Search

CN-121995464-A - Seismic parameter inversion method based on RBF proxy model and multi-algorithm collaborative optimization

CN121995464ACN 121995464 ACN121995464 ACN 121995464ACN-121995464-A

Abstract

The invention provides an earthquake parameter inversion method based on RBF proxy model and multi-algorithm collaborative optimization, which relates to the technical field of earthquake service, and comprises the steps of establishing an RBF proxy model, carrying out local real evaluation and proxy verification, dynamically dividing global exploration and local refinement stages according to the optimal adaptability change rate, and adopting a PSO particle swarm inertia weight linear decrementing strategy to smoothly transit; and in the global and local stages, continuously monitoring the inferior solution acceptance rate and cross entropy sampling distribution covariance of the annealing algorithm, and dynamically adjusting. The method can realize rapid prediction of complex stratum parameters and observed waveforms, avoid misleading search caused by agent failure, can perform extensive global exploration in the early stage and realize local refinement in the later stage through multi-algorithm cooperation and dynamic division optimization stages, effectively maintain search activity and diversity, remarkably improve search efficiency and inversion reliability, and reduce the dependence on expensive forward calculation.

Inventors

  • XIN JUNSHENG
  • ZHANG HENG

Assignees

  • 中国雅江集团有限公司
  • 中国科学院青藏高原研究所

Dates

Publication Date
20260508
Application Date
20251229

Claims (10)

  1. 1. The seismic parameter inversion method based on the RBF proxy model and the multi-algorithm collaborative optimization is characterized by comprising the following steps: Acquiring original seismic observation data, and constructing an RBF proxy model after standardization, thereby performing inversion iterative optimization; in the inversion iteration process, the time stamp of the sharing state among algorithms is acquired in real time, proxy verification is executed, the RBF proxy model adjustment requirement is determined, and a dynamic division optimization stage is carried out; When the optimization stage is a global exploration stage, determining an annealing temperature regulation strategy and a cross entropy distribution regulation strategy; when the optimization stage is a local refinement stage, determining a multi-algorithm cooperative adjustment strategy and executing; and after receiving the inversion execution signal, performing closed-loop optimization inversion to obtain a global optimal inversion result.
  2. 2. The seismic parameter inversion method based on the RBF proxy model and the multi-algorithm collaborative optimization of claim 1, wherein the construction of the RBF proxy model is characterized by comprising the following specific analysis method: Carrying out standardization processing on original seismic observation data, determining a reasonable range of inversion parameters according to geological priori knowledge, and generating initial sample points which are representative and uniformly cover the whole parameter space in the parameter space by using Latin hypercube sampling technology; Performing forward modeling calculation on each initial sample point, and evaluating the fitting degree of the initial sample points and observation data; And selecting a sample point with the best fitting effect from the initial sample points as a central point of the radial basis function model, constructing an RBF proxy model by using a cubic spline basis function, and determining the optimal weight parameter of the RBF proxy model by solving a linear equation set.
  3. 3. The seismic parameter inversion method based on the cooperative optimization of the RBF proxy model and the multiple algorithms as set forth in claim 1, wherein the executing proxy check determines the adjustment requirement of the RBF proxy model, and the specific analysis method is as follows: Constructing an initial sample set by each initial sample point, so as to perform multi-algorithm iterative optimization; And acquiring the time stamp of the shared state among algorithms in real time, and when the time stamp state read by any algorithm is detected to be inconsistent with the current time stamp, temporarily suspending the inter-algorithm data interaction of the corresponding algorithm, and firstly executing proxy verification, wherein the specific analysis process is as follows: Acquiring a current iteration stratum parameter vector as a sample, and inputting the current iteration stratum parameter vector into a forward calculation module for simulation to obtain a simulated seismic waveform; comparing the simulated seismic waveform with the observed seismic waveform, and calculating a mean square error; extracting a preset mean square error threshold value in a database; If the mean square error is greater than or equal to the mean square error threshold, marking the RBF proxy model regulation requirement as requirement regulation, regulating the RBF kernel width correction proxy model through the deviation value of the mean square error and the mean square error threshold, and refreshing the local state cache; And if the mean square error is smaller than the mean square error threshold, recording the RBF proxy model regulation requirement as no regulation, suspending using the sample to update the cross algorithm, and triggering the local state refresh.
  4. 4. The seismic parameter inversion method based on the RBF proxy model and the multi-algorithm collaborative optimization according to claim 1, wherein the dynamic partitioning optimization stage is performed by the following specific analysis method: calculating the optimal fitness change rate of each iteration: extracting a preset optimal adaptability change rate threshold value in a database; Dividing the optimization stage into a pre-global exploration stage when the change rate of the optimal fitness is greater than or equal to the change rate threshold; And when the optimal adaptability change rate is smaller than the change rate threshold, dividing the optimization stage into a later local refinement stage.
  5. 5. The seismic parameter inversion method based on the RBF proxy model and the multi-algorithm collaborative optimization of claim 1, further comprising performing PSO algorithm parameter adaptive adjustment, wherein the specific analysis method comprises the following steps: Extracting an initial inertia weight, a final inertia weight and a maximum iteration number which are preset in a database; And determining the proportional position of the current inertia weight between the initial inertia weight and the final inertia weight according to the proportion of the current iteration times to the maximum iteration times, thereby determining the current inertia weight of the PSO particle swarm.
  6. 6. The seismic parameter inversion method based on the collaborative optimization of RBF proxy model and multi-algorithm according to claim 1, wherein the determined annealing temperature adjustment strategy and cross entropy distribution adjustment strategy are as follows: In the early global exploration stage, the inferior solution acceptance rate of the acquisition annealing algorithm is marked as the first inferior solution acceptance rate of the annealing algorithm and the cross entropy sampling distribution covariance is marked as the cross entropy first sampling distribution covariance; Extracting a preset inferior solution acceptance rate threshold and a covariance threshold in a database; if the first inferior solution acceptance rate of the annealing algorithm is lower than the inferior solution acceptance rate threshold, marking as a global adjustment first condition; If the cross entropy first sampling distribution covariance is smaller than or equal to a covariance threshold, marking as a global adjustment second condition; if the first condition of global regulation exists and the second condition of global regulation does not exist, the global exploration phase regulation strategy is recorded as an annealing temperature regulation strategy; If the global adjustment second condition exists and the global adjustment first condition does not exist, marking the global exploration phase adjustment strategy as an execution cross entropy distribution adjustment strategy; If neither condition exists, the global exploration phase adjustment strategy is recorded as continuous iterative optimization, and an inversion execution signal is generated.
  7. 7. The seismic parameter inversion method based on the collaborative optimization of RBF proxy model and multi-algorithm according to claim 6, wherein the implementation of the annealing temperature adjustment strategy is characterized by the following specific analysis method: Obtaining an inferior solution acceptance rate deviation value according to a first inferior solution acceptance rate of the annealing algorithm and an inferior solution acceptance rate threshold value, determining an annealing temperature reduction rate reduction value according to the inferior solution acceptance rate deviation value, counting the inferior solution acceptance rate of the annealing algorithm in real time in a sliding window, and correcting a temperature attenuation coefficient according to the annealing temperature reduction rate reduction value; When the adjusted inferior solution acceptance rate of the annealing algorithm is still lower than the inferior solution acceptance rate threshold, rolling back to the last stable temperature, and determining a superposition short-term disturbance factor according to the secondary deviation value of the adjusted inferior solution acceptance rate of the annealing algorithm and the inferior solution acceptance rate threshold so as to restore the searching activity; The cross entropy distribution regulation strategy comprises the following specific analysis method: Expanding the covariance matrix and reducing the elite proportion, and determining the reducing amplitude of the elite proportion according to the maximum reducing amplitude allowed by the preset elite proportion in combination with the deviation value of the cross entropy first sampling distribution covariance and covariance threshold value, thereby obtaining an elite proportion updating value which is used in the next iteration.
  8. 8. The seismic parameter inversion method based on the collaborative optimization of RBF proxy model and multi-algorithm as set forth in claim 7, wherein said expansion covariance matrix is analyzed by the following method: Calculating an objective function value of a sample in sampling distribution obtained according to the previous iteration, and weighting the sample according to the fitness of the objective function value to obtain a new expected vector; carrying out outer product operation on the deviation vector between each sample and the expected vector, and carrying out weighted summation according to the weight of the deviation vector to form new weighted variance estimation; the original covariance matrix is replaced by the weighted variance estimation or weighted fusion is carried out on the original covariance matrix and the original covariance matrix according to a certain learning rate, so that covariance update is realized.
  9. 9. The seismic parameter inversion method based on the RBF proxy model and the multi-algorithm collaborative optimization of claim 1, wherein the multi-algorithm collaborative adjustment strategy is specifically analyzed as follows: Continuously monitoring the inferior solution acceptance rate and the cross entropy sampling distribution covariance of the annealing algorithm in the later local refinement stage, and respectively marking the inferior solution acceptance rate and the cross entropy sampling distribution covariance as the second inferior solution acceptance rate and the cross entropy sampling distribution covariance of the annealing algorithm; Extracting a preset fitness threshold value in a database; if the optimal fitness is greater than or equal to the fitness threshold, marking the multi-algorithm cooperative adjustment as non-adjustment; If the optimal fitness is smaller than the fitness threshold, marking the second inferior solution acceptance rate of the annealing algorithm lower than the inferior solution acceptance rate threshold as a local refinement first judgment criterion, and marking the cross entropy second sampling distribution covariance smaller than or equal to the covariance threshold as a local refinement second judgment criterion; If the first judgment criterion of local refinement and the second judgment criterion of local refinement do not exist, marking the cooperative adjustment of the multiple algorithms as non-adjustment, and continuing to perform iterative search to generate an inversion execution signal; if the first judgment criterion of local refinement exists and the second judgment criterion of local refinement does not exist, marking the multi-algorithm cooperative adjustment as executing an annealing temperature adjustment strategy; if the first judging criterion of local refinement does not exist and the second judging criterion of local refinement exists, marking multi-algorithm cooperative adjustment as executing a cross entropy distribution adjustment strategy; If the first judgment criterion of local refinement and the second judgment criterion of local refinement exist, the multi-algorithm cooperative adjustment is recorded as local locking adjustment.
  10. 10. The seismic parameter inversion method based on the collaborative optimization of RBF proxy model and multi-algorithm according to claim 9, wherein the local lock adjustment is performed by the following specific analysis method: Determining an annealing temperature lifting value and a freezing cycle proportion according to a deviation value of a second inferior solution acceptance rate of an annealing algorithm and a threshold value of the inferior solution acceptance rate, and determining an annealing temperature freezing cycle according to the freezing cycle proportion and a preset maximum freezing cycle; Determining the number of random samples according to the deviation value of the cross entropy second sampling distribution covariance and the covariance threshold, and determining the proportion of new random samples injected into the current cross entropy distribution neighborhood and the global range according to the deviation value of the optimal fitness and the fitness threshold; And simultaneously, a short-term tabu mechanism is started, a region with continuous local locking adjustment in the system log is extracted, a sampling weight reduction value is determined according to the local locking adjustment times of the region, and search is guided to migrate to an unexplored region.

Description

Seismic parameter inversion method based on RBF proxy model and multi-algorithm collaborative optimization Technical Field The invention relates to the technical field of seismic services, in particular to a seismic parameter inversion method based on RBF proxy model and multi-algorithm collaborative optimization. Background The seismic parameter inversion is to utilize directly observable seismic records to reversely infer physical parameters of the underground medium, namely, observe the propagation characteristics of seismic waves in the underground medium, infer stratum physical parameters including P wave speed, S wave speed and density, which determine the underground structural characteristics, reservoir properties, fault structures and underground medium changes, because the underground can not be directly observed, the core sampling is limited and the cost is extremely high, the underground structure must be inferred in the whole area by depending on the seismic inversion, and the applications such as seismic exploration, oil and gas reservoir prediction, engineering geological evaluation, seismic risk analysis and the like are realized. The method comprises the steps of deducing a reconstructed fluid saturation level-vertical orthogonal fracture medium petrophysical model based on a Chapman model of a plurality of groups of fracture-pore mediums to realize fluid saturation and two groups of orthogonal fracture medium petrophysical modeling, realizing fracture type reservoir frequency-dependent AVO forward modeling based on an anisotropic reflectivity algorithm and the fluid saturation level-vertical orthogonal fracture medium petrophysical model, calculating the frequency-dependent AVO reflection coefficient and synthetic seismic record of the fracture type reservoir under an interface reflection model, constructing a frequency-dependent AVO inversion theory based on a Bayesian frame by taking a forward flow of the frequency-dependent AVO as a drive, utilizing pre-stack seismic AVO data, introducing a simulated annealing-particle swarm algorithm in the inversion process to realize frequency-dependent AVO inversion based on the Bayesian inversion frame and carrying out reservoir parameter prediction. For example, the method for synchronous inversion of four parameters of pre-stack earthquake disclosed in Chinese patent publication No. CN106842310B comprises the steps of establishing a direct relation between a seismic wave reflection coefficient and a stratum longitudinal wave speed, a stratum transverse wave, a stratum density and a stratum absorption parameter Q on the basis of a viscoelastic medium theory, and realizing synchronous inversion of four parameters of the stratum longitudinal wave speed, the stratum transverse wave, the stratum density and the stratum absorption parameter Q through a stable pre-stack earthquake inversion algorithm. The traditional seismic parameter inversion method generally relies on numerical simulation technologies such as least square, regularization inversion or full waveform inversion, and along with development of agent models, machine learning and heuristic optimization algorithms, the seismic inversion method starts to introduce a method for constructing agent models based on observed seismic waveform data, map stratum parameters and corresponding waveform characteristics, search optimal solutions in a parameter space by combining particle swarm optimization and annealing algorithms, and in the algorithm iteration process, the algorithm feeds back optimization results, updates the agent models, so that closed-loop cooperation is formed between the models and the search algorithms. However, when the particle swarm algorithm is combined with the simulated annealing algorithm, the situation that samples generated in the optimization process of the two algorithms in a high-dimensional parameter space are distributed near a few local optimal solutions may occur, the samples lack of diversity, so that the algorithm search is easy to ignore a potential global optimal region, in order to increase the diversity of the samples, the search region is more comprehensive, the cross entropy algorithm is introduced to update probability distribution of diversified samples, but the annealing temperature of the annealing algorithm is often delayed by sharing information with the particle swarm algorithm and the cross entropy algorithm due to different updating mechanisms and step sizes, so that the annealing temperature dropping speed is out of control, the probability that the annealing algorithm accepts the inferior solutions is rapidly reduced, only the solutions which are currently seen to be better is reserved, the capability of escaping the current algorithm detection region by accidentally accepting the inferior solutions is lost, and meanwhile, when the cross entropy algorithm updates the probability distribution by using the limited superior solutions, the optimal so