CN111143984A - Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network - Google Patents
Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network Download PDFInfo
- Publication number
- CN111143984A CN111143984A CN201911336952.XA CN201911336952A CN111143984A CN 111143984 A CN111143984 A CN 111143984A CN 201911336952 A CN201911336952 A CN 201911336952A CN 111143984 A CN111143984 A CN 111143984A
- Authority
- CN
- China
- Prior art keywords
- network
- magnetotelluric
- dimensional
- model
- neural network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/12—Computing arrangements based on biological models using genetic models
- G06N3/126—Evolutionary algorithms, e.g. genetic algorithms or genetic programming
Abstract
The invention provides a magnetotelluric two-dimensional inversion method based on a genetic algorithm optimized neural network. The method comprises the following steps: firstly, establishing a BP neural network basic framework for a magnetotelluric two-dimensional geoelectric model to carry out learning training, inputting apparent resistivity parameters of a known geoelectric model by a network, and outputting the apparent resistivity parameters of the known geoelectric model; optimizing the neural network learning and training process by using a genetic algorithm, and calculating optimal solutions of various geoelectricity model network connection weights and thresholds; and finally, carrying out inversion test on the unknown model by using the optimal connection weight and the threshold, inputting apparent resistivity parameters of the unknown geoelectricity model by using a network, and outputting the apparent resistivity parameters of the unknown geoelectricity model. The invention can effectively improve the accuracy and speed of the magnetotelluric detection nonlinear inversion and has strong adaptability.
Description
The technical field is as follows:
the invention belongs to the field of geophysical exploration, and particularly relates to a magnetotelluric two-dimensional inversion method for optimizing a neural network based on a genetic algorithm.
Background art:
the magnetotelluric sounding method is widely applied to the fields of oil-gas exploration, mineral product general survey, exploration, geological engineering and the like, and the inversion of data is greatly developed. The classical inversion method has large dependence on an initial model and needs to calculate a sensitivity matrix. And the emergence of some new global optimization inversion methods provides a new approach for magnetotelluric inversion, and the common characteristics of the methods are that the dependence on an initial model is small, and a sensitivity matrix does not need to be calculated.
Since various inversion methods have various characteristics in principle, one method is not absolutely superior to the other method, and the hybrid optimization inversion can be widely adopted because the advantages of the methods can be combined to obtain a better inversion effect. For example, by combining an improved particle swarm optimization algorithm with a BP neural network, a nonlinear inversion of resistivity imaging is carried out; the teachers and the students combine the advantages of a multi-scale method and a genetic algorithm to establish a multi-scale successive approximation genetic algorithm; liu bin etc. combines the least square linear inversion and genetic algorithm nonlinear inversion based on inequality constraint, applies to three-dimensional resistivity detection, has improved the inversion effect. Practice shows that in geophysical data inversion, although certain effect is achieved by hybrid optimization inversion, the problems of low inversion speed, low accuracy and poor adaptability exist to a certain extent.
The invention content is as follows:
the invention provides a magnetotelluric two-dimensional inversion method based on a genetic algorithm optimized neural network, which aims to solve the problems of low inversion speed and accuracy and poor adaptability in the prior art.
In order to achieve the purpose of the invention, the technical scheme provided by the invention is as follows: a magnetotelluric two-dimensional inversion method based on a genetic algorithm optimized neural network comprises the following steps:
A. initializing and setting a neural network structure according to a magnetotelluric two-dimensional forward modeling model and apparent resistivity data, and optimizing network initial weight and threshold values through a genetic algorithm:
the network structure initialization setting means: using the number M of the acquired apparent resistivities as the number of nodes of an input layer, using the number S of grids in an inversion area of a two-dimensional model as the number of output nodes, and adopting double hidden layers, wherein the nodes are respectively H1、H2;
B. B, the network transmits in the forward direction, calculate the test error, and judge whether to meet the termination condition of iteration, if meet, get the optimal weight, threshold, otherwise the network transmits in the backward direction, revise weight, threshold, repeat step B;
C. and applying the trained neural network to magnetotelluric two-dimensional inversion.
The genetic algorithm optimization process in the step A comprises the following steps: selecting, inheriting and mutating, wherein the network prediction error in the process is used as an objective function and is expressed by the relative error between the network output value and the expected output value, and the expression is as follows:
wherein E (k) is the net prediction error of chromosome k, Ti(k) Expressed as expected output values, in the network learning training of magnetotelluric data, as the ground of a known modelTheoretical value of electrical parameter, Oi(k) And N is the total number of model parameters, and is the final output value of the network.
The fitness function represents the magnitude of the regeneration probability, and is given by:
where f (k) represents the fitness value for chromosome k, and E (k) is the net prediction error for chromosome k. And selecting the chromosome with the maximum fitness function value, decoding, and initializing a network weight and a threshold.
The iteration termination condition in the step B includes: iteration times and accuracy requirements.
The two-dimensional inversion for magnetotelluric described in the step C is to input apparent resistivity into a trained network to obtain corresponding geoelectric model parameters.
Compared with the prior art, the invention has the advantages that:
the invention fully combines the global optimization searching of the genetic algorithm and the local optimization searching of the neural network, improves the convergence success rate and the calculation speed of the understanding in the network learning training and can more accurately approach the real model in the inversion test compared with a single neural network inversion method. Meanwhile, compared with a common magnetotelluric data inversion method, the method does not need to calculate a sensitivity matrix, and is beneficial to improving the accuracy and speed of magnetotelluric nonlinear inversion. Network learning training and inversion testing are carried out through various two-dimensional geoelectric models, feasibility and effectiveness of the method on magnetotelluric nonlinear inversion are verified, and adaptability is high.
Description of the drawings:
FIG. 1 is a flow chart of the genetic algorithm optimized neural network of the present invention.
The specific implementation mode is as follows:
the present invention will be described in detail below with reference to the drawings and examples.
The invention introduces a genetic algorithm optimization neural network method into magnetotelluric two-dimensional nonlinear inversion. Firstly, establishing a BP neural network basic framework for a magnetotelluric two-dimensional geoelectric model to carry out learning training, inputting apparent resistivity parameters of a known geoelectric model by a network, and outputting the apparent resistivity parameters of the known geoelectric model; optimizing the neural network learning and training process by using a genetic algorithm, and calculating optimal solutions of various geoelectricity model network connection weights and thresholds; and finally, carrying out inversion test on the unknown model by using the optimal connection weight and the threshold, inputting apparent resistivity parameters of the unknown geoelectricity model by using a network, and outputting the apparent resistivity parameters of the unknown geoelectricity model. Referring to fig. 1, the whole process includes: initializing a network structure; determining a coding rule to form an initialization population; calculating a fitness function value, judging whether the fitness function value meets an iteration termination condition, and if not, operating a genetic algorithm; if yes, terminating iteration and selecting the best individual; initializing a network weight and a threshold; the network forward propagation is carried out, the network test error is calculated, whether the network test error meets the iteration termination condition is judged, if not, the network backward propagation is carried out, and the weight value and the threshold value are corrected; and obtaining an optimal weight threshold value for inversion of the unknown model.
Example (b): a magnetotelluric two-dimensional inversion method based on a genetic algorithm optimized neural network comprises the following steps:
a, initializing a neural network structure according to a magnetotelluric two-dimensional forward modeling model and apparent resistivity data, and optimizing a network initial weight and a threshold through a genetic algorithm:
the network structure initialization setting means: using the number M of the acquired apparent resistivities as the number of nodes of an input layer, using the number S of grids in an inversion area of a two-dimensional model as the number of output nodes, and adopting double hidden layers, wherein the nodes are respectively H1、H2(ii) a Specifically, in the magnetotelluric inversion process, the method aims at a two-dimensional underground medium, a finite element algorithm is adopted in the forward modeling process, the size of a grid is 30 x 31, the length of a measuring line is 16km, the distance between measuring points is 1km, 17 measuring points are totally adopted, 16 recording frequency points are adopted, the range is 0.016-512 Hz and are distributed at equal intervals according to logarithms, 544 visual resistivities are obtained by all the measuring points in the acquisition frequency range under a TE (transverse electric) and TM (transverse electric) mode, and the resistivity value of an epitaxial region of the grid is a true value and does not participate in inversion because the resistivity of a constant air layer and the real value is obtained, the final inversion grid is2. And determining the number of nodes of the input layer, the output layer and the double hidden layers as 544, 352, 12 and 8 respectively by using the apparent resistivity in all the acquisition frequency ranges as an input parameter and the electrical parameter of the inversion area grid, namely the resistivity value as an output and adopting the double hidden layers.
When the population size is set to 30 and the weight and threshold of the network are gene-coded, the gene coding length Num of each chromosome is (M × H)1+H1×H2+H2×S+H1+H2+S)×L=98120,M、S、H1And H2The number of input nodes is 544, the number of output nodes is 352, the number of nodes of a first hidden layer is 12, the number of nodes of a second hidden layer is 8, L is the number of coding bits of a variable, the value is 10, each chromosome comprises all weight values and threshold value information of a network, the value of each position of the chromosome is determined by a random function, and an initial population with a set scale is generated. The iteration precision is 0.01, and the maximum genetic algebra is 20 generations. The relative error between the network output value and the expected output value is used for representing the network prediction error, and the formula is shown as the formula (1):
in the formula (1), E (k) is the net prediction error of chromosome k, Ti(k) Expressed as the expected output value, in the network learning training of magnetotelluric data, is the theoretical value of the geoelectric parameter, O, of the known modeli(k) And N is the total number of model parameters, and is the final output value of the network. The smaller the value of E (k), the closer the inversion result is to the actual model.
The fitness function represents the magnitude of the regeneration probability, and the fitness function adopted by combining the characteristics of the method is shown as the formula (2):
where f (k) represents the fitness value of chromosome k, and E (k) is the net prediction error of chromosome k, which converts minimization of the objective function into maximization of the fitness function.
Selecting operators, randomly traversing and sampling, randomly matching individuals in a new population into pairs, and performing cross operation on each pair of individuals according to a cross probability of 0.7 to generate two new individuals, wherein the genetic gully is 0.95; and carrying out mutation evolution on each individual in the new population according to the mutation probability of 0.01 to generate new individuals. Through a plurality of genetic iteration processes, when the target convergence precision is met, the chromosome with the maximum fitness function value is selected and decoded, and the corresponding weight and threshold are obtained according to the arrangement sequence and the gene length of each parameter and are assigned to the neural network.
And B: and B, the network forwards propagates, a test error is calculated, whether an iteration termination condition is met is judged, if yes, the optimal weight and threshold are obtained, otherwise, the network backwards propagates, the weight and the threshold are corrected, and the step B is repeated:
in the neural network training process, the maximum iteration number is 1000, the mean square error of the inversion result and the model parameters is used as the fitting difference of a training target, the value is 0.01, and the formula is as shown in formula (3)
Wherein λTIs a normalized geoelectrical parameter theoretical value, lambda, of the test modelCCalculating the inversion parameter of the normalized model, wherein N is the total number of parameters, correcting the weight and the threshold value by a gradient descent method,
wherein η is learning rate, and its value is 0.1. Δ ω and Δ θ are weight and threshold of correction, and ω istIs thetatAs weight and threshold, ω, for the t-th iterationt+1And thetat+1The weight and threshold for the t +1 th iteration. Forward transmission over multiple networksAnd (5) broadcasting and back-propagating until the iterative fitting difference is less than 0.01, and obtaining the optimal weight and threshold.
And step C, inputting the apparent resistivity data of the unknown model into the trained neural network for inversion to obtain corresponding geoelectrical parameters.
Claims (4)
1. A magnetotelluric two-dimensional inversion method based on a genetic algorithm optimized neural network comprises the following steps:
A. initializing and setting a neural network structure according to a magnetotelluric two-dimensional forward modeling model and apparent resistivity data, and optimizing network initial weight and threshold values through a genetic algorithm:
the network structure initialization setting means: using the number M of the acquired apparent resistivities as the number of nodes of an input layer, using the number S of grids in an inversion area of a two-dimensional model as the number of output nodes, and adopting double hidden layers, wherein the nodes are respectively H1、H2;
B. B, the network transmits in the forward direction, calculate the test error, and judge whether to meet the termination condition of iteration, if meet, get the optimal weight, threshold, otherwise the network transmits in the backward direction, revise weight, threshold, repeat step B;
C. and applying the trained neural network to magnetotelluric two-dimensional inversion.
2. The two-dimensional magnetotelluric inversion method based on genetic algorithm optimized neural network as claimed in claim 1, wherein the genetic algorithm optimization process in step a comprises: selecting, inheriting and mutating, wherein the network prediction error in the process is used as an objective function and is expressed by the relative error between the network output value and the expected output value, and the expression is as follows:
wherein E (k) is the net prediction error of chromosome k, Ti(k) Expressed as expected output values, in the network learning training of magnetotelluric data, as geoelectric parameters of a known modelTheoretical value, Oi(k) And N is the total number of model parameters, and is the final output value of the network.
The fitness function represents the magnitude of the regeneration probability, and is given by:
where f (k) represents the fitness value for chromosome k, and E (k) is the net prediction error for chromosome k. And selecting the chromosome with the maximum fitness function value, decoding, and initializing a network weight and a threshold.
3. The two-dimensional magnetotelluric inversion method based on genetic algorithm optimized neural network as claimed in claim 1 or 2, wherein the iteration termination condition of step B comprises: iteration times and accuracy requirements.
4. The two-dimensional magnetotelluric inversion method based on genetic algorithm optimized neural network as claimed in claim 3, wherein the step C for two-dimensional magnetotelluric inversion is to input apparent resistivity into a trained network to obtain corresponding parameters of a geoelectric model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911336952.XA CN111143984A (en) | 2019-12-23 | 2019-12-23 | Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201911336952.XA CN111143984A (en) | 2019-12-23 | 2019-12-23 | Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN111143984A true CN111143984A (en) | 2020-05-12 |
Family
ID=70519412
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201911336952.XA Pending CN111143984A (en) | 2019-12-23 | 2019-12-23 | Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111143984A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111812732A (en) * | 2020-06-29 | 2020-10-23 | 中铁二院工程集团有限责任公司 | Magnetotelluric nonlinear inversion method based on convolutional neural network |
CN113158571A (en) * | 2021-04-26 | 2021-07-23 | 中国科学院地质与地球物理研究所 | Magnetotelluric inversion method based on full convolution neural network |
CN113204054A (en) * | 2021-04-12 | 2021-08-03 | 湖南工商大学 | Self-adaptive wide-area electromagnetic method induced polarization information extraction method based on reinforcement learning |
CN113486591A (en) * | 2021-07-13 | 2021-10-08 | 吉林大学 | Gravity multi-parameter data density weighted inversion method for convolutional neural network result |
CN115099144A (en) * | 2022-06-24 | 2022-09-23 | 无锡物联网创新中心有限公司 | Yarn raw material characteristic parameter inversion method and related device |
US11782183B2 (en) | 2021-04-26 | 2023-10-10 | Institute Of Geology And Geophysics, Chinese Academy Of Sciences | Magnetotelluric inversion method based on fully convolutional neural network |
-
2019
- 2019-12-23 CN CN201911336952.XA patent/CN111143984A/en active Pending
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111812732A (en) * | 2020-06-29 | 2020-10-23 | 中铁二院工程集团有限责任公司 | Magnetotelluric nonlinear inversion method based on convolutional neural network |
CN111812732B (en) * | 2020-06-29 | 2024-03-15 | 中铁二院工程集团有限责任公司 | Geoelectromagnetic nonlinear inversion method based on convolutional neural network |
CN113204054A (en) * | 2021-04-12 | 2021-08-03 | 湖南工商大学 | Self-adaptive wide-area electromagnetic method induced polarization information extraction method based on reinforcement learning |
CN113158571A (en) * | 2021-04-26 | 2021-07-23 | 中国科学院地质与地球物理研究所 | Magnetotelluric inversion method based on full convolution neural network |
CN113158571B (en) * | 2021-04-26 | 2022-04-15 | 中国科学院地质与地球物理研究所 | Magnetotelluric inversion method based on full convolution neural network |
US11782183B2 (en) | 2021-04-26 | 2023-10-10 | Institute Of Geology And Geophysics, Chinese Academy Of Sciences | Magnetotelluric inversion method based on fully convolutional neural network |
CN113486591A (en) * | 2021-07-13 | 2021-10-08 | 吉林大学 | Gravity multi-parameter data density weighted inversion method for convolutional neural network result |
CN113486591B (en) * | 2021-07-13 | 2022-04-19 | 吉林大学 | Gravity multi-parameter data density weighted inversion method for convolutional neural network result |
CN115099144A (en) * | 2022-06-24 | 2022-09-23 | 无锡物联网创新中心有限公司 | Yarn raw material characteristic parameter inversion method and related device |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111143984A (en) | Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network | |
CN110363344B (en) | Probability integral parameter prediction method for optimizing BP neural network based on MIV-GP algorithm | |
CN111126591B (en) | Magnetotelluric deep neural network inversion method based on space constraint technology | |
CN110968826B (en) | Magnetotelluric deep neural network inversion method based on spatial mapping technology | |
CN109993280B (en) | Underwater sound source positioning method based on deep learning | |
CN103077267B (en) | Parameter sound source modeling method based on improved BP (Back Propagation) neural network | |
CN105203565B (en) | A kind of power spectrum Overlapped peak resolution method | |
CN112949134B (en) | Earth-well transient electromagnetic inversion method based on non-structural finite element method | |
CN114595732B (en) | Radar radiation source sorting method based on depth clustering | |
CN112699596B (en) | Wide-area electromagnetic method induced polarization information nonlinear extraction method based on learning | |
CN111058840A (en) | Organic carbon content (TOC) evaluation method based on high-order neural network | |
CN115841076A (en) | Shallow sea layered seabed ground sound parameter inversion method based on BP neural network model | |
CN104422969A (en) | Method for reducing non-uniqueness of electromagnetic sounding inversion result | |
Dai et al. | Nonlinear inversion for electrical resistivity tomography based on chaotic DE-BP algorithm | |
CN115598714B (en) | Time-space coupling neural network-based ground penetrating radar electromagnetic wave impedance inversion method | |
Ilamah | A multiobjective dominance and decomposition algorithm for reservoir model history matching | |
CN116304565A (en) | Wavelet packet denoising and BP neural network-based transient electromagnetic inversion method | |
KR20200028801A (en) | Learning method and learning device for variational interference using neural network and test method and test device for variational interference using the same | |
CN103778327A (en) | Curve fitting method | |
CN115795994A (en) | Orientation electromagnetic wave logging while drilling data inversion method based on Unet convolution neural network | |
CN113640712A (en) | Prediction method for vertical component of vertical induced magnetic field of ship | |
CN117270072B (en) | Gravity magnetic potential field imaging inversion method and system based on improved differential evolution algorithm | |
Baun et al. | Hybrid stochastic genetic evolution-based prediction model of received input voltage for underground imaging applications | |
Xie et al. | 2D magnetotelluric inversion based on ResNet | |
Bao et al. | A Unified Filter Method for Jointly Estimating State and Parameters of Stochastic Dynamical Systems via the Ensemble Score Filter |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20200512 |
|
WD01 | Invention patent application deemed withdrawn after publication |