CN111812732B - Geoelectromagnetic nonlinear inversion method based on convolutional neural network - Google Patents
Geoelectromagnetic nonlinear inversion method based on convolutional neural network Download PDFInfo
- Publication number
- CN111812732B CN111812732B CN202010605022.6A CN202010605022A CN111812732B CN 111812732 B CN111812732 B CN 111812732B CN 202010605022 A CN202010605022 A CN 202010605022A CN 111812732 B CN111812732 B CN 111812732B
- Authority
- CN
- China
- Prior art keywords
- layer
- data
- pooling
- convolution
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 56
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 41
- 238000011176 pooling Methods 0.000 claims abstract description 51
- 238000000605 extraction Methods 0.000 claims abstract description 9
- 238000013507 mapping Methods 0.000 claims abstract description 8
- 238000012549 training Methods 0.000 claims description 27
- 230000006870 function Effects 0.000 claims description 24
- 230000004913 activation Effects 0.000 claims description 19
- 238000004364 calculation method Methods 0.000 claims description 18
- 238000012360 testing method Methods 0.000 claims description 11
- 230000002159 abnormal effect Effects 0.000 claims description 6
- 230000005540 biological transmission Effects 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 4
- 230000003213 activating effect Effects 0.000 claims description 3
- 239000011159 matrix material Substances 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 230000009467 reduction Effects 0.000 claims description 3
- 239000011435 rock Substances 0.000 claims description 3
- 230000009466 transformation Effects 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 238000011423 initialization method Methods 0.000 claims description 2
- 230000000644 propagated effect Effects 0.000 claims description 2
- 238000012546 transfer Methods 0.000 claims 1
- 238000007781 pre-processing Methods 0.000 abstract description 3
- 238000013528 artificial neural network Methods 0.000 description 8
- 230000008569 process Effects 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 230000009977 dual effect Effects 0.000 description 3
- 230000002068 genetic effect Effects 0.000 description 3
- 238000003384 imaging method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 230000003287 optical effect Effects 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 201000010099 disease Diseases 0.000 description 1
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 description 1
- 239000006185 dispersion Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000005672 electromagnetic field Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 229910052751 metal Inorganic materials 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 150000002739 metals Chemical class 0.000 description 1
- 239000002480 mineral oil Substances 0.000 description 1
- 235000010446 mineral oil Nutrition 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01V—GEOPHYSICS; GRAVITATIONAL MEASUREMENTS; DETECTING MASSES OR OBJECTS; TAGS
- G01V3/00—Electric or magnetic prospecting or detecting; Measuring magnetic field characteristics of the earth, e.g. declination, deviation
- G01V3/38—Processing data, e.g. for analysis, for interpretation, for correction
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/30—Assessment of water resources
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Geology (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Geophysics (AREA)
- Remote Sensing (AREA)
- Environmental & Geological Engineering (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Complex Calculations (AREA)
Abstract
The invention discloses a geomagnetic nonlinear inversion method based on a convolutional neural network, which applies the convolutional neural network to geomagnetic inversion in the geophysical field, takes a certain acquired observation data as input of the convolutional neural network, takes a geoelectric model parameter as output of the network, adopts a deep network model, can realize more accurate mapping on complex nonlinear data, avoids complex pre-processing on original data, has a weight sharing network structure, and can greatly reduce the complexity of the network model. In addition, the original input data can realize the extraction of different characteristics of the data through convolution so as to mine out the information of the deeper level of the data; after convolution, the pooling can reduce the resolution of the feature surface to obtain space invariance features, reduce the scale of the network and play a role in secondary feature extraction.
Description
Technical Field
The invention relates to a magnetotelluric sounding method, in particular to a magnetotelluric nonlinear inversion method based on a convolutional neural network.
Background
The geomagnetic sounding is a geophysical exploration method for researching an electrical structure of the earth by utilizing a natural alternating electromagnetic field, has low cost, convenient operation, larger exploration depth, high resolution to a low-resistance layer and high shielding resistance, and is widely applied to the fields of mineral oil and gas exploration, geothermal resource exploration, environmental engineering and the like. Inversion of data is an important technology in the process of processing and explaining the magnetotelluric sounding data, and comprises a linear inversion method and a nonlinear inversion method:
feng Deshan and Wang in China nonferrous metals journal 2013,23 (9): 2524-2531 published magnetotelluric biquadratic interpolation FEM forward and least square regularization joint inversion [ J ] a magnetotelluric linear inversion method based on least square is introduced, a regularization method for solving a disease state problem is applied to a least square optimization method by using an inversion theory, a smoothest constraint least square regularization inversion objective function is obtained, a least square inversion program based on smooth constraint is compiled, and various ground electric models are subjected to trial calculation by using the program, so that the effectiveness and feasibility of the inversion algorithm are verified, but the inversion accuracy is lower and the calculation speed is slower.
Wang He, liu Menglin, xi Zhenzhu, et al in geophysical journal, 2018,61 (04): 1563-1575. A genetic neural network-based magnetotelluric inversion [ J ] is disclosed, which is a magnetotelluric nonlinear inversion method based on a genetic neural network, wherein parameters of an input layer, an implicit layer and an output layer of the network are reasonably designed to construct a magnetotelluric network structure, a genetic algorithm is utilized to optimize weights and thresholds of the neural network, then a plurality of two-dimensional ground electric models are subjected to network learning training to obtain optimal solutions for connecting the weights and the thresholds, finally the optimal connecting weights and the thresholds are subjected to inversion test on unknown ground electric models, and the result shows that the method can realize accurate positioning and imaging on different ground electric models, but the success rate of convergence and the calculation speed are still improved.
Disclosure of Invention
The invention aims at: aiming at the problems of lower inversion precision and lower calculation speed in the prior art, the magnetotelluric nonlinear inversion method based on the convolutional neural network is provided.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a geoelectromagnetic nonlinear inversion method based on a convolutional neural network,
s100, constructing a convolutional neural network, wherein the convolutional neural network comprises an input layer, a double convolutional layer, a double pooling layer, a full connection layer and an output layer;
s200, acquiring a sample data set through a magnetotelluric two-dimensional forward model, wherein the sample data comprises input and output;
s300, inputting the sample data set into a convolutional neural network for training, comprising:
s310, inputting sample data into convolution;
s320, inputting the convolved sample data into a pooling device;
s330, the input of the sample data is convolved and pooled and then enters a full-connection layer, a mapping relation between the extracted features and an output layer is established in the full-connection layer, and the obtained output and the output of the sample data are subjected to difference, namely, the error between the output value and the theoretical value;
repeating the forward information transmission and the reverse error transmission until the error reaches a desired value or the iteration number reaches a preset value, and stopping training by the network at the moment and obtaining an optimal weight and a threshold value;
s400, inputting test set data into a network, and outputting the test set data after convolution, pooling and one forward propagation.
A magnetotelluric nonlinear inversion method based on a convolutional neural network applies the convolutional neural network to magnetotelluric inversion in the geophysical field, takes one acquired observation data as input of the convolutional neural network, takes a geoelectric model parameter as output of the network, adopts a deep network model, can realize more accurate mapping on complex nonlinear data, avoids complex pre-processing of original data, has a weight sharing network structure, and can greatly reduce the complexity of the network model.
Preferably, in the step S100, the convolution layers C1 and C2 and the pooling layers S1 and S2 are alternately arranged, that is, one convolution layer is connected to one pooling layer, and one convolution layer is connected after the pooling layer;
adopting an Xavier initialization method to carry out initial assignment on the weight and bias of the network, and initializing the parameter W of the weight [l] The mean μ=0, varianceNormal distribution of (c):
wherein l is the first layer of the network, n [l-1] Is the number of neurons in layer 1.
Preferably, the step S200 includes:
performing double interpolation on the grid by adopting a finite element method, and dividing the whole grid into two areas: the system comprises a target area and a grid extension area, wherein the target area is an occurrence area of a geologic body and is also a data acquisition area, the uniform grid is divided, the grid size NX multiplied by NY, and the grid step length of the grid extension area is increased by two times;
in a target area, fixing surrounding rock and abnormal volume resistivity values are unchanged, and moving a geologic body according to a certain step length to obtain a sample data set, wherein the sample data set comprises: containing N f Frequency point, N s Apparent resistivity data for individual stationsContaining N f Frequency point, N s Impedance phase data of individual measuring points->And corresponding earth model parameters +.>
The sample dataset is normalized using a zero-mean normalization method:
wherein x is * Represents normalized data, the mean value of which is 0, and the standard deviation of which is 1; mu and sigma represent the mean and standard deviation of the raw data, and the apparent resistivity after normalizationPhase data->As a two-channel input of the convolutional neural network, normalized corresponding electrical model parameters +.>As an output.
Preferably, the step S310 includes:
the normalized apparent resistivity, which is the input data in step S200Phase data->Performing convolution processing, namely performing convolution transformation mapping on the input data of each position to a new value;
each characteristic output graph can combine a plurality of characteristic convolution graphs, and the convolution calculation formula is as follows
Wherein the method comprises the steps ofIndicating the net activation of the j-th channel of the convolutional layer, by a characteristic map of the output of the previous layerThe +.A convolution sum and bias operation is performed to obtain +.>The output of the jth channel of the convolution layer l, f (·) represents the activation function, M j Representing a subset of feature maps, < >>Representing a convolution kernel matrix,/->Represents the bias term of the convolutional layer, "×" represents the convolutional symbol.
Preferably, the step S320 includes:
carrying out pooling treatment on the convolved data by adopting a maximum pooling method, namely reserving the maximum value of the data in a pooling window area, and activating by an activation function to obtain a result, wherein the step can be used for carrying out dimension reduction and secondary feature extraction on a data body; the pooling calculation formula is:
wherein,net activation of the jth channel of the pooling layer l is indicated by the feature map of the previous layer +.>Pooling, weighting and biasing to obtain the product; beta represents the pooling weight coefficient, +.>The bias term representing the pooling layer, pool (·) represents the pooling function.
Preferably, the information in the step S330 is propagated forward:
assume that the input of CNN is a i,l-1 The forward propagation algorithm calculation includes:
currently the full connection layer, there is a i,l ＝σ(z i,l )＝σ(W l a i,l-1 +b l ) Wherein a is i,l For forward propagation output, W l B is a network weight l As a network threshold, σ (·) is an activation function;
currently the convolutional layer, then there is a i,l ＝σ(z i,l )＝σ(W l *a i,l-1 +b l ) Wherein is a convolution operation;
currently the pooling layer, there is a i,l ＝pool(a i,l-1 ) Wherein pool (·) is a pooling operation.
For output layer L: a, a i,L ＝ReLU(z i,L )＝ReLU(W L a i,L-1 +b L ) Wherein ReLU is the activation function.
Finally, the total gradient error delta of the output layer is calculated through the loss function i,L 。
Preferably, the error in the step S330 is back-propagated:
s331 assumes that the gradient error of the previous layer is delta i,l+1 The back propagation algorithm calculation includes:
currently the full connection layer, then there is delta i,l ＝(W l+1 )Tδ i,l+1 ⊙σ'(z i,l ) Wherein delta i,l Represents the gradient error of the first layer, +.;
currently the convolutional layer, then there is delta i,l ＝δ i,l+1 *(W l+1 )T⊙σ'(z i,l )；
Currently the pooling layer, then there is delta i,l ＝upsample(δ i,l+1 )⊙σ'(z i,l ) Wherein upsample (·) is an upsampling function;
s332CNN updates W of the first layer l And b l Comprising:
currently is the fully connected layer:wherein alpha is iteration step length, and m is training sample number;
currently is a convolutional layer, with for each convolutional kernel:
preferably, before the test set data is input into the network in step S400, the following settings are performed: setting the proportion of training set, test set and verification set as 98:1:1, the number of network iterations is 200.
An electronic device comprising at least one processor, and a memory communicatively coupled to the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of the preceding claims.
In summary, due to the adoption of the technical scheme, the beneficial effects of the invention are as follows:
according to the geomagnetic nonlinear inversion method based on the convolutional neural network, the convolutional neural network is applied to geomagnetic inversion in the geophysical field, one acquired observation data is used as input of the convolutional neural network, the earth model parameters are used as output of the network, a deep network model is adopted, more accurate mapping of complex nonlinear data can be achieved, complex pre-processing of original data is avoided, a weight sharing network structure is provided, and complexity of the network model can be greatly reduced. In addition, the original input data can realize the extraction of different characteristics of the data through convolution so as to mine out the information of the deeper level of the data; after convolution, the pooling can reduce the resolution of the feature surface to obtain space invariance features, reduce the scale of the network and play a role in secondary feature extraction.
Drawings
Fig. 1 is a flow chart diagram of the present invention.
FIG. 2 is a schematic diagram of a convolutional neural network for magnetotelluric inversion.
FIG. 3 is an image of a convolutional neural network inversion high resistance model and a dual high resistance combined model.
FIG. 4 is an imaging result of the least square inversion high resistance model and the dual high resistance combined model.
FIG. 5 is a graph comparing observation data with inversion result model response data in a TM mode of a high-resistance model.
FIG. 6 is a graph comparing observation data with inversion result model response data in a dual high resistance combined model TM mode.
Fig. 7 is a schematic structural diagram of an electronic device provided by the present invention.
Detailed Description
The present invention will be described in detail with reference to the accompanying drawings.
The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
The implementation flow of the method of the invention is shown in the flow chart of the magnetotelluric nonlinear inversion method based on the convolutional neural network shown in fig. 1, and is specifically as follows:
s100: and building and initializing a convolutional neural network. The convolutional neural network basic framework is built according to the problem to be solved, and comprises an input layer I, double convolutional layers (C1 and C2), double pooling layers (S1 and S2), a full connection layer FA and an output layer O, as shown in figure 2. The convolution layers (C1, C2) and the pooling layers (S1, S2) are alternately arranged, namely one convolution layer is connected with one pooling layer, and one convolution layer is connected after the pooling layer; weights to networks using Xavier initialization methodsInitial assignment is carried out on the value and the bias, and the parameter W of the initial weight value [l] (l is the network layer I) meets the formula:b [l] =0. Wherein n is [l-1] Is the number of neurons in layer 1-1, that is to say the initial weights fit the mean μ=0, variance +.>Is a normal distribution of (c).
S200: sample dataset construction and input. The method comprises the steps of obtaining a sample data set through a geoelectromagnetic two-dimensional forward modeling, performing double interpolation on a grid by adopting a finite element method in order to improve the calculation efficiency and the accuracy, and dividing the whole grid into two areas: the target area is an occurrence area of the geologic body and is also a data acquisition area, the target area and the grid extension area are divided into uniform grids, the grid size NX multiplied by NY, and the grid step length of the grid extension area is increased by two times. In the target area, the resistivity values of the fixed surrounding rock and the abnormal body are unchanged, and the geologic body moves according to a certain step length to obtain a sample data set: containing N f Frequency point, N s Apparent resistivity data for individual stationsContaining N f Frequency point, N s Impedance phase data of individual measuring points->And corresponding earth model parameters +.>Normalization of the data set using zero-mean normalization method,/->Wherein x is * Represents normalized data, the mean value of which is 0, and the standard deviation of which is 1; mu and sigma represent the mean and standard deviation of the raw data, which willNormalized apparent resistivity +.>Phase data->As a two-channel input of the convolutional neural network, normalized corresponding electrical model parameters +.>As an output.
S300, inputting the sample data set into a convolutional neural network for training, comprising:
s310 convolution: normalized apparent resistivity, which is the input data in step S200Phase dataAnd performing convolution processing, namely performing convolution transformation on the input data of each position and mapping the input data into a new value. Each characteristic output graph can be combined with a plurality of characteristic convolution graphs, and the convolution calculation formula is +.> Wherein: />The net activation of the j-th channel of the convolution layer l is indicated by the characteristic map of the output of the previous layer +.>The +.A convolution sum and bias operation is performed to obtain +.>The output of the jth channel of the convolution layer l, f (·) represents the activation function, M j Representing a subset of feature maps, < >>Representing a convolution kernel matrix,/->Represents the bias term of the convolutional layer, "×" represents the convolutional symbol. The method comprises the steps that extraction of different characteristics of data can be achieved by selecting different convolution kernels, wherein the size of a convolution kernel of a first convolution layer is 5 multiplied by 5, the number of the convolution kernels is 32, and the convolution step length is 2 multiplied by 2; the second convolution layer has a size of 5×5, the number of convolution kernels is 64, and the convolution step size is 2×2.
S320 pooling: and (3) carrying out pooling treatment on the convolved data in the step (S310) by adopting a maximum pooling method, namely reserving the maximum value of the data in a pooled window area, and activating by an activation function to obtain a result, wherein the step can be used for carrying out dimension reduction and secondary feature extraction on the data body. The pooling calculation formula is Wherein: />Net activation of the jth channel of the pooling layer l is indicated by the feature map of the previous layer +.>Pooling, weighting and biasing to obtain the product; beta represents the pooling weight coefficient, +.>The bias term representing the pooling layer, pool (·) represents the pooling function. Depending on the problem to be solved, the data may be convolved again, i.e. the process returns to step S310.
In the learning training process, the S330 CNN performs supervised training through a BP algorithm, and two stages of forward information propagation and error reverse transmission are needed.
In the error forward propagation stage, assume that the input of CNN is a i,l-1 First, CNN performs forward propagation algorithm calculation according to the following 3 cases: a) If it is currently the fully connected layer: then there is a i,l ＝σ(z i,l )＝σ(W l a i,l-1 +b l ) Wherein: a, a i,l For forward propagation output, W l B is a network weight l As a network threshold, σ (·) is an activation function; b) If the current is a convolutional layer: then there is a i,l ＝σ(z i,l )＝σ(W l *a i,l-1 +b l ) Wherein: * Is convolution operation; c) If it is currently the pooling layer: then there is a i,l ＝pool(a i,l-1 ) Wherein: pool (·) is a pooling operation. For output layer L: ai ,L ＝ReLU(z i,L )＝ReLU(W L a i，L-1 +b L ) Wherein: reLU is an activation function. Finally, the total gradient error delta of the output layer is calculated through the loss function i，L 。
In the error back propagation stage, the gradient error of the previous layer is assumed to be delta i，l+1 First, the CNN performs the back-propagation algorithm calculation according to the following 3 cases, a) if it is currently a fully connected layer: then there is delta i，l ＝(W l+1 )Tδ i,l+1 ⊙σ'(z i,l ) Wherein: delta i,l Represents the gradient error of the first layer, +.; b) If the current is a convolutional layer: then there is delta i,l ＝δ i,l+1 *(W l+1 )T⊙σ'(z i,l ) The method comprises the steps of carrying out a first treatment on the surface of the c) If it is currently the pooling layer: then there is delta i,l ＝upsample(δ i,l+1 )⊙σ'(z i,l ) Wherein: upsampling (·) is an upsampling function. Thereafter, the CNN updates the W of the first layer according to the following two cases l And b l A) if it is currently a fully connected layer: wherein: alpha is iteration step length, m is training sample number; b) If a convolutional layer is present, for each convolution kernel there is: /> The CNN realizes network training through continuous forward information propagation and error reverse transmission, and when the error reaches an expected value or the iteration number reaches a preset value, the network stops training to obtain an optimal weight and a threshold value.
S400 model inversion: setting the sample data ratio of the training set test set and the verification set to be 98:1:1, the number of network iterations is 200. The optimal weight and the threshold value obtained by learning training are used as the initial weight and the threshold value of the network, unknown test set data are standardized and then directly input into the network, and output can be obtained through forward propagation once.
As shown in fig. 7, an electronic device (e.g., a computer server with program execution function) according to an exemplary embodiment of the present invention includes at least one processor, a power supply, and a memory and an input-output interface communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method disclosed in any one of the preceding embodiments; the input/output interface can comprise a display, a keyboard, a mouse and a USB interface, and is used for inputting and outputting data; the power supply is used for providing power for the electronic device.
Those skilled in the art will appreciate that: all or part of the steps for implementing the above method embodiments may be implemented by hardware related to program instructions, and the foregoing program may be stored in a computer readable storage medium, where the program, when executed, performs steps including the above method embodiments; and the aforementioned storage medium includes: a mobile storage device, a Read Only Memory (ROM), a magnetic disk or an optical disk, or the like, which can store program codes.
The above-described integrated units of the invention, when implemented in the form of software functional units and sold or used as stand-alone products, may also be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the embodiments of the present invention may be embodied in essence or a part contributing to the prior art in the form of a software product stored in a storage medium, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a removable storage device, a ROM, a magnetic disk, or an optical disk.
Example 1
The theoretical high-resistance abnormal body resistivity is 1000 omega-m, and the background resistivity value is 100 omega-m; the theoretical double-high-resistance combined model has the abnormal resistivity of 1000 Ω & m, the background resistivity value of 100 Ω & m, and the inversion results of the convolutional neural network under the TM mode of the two models shown in FIG. 3, wherein FIG. 3 (a) is the inversion result of the high-resistance model, and FIG. 3 (b) is the inversion result of the double-high-resistance combined model, and it can be seen that: the position, the size and the resistivity value of the abnormal body can be accurately reflected. Fig. 4 shows the inversion result of the least square method, where fig. 4 (a) shows the inversion result of the high-resistance model, and fig. 4 (b) shows the inversion result of the dual-high-resistance combined model, and it can be seen by comparison: the convergence degree of the least square inversion result is poor, the edge reflection of the model is not clear enough, and the magnetotelluric nonlinear inversion method based on the convolutional neural network can clearly describe the edge information of the model. The prediction data calculated by model inversion and the actual observation data are compared and analyzed, as shown in fig. 5 and 6, wherein fig. 5 (a) and 6 (a) are observation apparent resistivity, fig. 5 (b) and 6 (b) are prediction apparent resistivity, fig. 5 (c) and 6 (c) are observation impedance phase, and fig. 5 (d) and 6 (d) are prediction impedance phase, and both the observation apparent resistivity data and the impedance phase data are basically consistent with the actual observation data, so that the accuracy of the inversion method is illustrated.
The computer processor used to perform the inversion method is Intel (R) Core (TM) i5-8265U CPU@1.60GHz 1.80GHz, the programming software is python3.6.0, the training set data size 1160, the training round number 200 is set, the training batch size is 150, the training time of the convolutional neural network is 57s, the training time of the DNN is 168s, the training time of the ANN is 561s, and the calculation speed of the method is about 3 times of DNN and 10 times of ANN.
Relative to the prior art, it can be seen that: the method is simple in implementation process, network errors are continuously adjusted through network learning training, training can be stopped once accuracy requirements are met, and the trained network is directly used for magnetotelluric nonlinear inversion, so that inversion has instantaneity; compared with the traditional Artificial Neural Network (ANN) and the Deep Neural Network (DNN), the network learning convergence speed of the method is obviously faster; compared with classical least squares inversion, the method has better imaging effect and inversion efficiency. In the embodiment, an Adam optimizer is used in the data training process, so that the convergence speed is increased, and the problems of overfitting, gradient dispersion and the like of training are prevented by using a ReLU activation function and a node discarding technology.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.
Claims (7)
1. The magnetotelluric nonlinear inversion method based on the convolutional neural network is characterized by comprising the following steps of:
s100, constructing a convolutional neural network, wherein the convolutional neural network comprises an input layer, a double convolutional layer, a double pooling layer, a full connection layer and an output layer; wherein the convolution layer and the pooling layer are alternately arranged;
adopting an Xavier initialization method to carry out initial assignment on the weight and bias of the network, and initializing the parameter W of the weight [l] The mean μ=0, varianceNormal distribution of (c):
b [l] ＝0；
wherein l is the number of network layers, n [l-1] Is the number of neurons of layer 1, b [l] An initial value for a network threshold;
s200, acquiring a sample data set through a geoelectromagnetic two-dimensional forward model, wherein the sample data comprises apparent resistivity, phase data and ground model parameters;
specifically, a finite element method is adopted to perform double interpolation on the grid, and the whole grid is divided into two areas: the system comprises a target area and a grid extension area, wherein the target area is an occurrence area of a geologic body and is also a data acquisition area, the uniform grid is divided, the grid size NX multiplied by NY, and the grid step length of the grid extension area is increased by two times;
in a target area, fixing surrounding rock and abnormal volume resistivity values are unchanged, and moving a geologic body according to a certain step length to obtain a sample data set, wherein the sample data set comprises: containing N f Frequency point, N s Apparent resistivity data for individual stationsContaining N f Frequency point, N s Impedance phase data of individual measuring points->And corresponding earth model parameters +.>
The sample dataset is normalized using a zero-mean normalization method:
wherein x is * Representing normalized data, wherein the mean value is 0 and the variance is 1; mu and sigma represent the mean and variance of the raw data, and the normalized apparent resistivityPhase data->As a two-channel input of the convolutional neural network, normalized corresponding electrical model parameters +.>As an output;
s300, inputting normalized apparent resistivity, phase data and ground model parameters into a convolutional neural network for training, wherein the method comprises the following steps:
s310, performing primary feature extraction by performing convolution on the normalized apparent resistivity and phase data;
s320, the apparent resistivity and phase data after convolution enter a pool for data volume dimension reduction and secondary feature extraction;
s330, the normalized apparent resistivity and phase data enter a full-connection layer after convolution and pooling, a mapping relation between extracted features and an output layer is established in the full-connection layer, and the obtained earth model parameters and the earth model parameters of sample data are subjected to difference, namely, the error between an output value and a theoretical value;
repeating the forward information transmission and the reverse error transmission until the error reaches a desired value or the iteration number reaches a preset value, and stopping training by the network at the moment and obtaining an optimal weight and a threshold value;
s400, inputting test set data into a network, and obtaining required parameters of the ground model after convolution, pooling and one forward propagation of the test set data.
2. The method according to claim 1, wherein the step S310 includes:
the normalized apparent resistivity, which is the input data in step S200Phase data->Performing convolution processing, namely performing convolution transformation mapping on the input data of each position to a new value;
the convolution calculation formula is
Wherein the method comprises the steps ofThe net activation of the j-th channel of the first convolution layer is indicated by the characteristic map of the output of the previous layer +.>The i is the serial number of the input data, which is obtained after convolution summation and bias operation>The output of the jth channel representing the ith convolutional layer, f (·) representing the activation function, M j Representing a subset of feature maps, < >>Representing a convolution kernel matrix,/->The bias term representing the convolution layer, "×" is the convolution operation.
3. The method according to claim 2, wherein said step S320 comprises:
carrying out pooling treatment on the convolved data by adopting a maximum pooling method, namely reserving the maximum value of the data in a pooling window area, and activating by an activation function to obtain a result; the pooling calculation formula is:
wherein,net activation of the jth channel representing the ith pooling layer, by the feature map of the previous layer +.>Pooling, weighting and biasing to obtain the product; beta represents the pooling weight coefficient, +.>The bias term representing the pooling layer, pool (·) represents the pooling function.
4. A method according to claim 3, wherein the information in step S330 is propagated forward:
the input of CNN is a i，l-1 The information forward propagation algorithm calculation includes:
currently the full connection layer, there is a i，l ＝σ(z i，l )＝σ(W l a i，l-1 +b l ) Wherein a is i，l For forward propagation output, W l B is a network weight l As a network threshold, σ (·) is an activation function;
currently the convolutional layer, then there is a i，l ＝σ(z i，l )＝σ(W l *a i，l-1 +b l ) Wherein is a convolution operation;
currently the pooling layer, there is a i，l ＝pool(a i，l-1 ) Wherein pool (& gt) is poolingCalculating;
for output layer L: a, a i，L ＝ReLU(z i，L )＝ReLU(W L a i，L-1 +b L ) Wherein ReLU is an activation function;
finally, calculating the total gradient error delta of the output layer through a loss function i，L 。
5. The method according to claim 4, wherein the error in step S330 is transferred in reverse:
s331 the gradient error of the previous layer is delta i，l+1 The reverse transfer algorithm calculation includes:
currently the full connection layer, then there is delta i，l ＝(W l+1 ) T δ i，l+1 ⊙σ′(z i，l ) Wherein delta i，l Represents the gradient error of the first layer, +.;
currently the convolutional layer, then there is delta i，l ＝δ i，l+1 *(W l+i ) T ⊙σ′(z i，l )；
Currently the pooling layer, then there is delta i，l ＝upsample(δ i，l+1 )⊙σ′(z i，l ) Wherein upsample (·) is an upsampling function;
s332CNN updates W of the first layer l And b l Comprising:
currently is the fully connected layer:wherein alpha is iteration step length, and m is training sample number;
currently is a convolutional layer, with for each convolutional kernel:
6. the method according to claim 5, wherein before the step S400 of testing the set data input network, the following settings are made: setting the ratio of training set, test set and verification set to 98:1:1, and the number of network iterations to 200.
7. An electronic device comprising at least one processor, and a memory communicatively coupled to the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010605022.6A CN111812732B (en) | 2020-06-29 | 2020-06-29 | Geoelectromagnetic nonlinear inversion method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010605022.6A CN111812732B (en) | 2020-06-29 | 2020-06-29 | Geoelectromagnetic nonlinear inversion method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111812732A CN111812732A (en) | 2020-10-23 |
| CN111812732B true CN111812732B (en) | 2024-03-15 |
Family
ID=72855619
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010605022.6A Active CN111812732B (en) | 2020-06-29 | 2020-06-29 | Geoelectromagnetic nonlinear inversion method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111812732B (en) |
Families Citing this family (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112364911B (en) * | 2020-11-06 | 2022-11-11 | 东北石油大学 | Resistivity imaging inversion method and device and storage medium |
| CN112733386B (en) * | 2021-01-25 | 2021-08-17 | 吉林大学 | Multi-component reactive solute transport parameter inversion method based on deep neural network |
| CN112862837B (en) * | 2021-01-27 | 2023-06-23 | 南京信息工程大学 | Image processing method and system based on convolutional neural network |
| CN113158553B (en) * | 2021-03-25 | 2022-04-01 | 湖南师范大学 | Magnetotelluric signal noise suppression method and system based on CNN-LSTM |
| 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 |
| 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 |
| CN113484919B (en) * | 2021-06-30 | 2022-11-04 | 长江大学 | Magnetic anomaly inversion method, system, terminal and medium based on convolutional neural network |
| CN113568055B (en) * | 2021-07-16 | 2022-07-01 | 北京大学 | Aviation transient electromagnetic data inversion method based on LSTM network |
| CN113568056B (en) * | 2021-07-16 | 2022-07-01 | 北京大学 | Aviation transient electromagnetic data inversion method based on convolutional neural network |
| CN113807020B (en) * | 2021-09-28 | 2023-04-07 | 成都理工大学 | Magnetotelluric inversion method based on deep learning constraint |
| CN114048823A (en) * | 2021-11-25 | 2022-02-15 | 成都理工大学 | Resistivity inversion model establishment method based on full convolution network |
| CN116644658B (en) * | 2023-05-11 | 2024-03-15 | 华中科技大学 | Gas discharge temperature inversion method based on physical information neural network |
| CN117390961A (en) * | 2023-10-25 | 2024-01-12 | 中南大学 | Method for solving electromagnetic response of ground model by using physical information neural network |
| CN118194728A (en) * | 2024-05-16 | 2024-06-14 | 西南交通大学 | Magnetotelluric physical and data hybrid driving inversion method |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355244A (en) * | 2016-08-30 | 2017-01-25 | 深圳市诺比邻科技有限公司 | CNN (convolutional neural network) construction method and system |
| KR20180120413A (en) * | 2017-04-27 | 2018-11-06 | (주)지오룩스 | System for processing and analyzing magnetotellurics monitoring data using deep learning algorithm |
| CN109188536A (en) * | 2018-09-20 | 2019-01-11 | 成都理工大学 | Time-frequency electromagnetism and magnetotelluric joint inversion method based on deep learning |
| WO2019040288A1 (en) * | 2017-08-25 | 2019-02-28 | Exxonmobil Upstream Researchcompany | Automated seismic interpretation using fully convolutional neural networks |
| CN109490985A (en) * | 2018-12-24 | 2019-03-19 | 桂林理工大学 | A kind of geophysical reconnaissance system and method |
| CN109934272A (en) * | 2019-03-01 | 2019-06-25 | 大连理工大学 | A kind of image matching method based on full convolutional network |
| CN110346834A (en) * | 2019-07-22 | 2019-10-18 | 中国科学院地球化学研究所 | The forward modeling method of three-dimensional frequency domain controllable source electromagnetism, system |
| CN110515123A (en) * | 2019-08-22 | 2019-11-29 | 中国海洋大学 | Convolutional neural networks well-shooting joint inversion method based on small-sample learning |
| CN110968826A (en) * | 2019-10-11 | 2020-04-07 | 重庆大学 | Magnetotelluric deep neural network inversion method based on spatial mapping technology |
| CN111126591A (en) * | 2019-10-11 | 2020-05-08 | 重庆大学 | Magnetotelluric deep neural network inversion method based on space constraint technology |
| CN111143984A (en) * | 2019-12-23 | 2020-05-12 | 贵州大方煤业有限公司 | Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US10996372B2 (en) * | 2017-08-25 | 2021-05-04 | Exxonmobil Upstream Research Company | Geophysical inversion with convolutional neural networks |
-
2020
- 2020-06-29 CN CN202010605022.6A patent/CN111812732B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106355244A (en) * | 2016-08-30 | 2017-01-25 | 深圳市诺比邻科技有限公司 | CNN (convolutional neural network) construction method and system |
| KR20180120413A (en) * | 2017-04-27 | 2018-11-06 | (주)지오룩스 | System for processing and analyzing magnetotellurics monitoring data using deep learning algorithm |
| WO2019040288A1 (en) * | 2017-08-25 | 2019-02-28 | Exxonmobil Upstream Researchcompany | Automated seismic interpretation using fully convolutional neural networks |
| CN109188536A (en) * | 2018-09-20 | 2019-01-11 | 成都理工大学 | Time-frequency electromagnetism and magnetotelluric joint inversion method based on deep learning |
| CN109490985A (en) * | 2018-12-24 | 2019-03-19 | 桂林理工大学 | A kind of geophysical reconnaissance system and method |
| CN109934272A (en) * | 2019-03-01 | 2019-06-25 | 大连理工大学 | A kind of image matching method based on full convolutional network |
| CN110346834A (en) * | 2019-07-22 | 2019-10-18 | 中国科学院地球化学研究所 | The forward modeling method of three-dimensional frequency domain controllable source electromagnetism, system |
| CN110515123A (en) * | 2019-08-22 | 2019-11-29 | 中国海洋大学 | Convolutional neural networks well-shooting joint inversion method based on small-sample learning |
| CN110968826A (en) * | 2019-10-11 | 2020-04-07 | 重庆大学 | Magnetotelluric deep neural network inversion method based on spatial mapping technology |
| CN111126591A (en) * | 2019-10-11 | 2020-05-08 | 重庆大学 | Magnetotelluric deep neural network inversion method based on space constraint technology |
| CN111143984A (en) * | 2019-12-23 | 2020-05-12 | 贵州大方煤业有限公司 | Magnetotelluric two-dimensional inversion method based on genetic algorithm optimization neural network |
Non-Patent Citations (2)
| Title |
|---|
| 《基于深度神经网络的人脸识别方法研究》;赵远东;《CNKI优秀硕士学位论文全文库》;20190815(第08期);第9-18、26页 * |
| 大地电磁人工神经网络反演;王鹤等;《中南大学学报(自然科学版)》;20150526;第46卷(第05期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111812732A (en) | 2020-10-23 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111812732B (en) | Geoelectromagnetic nonlinear inversion method based on convolutional neural network | |
| Sun et al. | Physics-constrained bayesian neural network for fluid flow reconstruction with sparse and noisy data | |
| Pakravan et al. | Solving inverse-PDE problems with physics-aware neural networks | |
| Jiang et al. | Electrical resistivity imaging inversion: An ISFLA trained kernel principal component wavelet neural network approach | |
| CN113505534B (en) | Load prediction method considering demand response | |
| CN113568056B (en) | Aviation transient electromagnetic data inversion method based on convolutional neural network | |
| Oishi et al. | A surface-to-surface contact search method enhanced by deep learning | |
| Yao et al. | Enhanced supervised descent learning technique for electromagnetic inverse scattering problems by the deep convolutional neural networks | |
| CN116245013A (en) | Geological prediction model construction method, modeling method, equipment and storage medium | |
| Dalton et al. | Emulation of cardiac mechanics using Graph Neural Networks | |
| Mosegaard et al. | Inverse methods: Problem formulation and probabilistic solutions | |
| CN115496144A (en) | Power distribution network operation scene determining method and device, computer equipment and storage medium | |
| CN113486591B (en) | Gravity multi-parameter data density weighted inversion method for convolutional neural network result | |
| CN114779356A (en) | Formation electrical profile rapid imaging method based on array resistivity | |
| Jiang et al. | An E-PINN assisted practical uncertainty quantification for inverse problems | |
| Xie et al. | Automatic boundary fitting framework of boundary dependent physics-informed neural network solving partial differential equation with complex boundary conditions | |
| CN116151135B (en) | Electromagnetic simulation method and system for electric large-size target | |
| Laskar et al. | Artificial Neural Networks and Gene Expression Programing based age estimation using facial features | |
| CN114155354B (en) | Method and device for reconstructing capacitance tomography based on graph convolution network | |
| Xie et al. | 2D magnetotelluric inversion based on ResNet | |
| Aria et al. | Back propagation neural network for classification of IRS-1D satellite images | |
| Liu et al. | Applying machine learning methods to river topography prediction under different data abundances | |
| Ibrahim et al. | Determining optimum structure for artificial neural network and comparison between back-propagation and levenberg-marquardt training algorithms | |
| CN115951405B (en) | Speed change imaging method based on sparse convolution and adaptive moment estimation | |
| CN118171501B (en) | Three-dimensional geoelectromagnetic forward modeling method for pre-training model cyclic transfer learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |