阅读说明：本技术 一种基于卷积神经网络的航空瞬变电磁数据反演方法 (Aviation transient electromagnetic data inversion method based on convolutional neural network ) 是由 黄清华 吴思弘 于 2021-07-16 设计创作，主要内容包括：本发明公开了一种基于卷积神经网络的航空瞬变电磁反演方法,包括以下步骤：1)生成由航空瞬变电磁响应数据、收发装置高度和电阻率模型构成的合成数据集；2)根据合成数据集建立卷积神经网络,以航空瞬变电磁观测的时域信号为输入数据,以其对应介质模型的电阻率对数值为输出数据；3)根据误差下降曲线选取合适的训练集规模和训练周期,确保训练效果和计算效率；完成网络训练；4)根据网络反演结果及其正演响应与真实值的拟合情况判断网络的泛化能力；5)将新采集的航空瞬变电磁响应数据输入至卷积神经网络中,得到地下介质的电阻率分布情况。该方法反演结果准确,能够提高航空瞬变电磁数据解释效率,进而为航空瞬变电磁数据实时成像提供技术支持。(The invention discloses an aviation transient electromagnetic inversion method based on a convolutional neural network, which comprises the following steps of: 1) generating a synthetic data set consisting of the aviation transient electromagnetic response data, the height of the transceiver and the resistivity model; 2) establishing a convolution neural network according to the synthetic data set, taking the time domain signal of the aviation transient electromagnetic observation as input data, and taking the resistivity logarithm value of the corresponding medium model as output data; 3) selecting proper training set scale and training period according to the error reduction curve to ensure training effect and calculation efficiency; completing network training; 4) judging the generalization capability of the network according to the network inversion result and the fitting condition of the forward response and the true value thereof; 5) and inputting the newly acquired aviation transient electromagnetic response data into the convolutional neural network to obtain the resistivity distribution condition of the underground medium. The method has accurate inversion result, can improve the interpretation efficiency of the aviation transient electromagnetic data, and further provides technical support for real-time imaging of the aviation transient electromagnetic data.)

1. An aviation transient electromagnetic data inversion method based on a convolutional neural network comprises the following steps:

A. generating a synthetic data set comprising:

A1. generating a layered resistivity model according to the natural resistivity value range and the longitudinal grid adopted by inversion;

A2. randomly generating a height of the transceiver;

A3. forward modeling of the laminar resistivity model is carried out according to parameters of an observation system of the aviation transient electromagnetism and a sampling mode, and aviation transient electromagnetic response data are obtained;

the synthetic dataset samples comprise: containing N_tAirborne transient electromagnetic response data d of each time sampling point^LHeight h of the transceiver and N of the sheet resistivity model_ρLayer resistivity value m^L(ii) a The generated synthetic data set is divided into a training set and a testing set;

B. establishing a convolutional neural network: the convolutional neural network comprises an input layer, a convolutional layer, a full-link layer and an output layer, wherein the input data are aviation transient electromagnetic response observation data, the output data are resistivity model parameters predicted by the network, and the dimensions of the input layer and the output layer are determined according to time sampling points and the model parameters respectively; the convolutional layer comprises convolution, pooling and activation operations; unfolding the feature data after the convolutional layer operation into a one-dimensional vector, splicing the one-dimensional vector with the height of the transceiver device, and transmitting the one-dimensional vector into a full-connection layer; after the operation of the full connection layer, the network outputs the predicted resistivity model parameters;

C. selecting proper training set scale and training period: iteratively adjusting parameters in the network according to the difference between the network output result and the real resistivity model until the network training is converged; determining the scale and the training period of the training set according to the descending condition of inversion errors of the training sets of different scales along with the increase of the training period, ensuring the network convergence effect and the training efficiency, and obtaining a convolutional neural network after training;

D. testing the network inversion effect: calculating the error between the inversion result of the network to the test set and the real resistivity model and the error between the corresponding aviation transient electromagnetic response data, and evaluating the inversion effect of the network and the fitting condition of the response data;

E. resistivity model inversion: and inputting the newly acquired aviation transient electromagnetic response data into a convolutional neural network to obtain network output, namely a resistivity result of network inversion.

2. The aviation transient electromagnetic data inversion method based on the convolutional neural network as claimed in claim 1, wherein in the step A, a longitudinal uniform grid is adopted to subdivide a layered resistivity model; the resistivity values are distributed in 1-10000 omega-m and continuously change along with the depth, and logarithmic values of the resistivity values are taken as network target output; the height of the transceiver is randomly set between 25 m and 100 m.

3. The method for inverting aviation transient electromagnetic data based on a convolutional neural network as claimed in claim 2, wherein in step a, the resistivity value of the laminar resistivity model is continuously changed with the depth by an interpolation method.

4. The aviation transient electromagnetic data inversion method based on the convolutional neural network as claimed in claim 1, wherein in the step B, the convolutional neural network is structurally designed as follows:

for dimension N_tApplying a convolutional neural network structure comprising 3 convolutional layers and 3 fully-connected layers, with an output dimension of N_ρThe output data of (1); wherein:

the convolution kernel size of the first convolution layer is 1 × 15, and the number of convolution kernels isThe convolution step is 1;

the convolution kernel size of the second convolution layer is 1 × 15, and the number of convolution kernels isThe convolution step is 1;

the convolution kernel size of the third convolution layer is 1 × 5, and the number of convolution kernels isThe convolution step is 1;

the convolution layers are all subjected to average pooling with the size of 1 multiplied by 2 and the step length of 2;

the activation function takes f (x) max (0, x);

the first convolutional layer operation outputs a data dimension of

The second convolution layer operation outputs a data dimension of

The third convolution layer operation outputs data with dimensions of

WhereinRepresents rounding up;

after the three convolutional layers are operated, the method will be describedThe output data of the third convolutional layer is stretched into a one-dimensional vector with the length ofAnd splicing with the height of the transceiver.

5. The convolutional neural network-based airborne transient electromagnetic data inversion method as defined in claim 1, wherein in step C, N is included_tInputting the aviation transient electromagnetic response data of each time sampling point into the convolutional neural network established in the step B, and outputting N through an output layer_ρValue v^O，v^ON with layered resistivity model_ρInversion result m of layer resistivity value^OThe relationship of (1) is: v. of^O＝lg(m^O) (ii) a Iteratively adjusting network parameters to reduce the objective function such that v^OClose to lg (m)^L) And further completing the network training.

6. The aviation transient electromagnetic data inversion method based on the convolutional neural network as claimed in claim 5, wherein the objective function adopted for training the convolutional neural network in the step C is as follows:

wherein N is_sFor training set sample number, N_ρNumber of layers, v, of resistivity model^OFor the network output value, m^LIs the true resistivity value, λ_WThe value range is 0.001-1 for regularization parameters, W represents a transfer matrix and a convolution kernel contained in the network, b represents a bias vector in the convolution layer and the full connection layer, | | · |, L₂Indicating the Euclidian distance.

7. The aviation transient electromagnetic data inversion method based on the convolutional neural network as claimed in claim 6, wherein in the step C, the convolutional neural network is trained, network parameters W and b are iteratively adjusted by adopting an Adam algorithm, and the learning rate value range is 0.001-0.1.

8. The aviation transient electromagnetic data inversion method based on the convolutional neural network as claimed in claim 6, wherein in step C, the variation trend of the network inversion effect along with the training period is judged by using a Root Mean Square Error (RMSE), wherein the RMSE is defined as follows:

the training set size and training period are selected based on the trend of RMSE descent during training.

9. The convolutional neural network-based airborne transient electromagnetic data inversion method as claimed in claim 5, wherein in step D, the inversion effect is evaluated by using a relative root mean square error RMSPE, which is defined as follows:

wherein, RMSPE_modelRMSPE for resistivity model parameter relative root mean square error_signalRelative root mean square error of aviation transient electromagnetic response data; n is a radical of_ρNumber of layers for resistivity model, N_tIs the number of time sampling points, v^OOutput vector, m, for the network^LIs the true resistivity value; d^OOutputting to the network airborne transient electromagnetic response data corresponding to the resistivity model, d^LAnd obtaining aviation transient electromagnetic response data corresponding to the real medium model.

Technical Field

The invention provides a quick inversion method of terrestrial electromagnetic data, and particularly relates to an aviation transient electromagnetic data inversion method based on a convolutional neural network.

Background

The aviation transient electromagnetic method is an important active source electromagnetic exploration method, has the advantages of strong adaptability to terrain, high observation efficiency, sensitivity to low-resistance abnormal bodies and the like, and is widely applied to the fields of underground water exploration, environment monitoring, mineral exploration and the like.

At present, the aviation transient electromagnetic data interpretation method mainly comprises deterministic inversion and statistical inversion. The deterministic inversion method first sets an initial model and an objective function. The objective function is typically the sum of the regularization term and the fitting error between the forward modeling response and the observed data. The regularization term is used for fusing prior information such as a reference model and the smoothness of the model. And (4) iteratively adjusting the model parameters to make the objective function converge to a preset acceptable range to obtain a final interpretation model. However, due to the influence of multi-solution, the target function has a plurality of local minimum values, and the inversion result depends on the initial model and is easy to fall into a local optimal solution; and the selection of the regularization item is subjective. In addition, multiple forward modeling and Jacobian matrix calculation are needed when the model parameters are adjusted iteratively, and time is consumed. The statistical inversion method provides statistical characteristics of model parameters by sampling in a model space, but the model space is large, so that the requirement on computing resources is high, and the computing cost rises exponentially along with the increase of the model parameters.

The aviation transient electromagnetic method is wide in observation area, large in data collection amount and multi-source, forward response fitting needs to be carried out on each emission source during inversion, and huge calculation challenges are faced in data interpretation. Meanwhile, compared with a ground observation system, the airborne electromagnetic data has a low signal-to-noise ratio, is easily interfered by noise, and can aggravate the problem of multiple solutions of inversion. In general, interpretation of airborne transient electromagnetic data using conventional inversion methods is still limited by the multiplicity and computational cost.

The deep learning method is a data-driven nonlinear global optimization method, and is applied to an electromagnetic Neural Network (simulation) in Bai et al (2020, (Quasi-) Real-time inversion of airborne time-domain electronic data acquisition Neural Network. Remote sensing.12(20),3440), Noh et al (2020, Imaging sub-surface sensitivity structure from airborne electronic data acquisition using device Neural Network. expansion mapping, 51(2),214-220) and Feng et al (2020, resistance-depth Imaging with the airborne Neural Network. journal of the aviation Network. inversion of transient, 25) and applies to the electromagnetic Neural Network (N) is N) (N) N, N) is N, N. Li and the like (2020, Fast imaging of time-domain aircraft EM data using deep learning technologies, 85(5), E163-E170) combine a Convolutional Neural Network (CNN) with a Long-Short-Term Memory module (LSTM), and the aviation transient electromagnetic data inversion is realized by considering the flight height. The neural network inversion algorithms are simple in structure, mostly adopt a 3-5-layer ladder-shaped resistivity model, the thickness of a longitudinal grid is continuously increased along with the depth, the neural network inversion algorithms are difficult to flexibly adapt to the boundary of a resistivity abnormal body, and the neural network inversion algorithms are weak in applicability to a complex resistivity environment in practical application.

In summary, the existing aviation transient electromagnetic data interpretation method still has limitations, and an efficient and accurate data interpretation method needs to be researched.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides an aviation transient electromagnetic data inversion method based on a convolutional neural network, and the aviation transient electromagnetic inversion problem with large data volume and strong multi-solution is solved. The data interpretation method is high in applicability, high in efficiency and good in accuracy, and provides technical support for aviation transient electromagnetic real-time data imaging.

The technical scheme adopted by the invention is as follows:

an aviation transient electromagnetic data inversion method based on a convolutional neural network is characterized in that a synthetic data set is generated according to a natural resistivity rule and observation parameters; constructing a proper convolutional neural network structure according to the synthetic data set and finishing network training; migrating the trained network to newly acquired observation data to realize rapid interpretation of aviation transient electromagnetic data; the method specifically comprises the following steps:

A. generating a synthetic data set comprising:

A1. generating a layered resistivity model according to the natural resistivity value range and the longitudinal grid adopted by inversion;

A2. randomly generating a height of the transceiver;

A3. forward modeling of the laminar resistivity model is carried out according to parameters of an observation system of the aviation transient electromagnetism and a sampling mode, and aviation transient electromagnetic response data are obtained;

dividing the synthetic dataset samples into a network training dataset and a test dataset, wherein the samples comprise: containing N_tAirborne transient electromagnetic response data d of each time sampling point^LHeight h of the transceiver and N of the sheet resistivity model_ρLayer resistivity value m^L。

B. Establishing a convolutional neural network;

the convolutional neural network is composed of an input layer, a convolutional layer, a full-link layer and an output layer. The input data are time domain signals observed by aviation transient electromagnetism, the output data are resistivity model parameters (logarithm values of the resistivity values) predicted by a network, and the dimensions of the input layer and the output layer are determined according to time sampling points and the model parameters respectively. The convolutional layer contains convolution, pooling, and activation operations. And unfolding the characteristic data subjected to convolutional layer operation into a one-dimensional vector, splicing the one-dimensional vector with the height of the transceiver, and transmitting the one-dimensional vector to the full-connection layer. After the operation of the full connection layer, the network outputs the predicted resistivity model parameters.

C. Selecting proper training set scale and training period to complete network training;

training the network with training data sets of different sizes: iteratively adjusting parameters in the network according to the difference between the network output result and the real resistivity model until the network training is converged, specifically:

will contain N_tInputting the aviation transient electromagnetic response data of the time sampling points into the convolutional neural network established in the step B, and outputting N through an output layer_ρValue v^O，v^ON with layered resistivity model_ρInversion result m of layer resistivity value^OThe relationship of (1) is: v. of^O＝lg(m^O). Iteratively adjusting network parameters to reduce the objective function such that v^OClose to lg (m)^L) And further completing the network training.

And determining the scale of the training set and the training period according to the descending condition of inversion errors of the training sets with different scales along with the increase of the training period to obtain the convolutional neural network for completing the training, and ensuring the network convergence effect and the training efficiency.

D. Testing the network inversion effect;

and calculating the error between the inversion result of the network on the test set and the real resistivity model and the error between the corresponding aviation transient electromagnetic response data, evaluating the inversion effect of the network and the fitting condition of the response data, and judging the generalization capability of the network.

E. Inversion of the resistivity model;

and C, inputting newly acquired aviation transient electromagnetic response data into the convolutional neural network which is trained in the step C, and obtaining network output which is a resistivity result of network inversion.

Through the steps, aviation transient electromagnetic response data rapid inversion based on the convolutional neural network is achieved, and the resistivity distribution condition of the underground medium is obtained.

As a preferred scheme, in the step a, a longitudinal uniform grid is adopted to subdivide a layered resistivity model; the resistivity values are distributed in 1-10000 omega-m and continuously change along with the depth, and logarithmic values of the resistivity values are taken as network target output; the height of the transceiver is randomly set between 25 m and 100 m. Specifically, the resistivity value of the resistivity model is continuously changed along with the depth through interpolation, and the depth is finely divided by adopting a uniform grid, so that the longitudinal resolution is enhanced.

As a preferable scheme, in the step B, the dimension is N_tApplying a convolutional neural network structure comprising 3 convolutional layers and 3 fully-connected layers, with an output dimension of N_ρThe output data of (1). The convolutional neural network consists of an input layer, 3 convolutional layers, 3 pooling layers and an output layer; the number of convolution kernels in the 3 convolution layers is sequentiallyAndconvolution kernel sizes are 1 respectivelyX 15, 1 x 15 and 1 x 5, the convolution step size is 1. Adopting average pooling with the size of 1 × 2 and the step length of 2; the activation function takes f (x) max (0, x).

The first convolutional layer operation outputs a data dimension of

The second convolution layer operation outputs a data dimension of

The third convolution layer operation outputs data with dimensions of

WhereinRepresents rounding up;

after the three convolutional layers are operated, the output data of the third convolutional layer is stretched into a one-dimensional vector with the length ofAnd splicing with the height of the transceiver.

As a preferred scheme, the objective function used for training the convolutional neural network in step C is:

wherein N is_sFor training set sample number, N_ρNumber of layers, v, of resistivity model^OFor the network output value, m^LIs the true resistivity value, λ_WThe value range is 0.001-1 for regularization parameters, W represents a transfer matrix and a convolution kernel contained in the network, b represents a bias vector in the convolution layer and the full connection layer, | | · |, L₂Indicating the Euclidian distance.

As a preferred scheme, the network parameters W and b are iteratively adjusted by the training network in the step C by adopting an Adam algorithm, and the learning rate value range is 0.001-0.1.

As a preferable scheme, in the step C, for training sets of different scales, a Root Mean Square Error (RMSE) between an inversion result of the convolutional neural network on the training set and the test set in the training process and a theoretical model is used to quantitatively evaluate a variation condition of network convergence along with a training period. The root mean square error is defined as follows:

according to the trend that the RMSE of sample sets with different scales declines along with the increase of the training period, the proper training set scale and the training period are selected to obtain the convolutional neural network for completing the training, so that the fast training speed and the good convergence result are ensured to be obtained.

As a preferred scheme, in the step D, in order to determine the inversion effect of the network, fitting conditions of parameters in multiple orders of magnitude are considered in a balanced manner, and the inversion effect of the network in a test set and the fitting conditions of response data are evaluated by using a relative root mean square error (RMSPE), so as to reduce the influence of the order of magnitude difference of the parameters. RMSPE is defined as follows:

wherein, RMSPE_modelRMSPE for resistivity model parameter relative root mean square error_signalRelative root mean square error of aviation transient electromagnetic response data; d^OOutputting to the network airborne transient electromagnetic response data corresponding to the resistivity model, d^LAnd obtaining aviation transient electromagnetic response data corresponding to the real medium model.

The invention has the beneficial effects that: the method has the advantages that a proper convolutional neural network is established for aviation transient electromagnetic data, the height of a transceiver is used as a variable parameter to be introduced into an inversion process, a resistivity model with continuous change is considered, the inversion problem of the aviation transient electromagnetic data with strong multi-solution performance and large data volume is remarkably improved, the dependence of a traditional inversion method on an initial model is overcome, regularization items are avoided, and subjective influence is reduced. Numerical simulation results prove that the inversion result of the method is accurate, the interpretation efficiency of the aviation transient electromagnetic data can be improved, and technical support is further provided for real-time imaging of the aviation transient electromagnetic data.

Drawings

FIG. 1 is a schematic diagram of the structure of a convolutional neural network of the method of the present invention.

FIG. 2 shows the RMSE of the training set and the test set varying with the training period when the sizes of the training set are 8000, 40000, 80000, 120000 and 160000, respectively.

FIG. 3 is a RMSPE distribution diagram of inversion results and forward responses of a convolutional neural network and a Gaussian Newton method on a test set sample in an embodiment of the invention.

FIG. 4 is a comparison result of the convolution neural network and the Gaussian Newton method inversion with the real model for the same test model, and a corresponding aviation transient electromagnetic response comparison result.

Detailed Description

The aviation transient electromagnetic data inversion method based on the convolutional neural network provided by the invention is further explained with reference to the attached drawings. The idea of the invention is described as follows: generating a sample data set according to the aviation transient electromagnetic observation parameters and the natural resistivity characteristics; constructing a proper convolutional neural network structure according to the synthetic data set and finishing network training; and migrating the trained convolutional neural network to a newly acquired response data set, and aiming at realizing rapid quasi-real-time aviation transient electromagnetic data interpretation. The method has the advantages of high accuracy and high calculation speed.

A. Generating a synthetic data set;

A1. generating a resistivity model according to the inversion parameters;

the resistivity model adopts a one-dimensional layered medium structure, the number of layers is randomly set within 15 layers, and the depth of an interface is within 600 m; the distance between the center depths of adjacent resistive layers is greater than 15 m; no constraint is imposed on the thickness of the first layer of medium so as to improve the resolution of the shallow layer of medium; the resistivity value ranges from 1 to 10000 Ω · m. Randomly generating the number of dielectric layers, the resistivity value and the center depth of each layer of dielectric according to the conditions; obtaining a continuously changing resistivity value within a depth range of 600m by adopting a cubic spline interpolation method, wherein each 2m is one layer, and the total number of layers is 300; the depth below 600m is regarded as a uniform half-space medium, and the resistance value is consistent with that of the adjacent upper layer (598-600 m); discarding the model with the resistivity value exceeding the preset range after interpolation; taking a logarithmic value of the resistivity as an inversion parameter;

A2. randomly generating a transceiver height between 25-100 m;

A3. forward modeling is carried out according to parameters and sampling modes of an observation system of aviation transient electromagnetism to obtain the dB response of the aviation transient electromagnetism_zA dt component;

based on the resistivity model generated in the step A1 and the height of the transceiver generated in the step A2, according to a conventional aviation transient electromagnetic observation mode, the radius of a transmitting coil is 6m, and the central distance between a sensor and the transmitting coil is 4 m; after the analog current is turned off 10^-5-10^{- 1}s transient electromagnetic response dB_zThe/dt component, 100 time samples in total, is equally logarithmically spaced.

B. Establishing a convolutional neural network;

as shown in fig. 1, the network input is aviation transient electromagnetic response data, and the dimension of the input layer is equal to the number of time sampling points. According to the sample dimension of the step A, the dimension of the input layer of the network is 1 × 100, and the dimension of the output layer is 1 × 300. There are 3 convolutional layers and 3 fully-connected layers. The convolutional layer contains convolution, pooling, and activation operations. The convolution kernel sizes of the 3 convolutional layers are 1 × 15, 1 × 15 and 1 × 5, respectively, the numbers are 32, 64 and 128, respectively, and the step sizes are all 1. After the convolution operation, an average pooling operation and an activation operation are sequentially performed with a size of 1 × 2 and a step size of 2, and f (x) max (0, x) is used as an activation function. And unfolding the characteristic data subjected to convolutional layer operation into a one-dimensional vector, splicing the one-dimensional vector with the height h of the transceiver, inputting the one-dimensional vector into a full-connection layer, wherein the dimensions of the full-connection layer are respectively 1000 and 600, and finally outputting an inversion resistivity model by an output layer.

C. Selecting proper training set scale and training period to complete network training;

training the convolutional neural network by adopting training sets with sample numbers of 8000, 40000, 80000, 120000 and 160000 respectively, and judging the network convergence effect by adopting 4000 groups of samples which do not participate in the training as test sets. The objective function used for network training is as follows:

wherein N is_sFor training the number of samples, N_ρ300 is the number of parameters of the resistivity model, v^oFor network output, lg (m)^L) Is the logarithmic value of the resistivity of the real model, lambda_WIs 0.001. Parameters W and b in the network are iteratively adjusted by adopting an Adam algorithm, and the learning rate is 0.001.

The RMSE of the training and test sets as a function of training period is shown in fig. 2. And a network with the training sample number of 80000 and the training period of 600 times is selected as a final inversion network, so that a better convergence degree and shorter training time are ensured.

D. Testing the network inversion effect;

in order to evaluate the inversion effect of the network, the RMSPE between the inversion result of the convolutional neural network on 4000 groups of samples in the test set and the real resistivity model and the RMSPE between the real resistivity model and the corresponding aviation transient electromagnetic response are calculated, and compared with a classical traditional inversion method, namely a gauss-newton method, the RMSPE is distributed as shown in fig. 3. The initial model adopted by the gauss-newton method is a uniform half-space model of 100 ohm meters, and the inversion parameters are resistivity logarithm values of 300 layers of media. The inverse model of the gauss-newton method and the RMSPE of the response are widely distributed and have more extreme values. The RMSPE distribution of the convolutional neural network is concentrated and is close to the origin of coordinates, which shows that the inversion effect of the convolutional neural network on a test set is stable and the convolutional neural network has strong generalization capability. FIG. 4 is an inversion result of a sample in a test set and its response curve. The gauss-newton method is susceptible to multi-solution, and cannot be further converged to a real model under the condition that the forward response of the iterative model is very close to the observed value. The convolutional neural network does not depend on the selection of the initial model, and can accurately present the resistivity value. In addition, after the training of the convolutional neural network is finished, the inversion result of 4000 groups of data of the whole test set can be obtained only by one-time forward propagation, and the time consumption is less than 1 second.

The above description is only exemplary of the present invention and should not be taken as limiting the invention, as any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.