvolution models for layered media, wave equation migration models for complex structures, and wavelet domain thresholding models for noise suppression. Different models exhibit significantly varying performance on seismic facies data: parallel continuous facies are best suited to simple and efficient statistical models, while chaotic facies are better suited to deep learning models with strong nonlinear fitting capabilities. This approach pre-establishes a library containing various models and configures a corresponding model index table or rule set for each seismic facies category. After seismic facies identification, the system directly retrieves several models matching the identified category from the model library, forming a subset of candidate models. This step is equivalent to preliminary pruning of the model space, excluding models that are clearly unsuitable for the current geological type, thereby reducing the computational load in subsequent scoring stages and avoiding the risk of inappropriate models being incorrectly selected.

[0015] As described in steps S4-S6 above, the candidate model subset after screening in step four may still contain multiple models. To further determine which models are best suited for the current data segment, this scheme equips each model with a learnable scoring unit. This scoring unit can be a lightweight neural network, a small regression model, or any function that can output a matching degree value. The input to the scoring unit is the multidimensional features of the current data segment. These features can be statistical features of the signal (such as spectral centroid, skewness, and kurtosis) or geometric attributes used in previous seismic facies identification. Each scoring unit outputs a score between 0 and 1 based on its own learning experience, representing the applicability of the model to the current data segment. The higher the score, the more confident the model is in processing this data segment. These scoring units are not fixed but can be updated online based on feedback from subsequent processing results, thereby gradually adapting to the geological characteristics of different work areas. Step five: In a real computing environment, resources such as the number of processor cores, memory capacity, and storage bandwidth are limited. At the same time, seismic data processing often has real-time requirements; for example, each data segment must be completed within a few seconds. Therefore, this scheme cannot rely solely on model scores when making final scheduling decisions; it must also consider the current system resource status and preset time constraints. Specifically, it first obtains information such as the number of available processor cores, memory capacity, and storage bandwidth, as well as pre-set constraints such as the maximum processing time per segment, the maximum allowed number of concurrent models, and the minimum acceptable score threshold. Then, it sorts the models in the candidate model subset according to their scores from highest to lowest. Next, it attempts to add models to the set to be called sequentially, simulating and estimating the total resource consumption and total processing time required to execute each model. If adding a new model results in simulated resources or time exceeding the available limit, or reaching the maximum allowed concurrency, the addition is stopped. The model set at this point is the final set to be called. Furthermore, the execution mode needs to be determined based on the number of models in the model set and the correlation between their results: if there is only one model, it is naturally executed exclusively; if there are multiple models and their processing results have low similarity, it indicates that they can complement each other from different perspectives, making parallel execution suitable to leverage their respective advantages; conversely, if the results are highly similar, serial execution is sufficient to avoid meaningless parallel waste. In this way, this solution maximizes its adaptability to hardware resources and real-time requirements while ensuring processing quality. Step six, after determining the set of models to be invoked and their execution modes, allocates specific computing resources to each model based on the multidimensional characteristics of the current data segment and the execution mode. This includes specifying which CPU cores to use, the amount of memory allocated, and whether GPU acceleration is enabled. Then, according to the predetermined execution mode, these models are invoked to process the current data segment. After processing, the output results are provided for subsequent fusion or direct use.This step ensures that scheduling decisions can be implemented and that resource allocation matches actual needs, avoiding resource idleness or excessive competition.

[0016] As mentioned above, this scheme transforms seismic data processing from traditional signal statistics-driven to a combined approach driven by prior geological structure and resource constraints. Specifically, traditional methods suffer from several problems: first, they fail to fully utilize information from subsurface reflection structures during model selection, leading to the inclusion of mismatched models in subsequent calculations; second, scheduling decisions often only consider model scores while ignoring system resource and time constraints; and third, the selection of execution modes lacks consideration of the relationships between models. This scheme, by introducing seismic facies identification, allows the system to first understand the geological structure type of the current data segment, and then selects matching candidate models accordingly, eliminating a large number of unreasonable models at the source. Subsequently, a refined matching degree evaluation is performed using learnable scoring units, and combined with real-time available computing resources and preset processing constraints, a greedy resource simulation method is employed to determine the final set of models to be invoked and the execution mode. Finally, processing is executed according to the allocated resources. This invention provides geological priors for seismic facies identification, data-driven adaptability through learning scoring units, and engineering feasibility through resource constraints. The combined effect of these three factors enables the model selection and scheduling to no longer rely on fixed rules or coarse statistical thresholds, but to dynamically adjust adaptively with changes in geological structure and system load.

[0017] It should be added that the training of the scoring unit can employ supervised learning. First, a large number of historical earthquake data segments are collected as training samples. For each sample, various models from the model library are used for processing, and objective evaluation metrics such as signal-to-noise ratio improvement and enhancement of phase axis continuity are used as the model's processing quality score for that sample. This quality score is normalized to the 0-1 range and used as a supervision label. Then, for each model, a regression model is trained using the multidimensional features of its corresponding training samples as input and the normalized quality score as output. This regression model can employ linear regression, support vector regression, or shallow neural network structures. After training, this regression model can serve as the scoring unit for the model, outputting predicted scores for new data segments during actual operation.

[0018] Reference Figure 2 In one embodiment, the step of performing seismic facies identification on each data segment to determine the seismic facies category to which each data segment belongs includes: S21. Extract the seismic phase feature set for each data segment. The seismic phase feature set includes in-phase axis dip angle, continuity measure, waveform complexity index, and amplitude envelope change rate. S22. Input the set of seismic phase features into a preset seismic phase classifier, and output the probability distribution of each data segment belonging to multiple preset seismic phase categories through the seismic phase classifier; S23. Based on the probability distribution, select the seismic facies category with the highest probability as the identification result of the data segment, or when the maximum probability is lower than a preset probability threshold, mark the data segment as a transitional facies category.

[0019] As described above, the first step, to distinguish different seismic facies categories, requires extracting quantitative features reflecting the morphology of the reflection structure from each data segment. The features selected in this scheme include phase axis dip, continuity measure, waveform complexity index, and amplitude envelope variation rate. Phase axis dip refers to the angle of the seismic reflection wave's phase axis relative to the horizontal direction. In parallel sedimentary strata, the dip angle is usually small and changes gradually; however, near faults or folds, the dip angle shows significant changes. The dip angle can be calculated using methods such as local slope scanning or structural tensor eigenvalue decomposition. The continuity measure characterizes the spatial extent of the phase axis. Continuous seismic reflections usually correspond to stable sedimentary layers, while fault depressions or chaotic infill areas show poor continuity. Commonly used calculation methods include calculating the maximum cross-correlation value or local similarity coefficient between adjacent traces along the phase axis direction. The waveform complexity index reflects the degree of drastic change in the waveform morphology of the seismic trace. Simple reflection waves often exhibit a single-peak sinusoidal waveform, while complex interferometric regions show multiple peaks, side lobes, or oscillatory attenuation. This indicator can be quantitatively described by calculating the decay rate or information entropy of the waveform's autocorrelation function. The amplitude envelope change rate refers to the gradient of the instantaneous amplitude envelope of the seismic signal over time. Regions with rapidly decaying amplitude envelopes often correspond to highly absorbing media or oil and gas reservoirs, while regions with stable changes reflect relatively homogeneous lithology. This feature helps distinguish different lithofacies or fluid types. When extracting these features, a sliding window approach can be used to calculate each point in the entire data segment, and then the mean, variance, peak value, and other statistical measures are taken as the final feature vector. These features are highly complementary, collectively forming an attribute set that effectively characterizes the spatial morphology of seismic facies. After obtaining the feature vector for each data segment in the second step, it needs to be converted into a specific seismic facies category. This scheme pre-trains a seismic facies classifier, which can be a common model such as a support vector machine, random forest, or shallow neural network, as long as it can output the probability distribution of the categories. Training the classifier requires a labeled seismic facies sample library. These samples are from seismic profiles with known geological conditions. Experts categorized each segment into seismic facies types based on the reflection structure morphology, such as parallel continuous facies, mound-like progradation facies, chaotic infill facies, and lenticular facies. The classifier learns the mapping relationship between features and categories, ultimately enabling automatic identification of new, unseen data segments. In actual operation, the feature vector extracted in the first step is input into the classifier, which outputs a probability vector, where each component corresponds to a confidence level for a preset seismic facies category. For example, the output of a data segment might have a probability of 0.7 for parallel continuous facies, 0.2 for mound-like progradation facies, and 0.1 for chaotic facies. This probability output method preserves uncertainty information, providing richer evidence for subsequent decision-making. After obtaining the probability distribution in the third step, this scheme employs a decision rule with a safety mechanism: under normal circumstances, the category with the highest probability is directly selected as the identification result for that data segment.In the example above, the segment was identified as a parallel continuous phase because the probability of 0.7 was the highest. However, seismic data sometimes contains regions with ambiguous features, such as transition zones between two different phases or areas with strong noise interference. In such regions, the maximum probability output by the classifier may be very low; for example, the probabilities of the three categories might be 0.35, 0.33, and 0.32, close to each other and not dominant. Forcibly assigning a category in this case is likely to be incorrect and lead to inaccurate subsequent model selection. Therefore, this scheme sets a preset probability threshold, such as 0.5 or 0.6. If the maximum probability is lower than this threshold, a specific seismic phase category is no longer forcibly assigned; instead, the data segment is marked as a transitional phase category. The transitional phase category serves as a special identifier, indicating that the subsequent scheduling module cannot rely on a single seismic phase prior for selection and needs to adopt a more conservative or comprehensive strategy. For example, the selection range of the candidate model subset can be broadened, or the fusion weight of multiple model results can be increased.

[0020] In one embodiment, the step of determining the set of models and execution mode to be invoked for the current data segment based on the scores of each model, currently available computing resources, and preset processing constraints includes: Obtain the scores of each model in the candidate model subset and sort them from highest to lowest; Obtain information on currently available computing resources and preset processing constraints; In descending order of scores, models are added to the candidate set one by one, and the computational resources required and the estimated processing time are simulated and calculated after addition. Based on the simulation results, when adding the next model would cause the required computing resources to exceed the currently available computing resources, or the estimated processing time to exceed the preset maximum processing time, or the number of models in the candidate set has reached the maximum allowed concurrency, the addition of the model to the candidate set will be stopped, and the current candidate set will be determined as the set of models to be called. The execution mode is determined based on the number of models in the model set being called and the correlation indicators between the models.

[0021] As mentioned above, after selecting a subset of candidate models through seismic facies analysis and assigning a score to each model by the learning scoring unit, the next step is to decide which models to use and how they should be executed in combination. This step faces three practical constraints: limited computational resources, a maximum processing time, and potential result conflicts or redundancy between models. To address this issue, specifically, each model output by the learning scoring unit in the first step has a score between 0 and 1, reflecting its suitability for the current data segment. This scheme first collects the scores of all models in the candidate model subset and then sorts them in descending order of score. The purpose of this is to prioritize the models most likely to perform well. The sorted sequence provides a clear priority order for subsequent greedy selection. In the second step, in a real-world operating environment, computational resources are not unlimited. This scheme needs to obtain the current system availability, mainly including the number of available processor cores, remaining memory capacity, and storage bandwidth. These values ​​can be read in real time through the operating system or runtime environment. In addition, pre-defined processing constraints need to be obtained. These constraints are set in advance by users or system administrators according to task requirements, such as the maximum allowed processing time for each data segment, the maximum number of concurrent models, and a minimum acceptable model score threshold. Models below this threshold, even if ranked highly, should not be considered because their processing performance for the current data segment may be poor. These resource information and constraints together constitute the boundary conditions for scheduling decisions, making the decision no longer a simple ranking of model quality, but a trade-off process combining actual carrying capacity. After obtaining the ranked model list and boundary conditions in the third step, this solution adopts a greedy strategy to select the final set of models to be called. Specifically, starting with the model with the highest score, models are sequentially added to a temporary set, i.e., the candidate set. For each model added, the total computing resources and estimated processing time required to execute all models in the current candidate set are simulated and calculated. The specific method of simulation calculation can be based on the historical running statistics of the models, such as recording the CPU time, memory usage, I / O bandwidth, etc., consumed by each model when processing similar data segments, and then accumulating the data for the models in the current candidate set. Alternatively, the complexity parameters of the model itself can be used for estimation, such as the number of floating-point operations in a convolutional neural network or the product of the window length and the number of channels in a traditional deconvolution. This simulation process does not require actually running the model; it can be completed simply by looking up tables or using lightweight computation, thus incurring minimal overhead. When adding a new model, three conditions are checked: whether the total required computing resources exceed the currently available resources, whether the estimated total processing time exceeds the preset maximum processing time, and whether the number of models in the candidate set has reached the maximum allowed concurrency.If any one of these three conditions is met, the process stops immediately, no further models are tried, and the current candidate set is determined as the final set of models to be called. If none of the conditions are met, the model is officially added to the candidate set, and the process continues to try the next model. Within the limits, as many high-quality models as possible are called to obtain better processing results. Because high-scoring models are prioritized, the final set is the optimal or near-optimal combination under resource constraints. After determining the set of models to be called in the fourth step, it is also necessary to decide how these models should be executed. There are three common execution methods: single-model exclusive execution, multi-model parallel execution, and multi-model serial execution. This scheme dynamically selects based on the number of models in the model set and the correlation index between models. If there is only one model in the model set, then obviously only the single-model exclusive execution mode can be used. If there are more than one model, it is necessary to further determine the correlation between the models. The correlation index is used to measure the similarity of the processing results of different models on the same data segment. If the output results of two models are very similar, for example, if they are essentially linear filtering or based on similar feature extraction methods, then running them simultaneously may not bring significant gains, but rather waste resources. In this situation, a more suitable approach is sequential execution, running one model first and using its result as input or reference for the other, forming a cascaded processing. Conversely, if the outputs of the two models have a low correlation, it indicates that they analyze the data from different perspectives—one model excels at capturing layered features, while the other excels at identifying diffraction points. In this case, parallel execution can fully leverage their respective strengths, ultimately fusing them to obtain a more comprehensive result. The specific value of the correlation index can be calculated based on the similarity between the outputs of each model on historical data segments, such as using cosine similarity, mutual information, or normalized cross-correlation coefficients. A correlation threshold is pre-set; if the calculated correlation index is below this threshold, it is considered low correlation, and a multi-model parallel execution mode is adopted; if it is not below the threshold, it is considered high correlation, and a multi-model sequential execution mode is adopted. This decision-making ensures the effective use of computing resources and avoids duplication of work or unnecessary competition.

[0022] In one embodiment, the step of determining the execution mode based on the number of models in the called model set and the correlation index between the models includes: Get the number of models in the collection of models called; If the number of models is equal to 1, then the execution mode is determined to be the single-model exclusive execution mode; If the number of models is greater than 1, then calculate the correlation index between the models in the model set; The correlation index is compared with a preset correlation threshold; Based on the comparison results, if the correlation index is lower than the preset correlation threshold, the execution mode is determined to be a multi-model parallel execution mode. If the correlation index is not lower than the preset correlation threshold, then the execution mode is determined to be the multi-model serial execution mode; The correlation index is used to characterize the similarity of the processing results of different models on the same data segment, and is calculated based on the similarity between the output results of each model on the historical data segment.

[0023] As mentioned above, after determining the final set of models to be invoked, the next step is to decide how these models should be executed. Different execution methods directly affect processing speed, resource consumption, and the final fusion effect. Specifically, the first step is the most intuitive judgment. First, count how many models are in the determined model set. This number could be one, two, or more, depending on the combined effect of resources and scoring in the previous steps. Knowing the number of models is the starting point for all subsequent judgments. The second step is that if there is only one model in the model set, there is no choice; simply let that model process the current data segment alone. This is the single-model exclusive execution mode. At this time, there is no need to consider the cooperation or resource competition between models, nor is any correlation calculation required; the process is simple and direct. The third step is that when the model set contains two or more models, the problem becomes more complex. These models may complement each other, or their functions may overlap or even interfere with each other. To determine how they should cooperate, this scheme introduces a correlation index. This index is used to measure how similar the processing results of different models on the same seismic data segment are. The specific calculation method is based on the output results of each model on historical data segments. For example, the system records the denoising results, deconvolution results, or migration imaging results generated by each model when processing past data segments. For two models, their outputs on the same historical data segment can be treated as two vectors, and then the cosine similarity, mutual information, or normalized cross-correlation coefficient between these two vectors can be calculated. If the outputs of two models on a large amount of historical data are very similar, then their correlation index is high, meaning they are doing almost the same thing. Conversely, if the outputs of two models differ greatly, for example, one model is good at suppressing random noise while the other model is good at enhancing fault edges, then their correlation index will be low. To obtain a comprehensive correlation value, the average similarity between all pairs of models in the model set can be calculated, or the maximum or minimum value can be taken, depending on the specific strategy. This scheme does not limit the unique calculation method, as long as it can reflect the overall similarity between models. The fourth step is to pre-set a correlation threshold, such as 0.6 or 0.7. This threshold can be adjusted according to the actual application scenario. When it is desirable to encourage model diversity, the threshold can be set lower, and when it is desirable to avoid redundant calculations, the threshold can be set higher. The correlation index calculated in the previous step is compared with this threshold. If the correlation index is below the preset correlation threshold, it indicates that the processing results of the various models in the model set are not very similar, and they each capture different aspects of the data. In this case, a multi-model parallel execution mode is most advantageous. Parallel execution means that all models run simultaneously without waiting for each other, each producing its own processing result. Ultimately, these results can be integrated into a single output through weighted averaging, voting, or a more intelligent fusion method.Parallel execution fully leverages the complementarity between models, enriching output information from multiple perspectives, and is particularly suitable for processing complex geological structures. If the correlation index is not lower than a preset correlation threshold, it indicates that the models in the model set are highly similar in their processing results, essentially repeating similar calculations. Continuing to execute them in parallel in this case not only wastes computational resources, but the final fusion result will not be much better than that of a single model. A more reasonable approach is to adopt a multi-model serial execution mode. Serial execution refers to running these models in a specific order, with each subsequent model receiving the output of the previous model as input or as a reference. For example, one model can be used for denoising, followed by another model for deconvolution, forming a pipeline-like cascaded processing. Serial execution is more efficient when model functions are dependent or overlapping, and can also achieve progressively refined processing effects.

[0024] It's worth noting that a frequently overlooked issue in conventional multi-model scheduling is how to execute multiple models after they have been invoked. Existing technologies often default to parallel execution or simple sequential execution, rarely proactively selecting the execution mode dynamically based on the relationships between models. This neglect leads to two undesirable consequences: firstly, when models are highly correlated, parallel execution wastes valuable computational resources and time; secondly, when models are highly complementary, sequential execution may not fully utilize their respective advantages, resulting in biased final results. This solution introduces a correlation index to quantify the relationships between models and integrate them into the decision-making process, achieving intelligent matching of execution modes. In particular, this correlation index is not arbitrarily set but calculated based on the actual output similarity of each model on historical seismic data segments, thus possessing realism and adaptability. After the system has processed for a period of time, the correlation index gradually stabilizes, accurately reflecting the intrinsic relationships between models.

[0025] In one embodiment, before the step of determining the set of models to be invoked and the execution mode for the current data segment based on the scores of each model, currently available computing resources, and preset processing constraints, the method further includes: Obtain the seismic facies category to which the current data segment belongs; Based on the seismic facies category to which the current data segment belongs, analyze the reflection structure complexity level corresponding to the seismic facies category; Based on the reflection structure complexity level corresponding to the earthquake facies category, the scores of each model in the candidate model subset are weighted and corrected; wherein, the earthquake facies category with higher reflection structure complexity is assigned a higher weight coefficient to the model that focuses on structural information processing.

[0026] As mentioned above, in the aforementioned scheme, each model in the candidate model subset obtains a score through learning a scoring unit. This score reflects the model's adaptability to the current data segment. However, the complexity of subsurface structures varies greatly depending on the seismic facies category. Some models perform well in simple tectonic regions but may struggle in complex ones; others are the opposite, excelling at handling complex reflection structures but overcomputing in simple ones. Before formally determining the model set, this scheme introduces a scoring correction step based on seismic facies category, allowing the model score to be dynamically adjusted according to the current geological complexity. This process is explained in detail below. The first step involves identifying the seismic facies for each data segment, obtaining clear category labels, such as parallel continuous facies, mound-like pre-accretionary facies, and chaotic infill facies. This step directly reads the identification result as the basis for subsequent corrections. The second step involves different seismic facies categories reflecting different subsurface sedimentary and tectonic forms, with varying degrees of reflection structure complexity. This scheme pre-defines a reflection structure complexity level for each seismic facies category. This level can be one of three grades: simple, medium, and complex, or a continuous numerical score. For example, parallel continuous facies typically correspond to stable sedimentary environments with gently dipping strata and regularly arranged reflection phase axes; this type of structure has low complexity. Mound-like progradational facies reflect progradational sedimentary bodies with internal reflections exhibiting progradational or imbricate patterns; this type of structure has moderate complexity. Disorganized infill facies often occur within fault zones, landslides, or reefs, with chaotic reflections and discontinuous or broken phase axes; this type of structure has high complexity. This complexity level can be pre-defined based on prior geological knowledge or dynamically learned by statistically analyzing the variance or error of model processing results under different seismic facies categories in historical data. Regardless of the method used, the core idea is to transform abstract seismic facies categories into a quantifiable structural complexity index. After obtaining the reflection structural complexity level in the third step, this scheme uses this information to weight and correct the scores of each model in the candidate model subset. The correction method is to assign different weight coefficients to different types of models based on their functional emphasis. Specifically, seismic data processing models can be broadly divided into two categories. One type focuses on structural information processing, such as sparse transform models for preserving fault edges, deep learning networks for identifying diffraction points, and structure-guided filtering models for enhancing the continuity of phase axes. These models are more valuable in complex tectonic regions because the requirements for structural preservation are more stringent in complex regions. Another type focuses on amplitude fidelity or statistical property processing, such as mean filtering for suppressing random noise and Wiener filtering for deconvolution. These models are sufficient for simple tectonic regions, but may actually smooth out the true structure in complex regions.Therefore, when the reflection structure is highly complex, such as encountering chaotic infill facies or mound-like preaccelerated facies, this scheme assigns higher weight coefficients to models that focus on structural information processing, thus improving their corrected scores; while models that focus on amplitude or statistical characteristics may maintain their original weights or slightly reduce them. When the reflection structure is less complex, such as encountering parallel continuous facies, there is no need to particularly emphasize structural models; the original scores can be maintained or efficiency-type models can be appropriately increased. The specific calculation method for weighted correction can be multiplicative, that is, the original score multiplied by a weight coefficient related to the complexity level and model type; or additive, that is, adding a correction term to the original score. The value of the weight coefficient can be set empirically or automatically optimized during operation through reinforcement learning. After this correction step, the score of each model is no longer a purely data-driven result, but a comprehensive evaluation value that integrates prior knowledge of seismic facies. Subsequent steps will use these corrected scores for sorting and greedy selection.

[0027] In one embodiment, after the steps of acquiring the seismic data stream to be processed and dividing the seismic data stream into multiple data segments, the method further includes: Identify whether each data segment belongs to a fault region or a diffraction region; If the data segment belongs to a fault region or a diffraction region, the execution mode of the data segment is set to a multi-model parallel execution mode, and at least one wave equation forward model is forcibly invoked to participate in the processing.

[0028] As mentioned above, not all regions in seismic data share similar physical characteristics. Fault regions exhibit discontinuities or breaks in the phase axis, while diffracted regions display hyperbolic diffracted wave morphologies. The seismic responses of these two types of regions are highly unique, often proving difficult for conventional data-driven models to accurately handle, easily leading to false structures or omissions of genuine geological information. This solution adds an identification and forced processing method after data segmentation and before formal scheduling to address these special regions. Specifically, after data segmentation, the system first performs a rapid detection on each segment to determine if it contains fault or diffracted features. Typical characteristics of faults include a clear break or termination of the phase axis and inconsistent reflection wave morphologies on both sides, which can be automatically identified through methods such as coherence volume analysis, variance volume calculation, or dip change detection. Typical characteristics of diffracted waves include a hyperbolic morphology and relatively weak energy, easily submerged in conventional stacking profiles, requiring detection using plane wave deconstruction filters or diffracted wave separation techniques. This identification step does not require very high precision; it only needs to roughly determine the possibility of a fault in the data segment or whether it contains diffracted wave energy. Lightweight and fast algorithms can be used to avoid introducing excessive computational overhead. The identification result is a Boolean flag or confidence score for subsequent decision-making. In the second step, if a data segment is identified as a fault region or diffraction region, this scheme adopts a mandatory scheduling strategy instead of relying on conventional scoring and resource simulation processes. Specifically, regardless of the data segment's score in the candidate model subset or the current computational resource constraints, the system will force the execution mode of that data segment to be set to multi-model parallel execution mode. This means that multiple models will process the data segment simultaneously, rather than relying solely on the highest-scoring model or a serial approach. More importantly, the system will require that the model set used in this call must contain at least one wave equation forward model. The wave equation forward model is a computational method based on physical laws; it generates theoretical seismic records by simulating the propagation of seismic waves in the subsurface medium. Unlike purely data-driven models, the wave equation forward model does not depend on training samples but is directly controlled by the subsurface velocity model and the wave equation itself. In fault and diffraction regions, data-driven models may fail due to insufficient similar samples in the training data, but wave equation forward modeling can guarantee the geological rationality of the output results from a physical perspective. The purpose of forced parallel execution is to allow multiple models to process the same data from different angles, producing multiple results. These results can corroborate each other, and the reliability of other models can be judged by comparing the results of the wave equation forward model. The mandatory inclusion of the wave equation forward model aims to introduce a physical constraint benchmark, preventing purely data-driven models from outputting results that deviate from physical laws. It should be noted that this mandatory strategy does not completely negate conventional scheduling logic. Under the premise of meeting the mandatory requirements, the system can still supplement and call other models based on scoring and resource availability.For example, after ensuring at least one forward model of the wave equation and using parallel mode, data-driven models with higher scores can still be added, as long as resources allow. If the data segment is not identified as a fault or diffraction region, this step has no impact, and the system continues to execute according to the original scheduling process, that is, selecting the model set and execution mode normally based on scores, resources, and constraints.

[0029] In one embodiment, the steps for determining the required computing resources and estimated processing time after the simulation calculation are added include: Obtain the seismic phase category to which the current data segment belongs, and determine the corresponding reflection structure complexity level based on the seismic phase category; Based on the complexity level of the reflection structure, the computational resource security factor and the duration security factor are determined respectively; wherein, the higher the complexity level of the reflection structure, the larger the values ​​of the computational resource security factor and the duration security factor. The required computing resources obtained from the simulation calculation are weighted using the aforementioned computing resource security factor to obtain the adjusted required computing resources. The estimated processing time obtained from the simulation calculation is weighted using the aforementioned time safety factor to obtain the adjusted estimated processing time. Based on the adjustment results, the adjusted required computing resources are compared with the currently available computing resources, and the adjusted estimated processing time is compared with the preset maximum processing time.

[0030] As mentioned above, when simulating and estimating the computational resources and processing time required after adding a model, the conventional approach is often to use fixed values ​​or historical averages. However, the complexity of the reflection structure in seismic data has a significant impact on the actual execution efficiency of the model. Processing a parallel, continuous data segment with a high signal-to-noise ratio, the model may converge with only a few iterations; while processing a data segment with messy reflections or containing a large number of diffracted waves, the same model may require more iterations, more refined scan parameters, or a larger computation window to achieve acceptable results. If this difference is ignored, resource estimates will be overly optimistic, leading to exceeding available resources or processing time limits during actual operation. This solution introduces a safety factor mechanism based on seismic facies categories in the simulation calculation stage, making the estimated results closer to real-world requirements. The first step, seismic facies identification, has already been completed in the previous steps, assigning each data segment a specific seismic facies category. This solution first obtains this category label, and then determines the reflection structure complexity level corresponding to that seismic facies category based on a pre-established mapping relationship. This level can be divided into three grades: simple, medium, and complex, or it can be a continuously varying value. For example, parallel continuous facies are typically classified as simple, progradational facies as medium, and chaotic infill facies or rift zones as complex. Different complexity levels represent the degree of disorder and drastic structural changes in the reflection morphology within the data segment, which directly affect the computational load of the model. After establishing the complexity levels in the second step, this scheme determines two safety factors: one for computational resources and one for processing time. These two safety factors share the characteristic that their values ​​increase with the higher the complexity of the reflection structure. The safety factor is greater than or equal to 1; for example, a safety factor of 1.0 can be set for simple structures, 1.3 for medium structures, and 1.6 or higher for complex structures. The physical implication of this design is that more complex geological structures may require more memory, more computational cores, and longer runtime for the model. The safety factor acts as a buffer, amplifying the original estimate based on historical averages, making the prediction closer to the worst-case scenario rather than the average level. The computational resource safety factor and the time safety factor can be set separately because the expansion effect of complex structures on resources and time may not be completely synchronized. A deep neural network model might require significantly more memory to store intermediate feature maps in complex regions, while the processing time might only increase by a few tens of percent; conversely, an iterative optimization model might have a substantial increase in processing time but little change in memory requirements. Allowing both to be set independently provides more granular control. In the third step of regular simulation computation, the system estimates, based on the model's historical running records or complexity metrics, the additional processor cores, memory capacity, storage bandwidth, and estimated processing time required to add the current model to the candidate set.This solution multiplies these estimated values ​​by the corresponding safety factor to obtain the adjusted required computing resources and the adjusted estimated processing time. Specifically, the adjusted required computing resources equal the original required computing resources calculated from the simulation multiplied by the computing resource safety factor. The adjusted estimated processing time equals the original estimated processing time calculated from the simulation multiplied by the processing time safety factor. This multiplication operation is equivalent to adding a protective boundary to the estimated values, making the originally potentially ideal values ​​more conservative and reliable. After the adjustment is completed in the fourth step, this solution no longer uses the original simulation estimates, but instead uses these adjusted values ​​to compare with the currently available computing resources and the preset maximum processing time. That is to say, the original determination of whether available resources were exceeded used an unadjusted optimistic estimate; now, a conservative estimate amplified by the safety factor is used. The result of this conservative estimation is that the system will recognize earlier that resources or time may be insufficient, thus stopping the addition of models to the candidate set. Although this may reduce the number of models ultimately called, it gains a higher safety margin, preventing resource overruns, processing timeouts, or even system crashes during actual operation. In simple data regions, the safety factor is close to 1, so the adjusted value is not much different from the original value, and the scheduling strategy is basically unaffected, allowing the system to still call as many high-scoring models as possible. In complex data regions, the safety factor is significantly greater than 1, so the adjusted value will be significantly higher than the original value. The system will then be more cautious in controlling the number of models, and may even call only the highest-scoring model to avoid overload on complex data segments.

[0031] In one embodiment, this method is deployed in a multi-agent collaborative system. This system breaks down seismic interpretation operations into multiple sub-tasks according to different stages, with a master agent responsible for the overall process orchestration and scheduling, and each sub-agent undertaking a specific role.

[0032] Specifically, the SGY data parsing sub-agent is responsible for reading seismic data, confirming parameters such as trace number and sampling interval, and generating seismic profiles for viewing. The interpretation model selection sub-agent is responsible for executing the core model scheduling process of this invention, including seismic facies identification, candidate model screening, scoring calculation, resource constraint judgment, and determination of the final model set and execution mode. The well logging analysis sub-agent is responsible for the well-seismic combined analysis, and the report generation sub-agent is responsible for automatically outputting interpretation results. The model scheduling method of this invention mainly runs in the interpretation model selection sub-agent. This sub-agent internally maintains multiple processing models as callable tools. Seismic facies identification can utilize classification rules in the knowledge base, while model scoring and resource scheduling are executed according to the aforementioned steps. The main agent dynamically allocates computing resources according to the task progress and aggregates and merges the processing results of each sub-agent. The system also integrates a human-in-the-loop mechanism. The system will proactively request manual intervention in the following situations: first, when the maximum confidence score for seismic facies identification is lower than a preset threshold, and the data segment is marked as a transitional facies; second, when the data segment is identified as a fault or diffraction area, and the system forces parallel execution, manual confirmation is needed to determine which model's result is more reasonable; and third, when all models' scores are low or closely related, making decision-making difficult. The results of manual confirmation will be recorded and used to update the parameters of the scoring units, forming feedback optimization. In addition, the system integrates a seismic interpretation knowledge base. The knowledge base stores definitions of professional terms, standard characteristic descriptions of each seismic facies, descriptions of applicable model scenarios, and historical processing cases. During model selection and manual confirmation, the knowledge base can provide reference and also understand and answer users' professional questions. The scheduling method of this invention, as the core capability of the sub-agent, ensures the automation and adaptability of model selection; on the other hand, the main agent's process orchestration and human-in-the-loop intervention mechanism compensate for potential judgment errors in complex geological scenarios. The combination of these two aspects retains the efficiency advantages of the algorithm while introducing the professional judgment capabilities of humans.

[0033] Reference Figure 3 This application also provides a seismic data multi-model adaptive selection and scheduling device, including: Acquisition module 1 is used to acquire the seismic data stream to be processed and divide the seismic data stream into multiple data segments; The identification module 2 is used to identify seismic facies for each data segment based on the segmentation results, and determine the seismic facies category to which each data segment belongs; The filtering module 3 is used to filter out a subset of candidate models that match the identified seismic facies category from a preset model library. Input module 4 is used to input the multidimensional features of the current data segment into the learning scoring unit corresponding to each model in the candidate model subset, and output the scoring result through the learning scoring unit; Module 5 is used to obtain information on currently available computing resources and preset processing constraints. Based on the scores of each model, currently available computing resources, and preset processing constraints, it determines the set of models to be called and the execution mode for the current data segment. The allocation module 6 is used to allocate computing resources to each model in the model set according to the multidimensional characteristics of the current data segment and the determined execution mode, and to process the current data segment through each model in the model set.

[0034] As described above, it is understood that each component of the seismic data multi-model adaptive selection and scheduling device proposed in this application can realize the function of any of the seismic data multi-model adaptive selection and scheduling methods described above, and the specific structure will not be described in detail.

[0035] Reference Figure 4 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores monitoring data and other data. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-model adaptive selection and scheduling method for seismic data.

[0036] The processor described above executes the aforementioned seismic data multi-model adaptive selection and scheduling method, including: acquiring the seismic data stream to be processed and dividing the seismic data stream into multiple data segments; based on the division results, performing seismic facies identification on each data segment to determine the seismic facies category to which each data segment belongs; according to the identified seismic facies category, selecting a subset of candidate models matching the seismic facies category from a preset model library; inputting the multidimensional features of the current data segment into the learning scoring unit corresponding to each model in the candidate model subset, and outputting the scoring result through the learning scoring unit; acquiring information on currently available computing resources and preset processing constraints, and determining the model set and execution mode to be called by the current data segment based on the scores of each model, the currently available computing resources, and the preset processing constraints; allocating computing resources to each model in the model set according to the multidimensional features of the current data segment and the determined execution mode, and processing the current data segment through each model in the model set.

[0037] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements a method for adaptive selection and scheduling of multiple seismic data models, including the following steps: acquiring a seismic data stream to be processed and dividing the seismic data stream into multiple data segments; based on the division results, performing seismic facies identification on each data segment to determine the seismic facies category to which each data segment belongs; according to the identified seismic facies category, selecting a subset of candidate models matching the seismic facies category from a preset model library; inputting the multidimensional features of the current data segment into the learning scoring unit corresponding to each model in the candidate model subset, and outputting the scoring result through the learning scoring unit; acquiring information on currently available computing resources and preset processing constraints, and determining the set of models to be called and the execution mode of the current data segment based on the scores of each model, the currently available computing resources, and the preset processing constraints; allocating computing resources to each model in the set of models based on the multidimensional features of the current data segment and the determined execution mode, and processing the current data segment through each model in the set of models.

[0038] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0039] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0040] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A multi-model adaptive selection and scheduling method for seismic data, characterized in that, The method includes: Acquire the seismic data stream to be processed and divide the seismic data stream into multiple data segments; Based on the segmentation results, seismic facies identification is performed on each data segment to determine the seismic facies category to which each data segment belongs; Based on the identified seismic facies category, a subset of candidate models matching the seismic facies category is selected from a pre-defined model library; The multidimensional features of the current data segment are input into the learning and scoring units corresponding to each model in the candidate model subset, and the scoring results are output through the learning and scoring units. Obtain information on currently available computing resources and preset processing constraints. Based on the scores of each model, currently available computing resources, and preset processing constraints, determine the set of models to be called and the execution mode for the current data segment. Based on the multidimensional characteristics of the current data segment and the determined execution mode, computing resources are allocated to each model in the model set, and the current data segment is processed by each model in the model set.

2. The seismic data multi-model adaptive selection and scheduling method according to claim 1, characterized in that, The steps of identifying seismic facies for each data segment and determining the seismic facies category to which each data segment belongs include: Extract the seismic phase feature set for each data segment. The seismic phase feature set includes in-phase axis dip, continuity measure, waveform complexity index, and amplitude envelope change rate. The set of seismic facies features is input into a preset seismic facies classifier, and the seismic facies classifier outputs the probability distribution of each data segment belonging to multiple preset seismic facies categories. Based on the probability distribution, the seismic facies category with the highest probability is selected as the identification result of the data segment, or when the maximum probability is lower than a preset probability threshold, the data segment is marked as a transitional facies category.

3. The seismic data multi-model adaptive selection and scheduling method according to claim 1, characterized in that, The step of determining the set of models to be invoked and the execution mode for the current data segment based on the scores of each model, currently available computing resources, and preset processing constraints includes: Obtain the scores of each model in the candidate model subset and sort them from highest to lowest; Obtain information on currently available computing resources and preset processing constraints; In descending order of scores, models are added to the candidate set one by one, and the computational resources required and the estimated processing time are simulated and calculated after addition. Based on the simulation results, when adding the next model would cause the required computing resources to exceed the currently available computing resources, or the estimated processing time to exceed the preset maximum processing time, or the number of models in the candidate set has reached the maximum allowed concurrency, the addition of the model to the candidate set will be stopped, and the current candidate set will be determined as the set of models to be called. The execution mode is determined based on the number of models in the model set being called and the correlation indicators between the models.

4. The seismic data multi-model adaptive selection and scheduling method according to claim 3, characterized in that, The step of determining the execution mode based on the number of models in the called model set and the correlation indicators between the models includes: Get the number of models in the collection of models called; If the number of models is equal to 1, then the execution mode is determined to be the single-model exclusive execution mode; If the number of models is greater than 1, then calculate the correlation index between the models in the model set; The correlation index is compared with a preset correlation threshold; Based on the comparison results, if the correlation index is lower than the preset correlation threshold, the execution mode is determined to be a multi-model parallel execution mode. If the correlation index is not lower than the preset correlation threshold, then the execution mode is determined to be the multi-model serial execution mode; The correlation index is used to characterize the similarity of the processing results of different models on the same data segment, and is calculated based on the similarity between the output results of each model on the historical data segment.

5. The seismic data multi-model adaptive selection and scheduling method according to claim 3, characterized in that, Before the step of determining the set of models and execution mode to be invoked for the current data segment based on the scores of each model, currently available computing resources, and preset processing constraints, the method further includes: Obtain the seismic facies category to which the current data segment belongs; Based on the seismic facies category to which the current data segment belongs, analyze the reflection structure complexity level corresponding to the seismic facies category; Based on the reflection structure complexity level corresponding to the earthquake facies category, the scores of each model in the candidate model subset are weighted and corrected; wherein, the earthquake facies category with higher reflection structure complexity is assigned a higher weight coefficient to the model that focuses on structural information processing.

6. The seismic data multi-model adaptive selection and scheduling method according to claim 1, characterized in that, After the steps of acquiring the seismic data stream to be processed and dividing the seismic data stream into multiple data segments, the method further includes: Identify whether each data segment belongs to a fault region or a diffraction region; If the data segment belongs to a fault region or a diffraction region, the execution mode of the data segment is set to a multi-model parallel execution mode, and at least one wave equation forward model is forcibly invoked to participate in the processing.

7. The seismic data multi-model adaptive selection and scheduling method according to claim 3, characterized in that, The simulation calculations, including the required computing resources and estimated processing time, include the following steps: Obtain the seismic phase category to which the current data segment belongs, and determine the corresponding reflection structure complexity level based on the seismic phase category; Based on the complexity level of the reflection structure, the computational resource security factor and the duration security factor are determined respectively; wherein, the higher the complexity level of the reflection structure, the larger the values ​​of the computational resource security factor and the duration security factor. The required computing resources obtained from the simulation calculation are weighted using the aforementioned computing resource security factor to obtain the adjusted required computing resources. The estimated processing time obtained from the simulation calculation is weighted using the aforementioned time safety factor to obtain the adjusted estimated processing time. Based on the adjustment results, the adjusted required computing resources are compared with the currently available computing resources, and the adjusted estimated processing time is compared with the preset maximum processing time.

8. A seismic data multi-model adaptive selection and scheduling device, characterized in that, include: The acquisition module is used to acquire the seismic data stream to be processed and divide the seismic data stream into multiple data segments; The identification module is used to identify seismic facies for each data segment based on the segmentation results, and determine the seismic facies category to which each data segment belongs; The filtering module is used to filter out a subset of candidate models that match the identified seismic facies category from a preset model library. The input module is used to input the multidimensional features of the current data segment into the learning and scoring unit corresponding to each model in the candidate model subset, and output the scoring result through the learning and scoring unit; The determination module is used to obtain information on currently available computing resources and preset processing constraints. Based on the scores of each model, currently available computing resources, and preset processing constraints, it determines the set of models to be called and the execution mode for the current data segment. The allocation module is used to allocate computing resources to each model in the model set according to the multidimensional characteristics of the current data segment and the determined execution mode, and to process the current data segment through each model in the model set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images