Search

CN-122017986-A - Air gun seismic source array inversion optimization method combining deep learning calculation model and particle swarm optimization

CN122017986ACN 122017986 ACN122017986 ACN 122017986ACN-122017986-A

Abstract

The invention discloses an air gun seismic source array inversion optimization method combining a deep learning calculation model and a particle swarm algorithm, and belongs to the technical field of marine geophysical exploration. The method comprises the steps of 1, constructing a high-fidelity physical sample library, 2, constructing a forward calculation model of a deep neural network and training, 3, constructing a multi-target constrained particle swarm inversion frame based on the forward calculation model in the step 2, and 4, realizing global optimization of the multi-target constrained particle swarm inversion frame in the step 3 based on a particle swarm algorithm until the optimal air gun array arrangement combination parameters meeting the target spectrum requirements are obtained. The method is used for solving the problems of slow calculation and strong inversion multi-solution of the existing air gun array optimization design.

Inventors

  • ZHANG SHUAI
  • Ning Yichi
  • REN SHAOFEI
  • ZHANG AMAN

Assignees

  • 哈尔滨工程大学
  • 哈尔滨工程大学三亚南海创新发展基地

Dates

Publication Date
20260512
Application Date
20260120

Claims (10)

  1. 1. An air gun seismic source array inversion optimization method combining a deep learning calculation model and a particle swarm algorithm is characterized by comprising the following steps of: Step1, constructing a high-fidelity physical sample library; Step 2, constructing a forward calculation model of the deep neural network and training; Step 3, constructing a multi-target constrained particle swarm inversion frame based on the forward direction calculation model in the step 2; and 4, realizing global optimization of the multi-target constrained particle swarm inversion frame in the step 3 based on a particle swarm algorithm until the optimal air gun array arrangement combination parameters meeting the target spectrum requirements are obtained.
  2. 2. The method according to claim 1, wherein step 1 is specifically that a latin hypercube sampling strategy is adopted to generate input parameter sets containing different air gun numbers, volumes, pressures, depths of deposition and spatial positions, and corresponding far-field pressure sweep frequency spectrums are calculated as output labels based on an air gun bubble dynamics model adopting a Zhang equation.
  3. 3. The method for optimizing inversion of air gun source array according to claim 1, wherein in step 2, the Deep neural network is one or a combination of a graph neural network, a long-term and short-term memory network adopting an attention mechanism, a fully-connected neural network or a Deep set network, the air gun array parameters are taken as input, the frequency spectrum curve is taken as output, and the model can establish a fast forward mapping relation from 'array parameters' to 'frequency spectrum characteristics'.
  4. 4. The method for optimizing inversion of an air gun seismic source array according to claim 1, wherein the step 3 is specifically to build an inversion optimization design model with the goal of reducing corner frequency and fusing engineering constraints.
  5. 5. The method for optimizing inversion of an air gun seismic source array according to claim 4, wherein the step 3 specifically comprises the following steps: step 3.1, defining an optimization target; And 3.2, defining the fitness function.
  6. 6. The method of claim 5, wherein the step 3.1 is specifically performed assuming that the design objective is to obtain an array with a corner frequency of 3Hz, and that parameters are required to be adjusted within a fixed parameter range; the step 3.2 is specifically to construct the following fitness function F (X): wherein the first term is a spectral error term For forcibly meeting the 3Hz requirement, wherein Is the angular frequency calculated by calculating the frequency spectrum output by the model, A second term is a regularization term For screening optimal solutions from multiple solutions, the third term being a geometric constraint term If the air gun distance D ij is smaller than the safety distance D min , a huge penalty is applied to prevent air gun bubbles from fusing: 。
  7. 7. the method for inverting and optimizing the air gun source array according to claim 1 is characterized in that the step 4 is specifically that firstly an initial particle group is randomly generated, meaning of single particles is represented by a group of parameters to be optimized, secondly fitness of each particle is calculated by utilizing a forward depth neural network model, and searching of the particle group to an optimal solution area is guided by continuously and iteratively updating the speed and the position of the particles until the optimal air gun array arrangement combination parameters meeting the target spectrum requirements are obtained.
  8. 8. The method for optimizing inversion of an air gun seismic source array according to claim 7, wherein the step 4 specifically comprises the steps of: step 4.1, particle coding and initialization; Step 4.2, calling a forward computing model to evaluate the adaptability; step 4.3, updating the individual and global optimum; Step 4.4, iterating the particle position and the particle speed; And 4.5, terminating the condition and verifying.
  9. 9. The method of claim 8, wherein the step 4.1 is specifically defined by particles, each particle representing a set of potential airgun array schemes, and the position vector X=[V 1 , P 1 , x 1 , y 1 , z 1 ,…,V N , P N , x N , y N , z N ] comprises all parameters to be optimized of N airguns; initializing a population, namely randomly generating N initial particles in a preset physical boundary; Speed initialization, namely assigning an initial search speed V k to each particle, and determining the exploration step length of the particle in a parameter space; in each iteration, inputting all N array parameters in a particle swarm into a trained forward proxy model in batches; The agent model outputs far-field spectrum curves corresponding to the N arrays in millisecond time; Calculating a fitness value F (X) by calculating a score of each particle according to a predicted spectrum by using a formula; The step 4.3 is specifically that an individual extremum pbest is recorded at the optimal fitness position reached since searching each particle; global extremum gbest recording the "particles" currently found in the whole population that are closest to the target spectrum and most in line with engineering constraints; the step 4.4 is specifically to update the particle state according to the following classical formula to achieve convergence to the optimal solution: The physical significance is that the particles not only keep inertia, but also can be closed to the historical best scheme of the particles and the best scheme of the group; step 4.5 is specifically that when the fitness value corresponding to gbest tends to be stable and has no obvious improvement in a plurality of continuous iterations, the result is determined to be converged, and then the search is ended, and at this time, the obtained gbest is the finally determined inversion design scheme.
  10. 10. An air gun source array inversion optimization system combining a deep learning calculation model and a particle swarm algorithm, wherein the optimization system uses the air gun source array inversion optimization method combining the deep learning calculation model and the particle swarm algorithm according to any one of claims 1 to 9, and the optimization system comprises the following steps: The sample library construction module is used for constructing a high-fidelity physical sample library; the calculation model construction module is used for constructing a forward calculation model of the deep neural network and training; The particle swarm inversion frame construction module is used for constructing a multi-target constrained particle swarm inversion frame based on the forward calculation model constructed by the calculation model construction module; And the calculation module is used for performing particle swarm optimization on the frame constructed by the particle swarm inversion frame construction module to realize global optimization until the optimal air gun array arrangement combination parameters meeting the target spectrum requirements are obtained.

Description

Air gun seismic source array inversion optimization method combining deep learning calculation model and particle swarm optimization Technical Field The invention belongs to the technical field of marine geophysical exploration, and particularly relates to an air gun seismic source array inversion optimization method combining a deep learning calculation model and a particle swarm algorithm. Background As marine surveys enter the ultra-deep water age, increasingly high quality seismic wave data become key resources for marine surveys, which are critical for accurately detecting subsea geologic structures, evaluating hydrocarbon resources, and guiding drilling operations. The performance of the airgun array, which is used as a main artificial seismic wave source, directly affects the quality and accuracy of the survey data. The arrangement of the airgun array has important influence on the performance of the airgun array, such as the interval between airguns, the depth of sinking of the airguns, the excitation time of the airguns and the like. The spectral characteristics (e.g., corner frequency, bandwidth, etc.) of the seismic waves excited by the airgun array directly determine the resolution and penetration depth of the survey. In order to meet the exploration requirements of deep sea formations (e.g., requiring a corner frequency of up to 3Hz to detect deeper subsea structures), it is necessary to optimally design parameters such as volume, firing pressure, depth of deposition, air gun spacing, etc. of the air gun array, but it is a difficult matter how to obtain an ideal air gun array from existing air gun libraries. In the aspect of air gun array calculation and optimal arrangement, the prior art mainly has the problems that (1) the calculation efficiency is low, the traditional design depends on solving a complex nonlinear bubble dynamics equation (such as Gilmore equation), and when thousands of parameter iterative searches are carried out, the calculation time is huge, and the real-time requirement is difficult to meet. (2) Inversion difficulty and multiple solutions air gun array design is essentially a nonlinear inverse problem. For a given target pressure wave or spectrum (e.g., corner frequency up to 3 Hz), there are a variety of air gun combinations that can be satisfied. The traditional gradient descent inversion algorithm is easy to sink into local optimization, and solutions which are optimal in engineering (lowest energy consumption and smallest volume) cannot be effectively screened out. (3) And when the number of air guns is increased, the parameters to be optimized are increased in multiple, the search space is enlarged sharply, and the traditional experience trial-and-error method fails. Therefore, an array optimization design method capable of rapidly and accurately simulating bubble pressure wave and sound pressure level frequency spectrum, effectively solving the problem of multiple solutions and outputting engineering optimal solutions is needed. Disclosure of Invention The invention provides an air gun seismic source array inversion optimization method combining a deep learning calculation model and a particle swarm algorithm, which is used for solving the problems of slow calculation and strong inversion multi-solution of the existing air gun array optimization design. The invention is realized by the following technical scheme: An air gun seismic source array inversion optimization method combining a deep learning calculation model and a particle swarm algorithm, the optimization method comprising the following steps: Step1, constructing a high-fidelity physical sample library; Step 2, constructing a forward calculation model of the deep neural network and training; Step 3, constructing a multi-target constrained particle swarm inversion frame based on the forward direction calculation model in the step 2; and 4, realizing global optimization of the multi-target constrained particle swarm inversion frame in the step 3 based on a particle swarm algorithm until the optimal air gun array arrangement combination parameters meeting the target spectrum requirements are obtained. Furthermore, in the step 1, an input parameter set including the number, volume, pressure, depth of sinking and space position of different air guns is generated by using Latin hypercube sampling strategy, and a corresponding far-field pressure sweep spectrum is calculated as an output tag based on an air gun bubble dynamics model using a Zhang equation. Furthermore, in the step 2, the Deep neural network may be one or a combination of a graph neural network, a long-term and short-term memory network adopting a attention mechanism, a fully connected neural network or a Deep Sets network, and is trained by taking air gun array parameters as input and spectrum curves as output, and the model can establish a fast forward mapping relation from 'array parameters' to 'spectrum characteristics'. Further, in the step 3