Fast time domain aviation electromagnetic data variable dimension Bayesian inversion method, system and device and medium

Abstract

The invention relates to the technical field of aviation electromagnetic detection, in particular to a rapid time domain aviation electromagnetic data variable-dimension Bayesian inversion method, system and device and a medium. According to the method, a mapping relation between geoelectric model parameters and time domain electromagnetic response data is constructed by adopting a one-dimensional convolutional neural network according to a deep learning network, and rapid prediction of the time domain aviation electromagnetic response data in a variable-dimension Bayesian inversion process is realized; secondly, during parallel calculation, only one Markov chain is operated in the preheating period, and N Markov chains are executed in parallel after the preheating period is ended, so that the parallel strategy can avoid the problems that the N Markov chains are executed in parallel from the beginning of the preheating period, and the sampling of the Markov chains is not convergent and the calculation time is wasted due to a relatively poor initial state; the condition that Markov chain sampling is not convergent can be found in time in the sampling process of the preheating period, so that the next operation is stopped, and wrong calculation is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of airborne electromagnetic detection, and in particular to a fast variable-dimensional Bayesian inversion method, system, equipment and medium for time-domain airborne electromagnetic data. Background Art

[0002] In mineral resource exploration, this method can quickly process time-domain airborne electromagnetic data, obtain inversion results and uncertainty information of inversion parameters, effectively identify the location of underground ore bodies, and evaluate the reliability of inversion results to improve the efficiency and accuracy of prospecting. In the field of environmental monitoring, this method can be used for groundwater resource surveys, pollution area division, etc. Through airborne electromagnetic detection, combined with variable-dimensional Bayesian inversion, the distribution and pollution range of groundwater can be accurately obtained, providing a scientific basis for environmental monitoring and protection.

[0003] The variable-dimensional Bayesian inversion method has been applied to the inversion of frequency-domain and time-domain airborne electromagnetic data. For example, Yin Changchun et al. introduced weighting coefficients in the variable-dimensional Bayesian inversion method in the article "Research on variable-dimensional Bayesian inversion of frequency-domain airborne electromagnetic data" to adjust the constraint strength of the inverse model to improve the effect of variable-dimensional Bayesian inversion. Yu Xiaodong et al. introduced an adaptive sampling step adjustment method in the article "Research on adaptive variable-dimensional Bayesian inversion of time-domain airborne electromagnetic data" to solve the impact of unreasonable sampling step settings, so as to improve the efficiency of model sampling and improve the accuracy of inversion results.

[0004] The existing variable-dimensional Bayesian inversion of time-domain airborne electromagnetic data mainly solves the effect of the inversion result. However, due to the dense sampling of time-domain airborne electromagnetic data, there is a lot of data to be inverted, and each measuring point needs to be sampled millions of times when performing variable-dimensional Bayesian inversion. As a result, the existing variable-dimensional Bayesian inversion method is computer-intensive when processing airborne electromagnetic data, and the calculation time required is very long, which makes this method difficult to apply to the inversion of airborne electromagnetic data in actual engineering areas. Summary of the invention

[0005] Aiming at the problem that the variable-dimensional Bayesian inversion method of time-domain airborne electromagnetic data has large calculation amount and long inversion time, the present invention proposes a fast variable-dimensional Bayesian inversion method, system, equipment and medium for time-domain airborne electromagnetic data; a one-dimensional convolutional neural network is adopted to construct a mapping relationship between geoelectric model parameters and time-domain electromagnetic response data according to a deep learning network, so as to realize the fast prediction of time-domain airborne electromagnetic response data in the variable-dimensional Bayesian inversion process; secondly, in parallel calculation, only one Markov chain is run in the preheating period, and N Markov chains are executed in parallel after the preheating period ends. This parallel strategy can avoid the problem that N Markov chains are executed in parallel from the beginning of the preheating period, and the initial state is poor, resulting in the Markov chain sampling non-convergence and the waste of calculation time. The situation that the Markov chain sampling does not converge can be found in time during the sampling process of the preheating period, so as to stop the next step of calculation and avoid erroneous calculation.

[0006] The specific implementation contents of the present invention are as follows:

[0007] A fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data, comprising the following steps:

[0008] Step S1: Initialize the variable-dimensional Bayesian inversion parameter range, the warm-up period model step length s1 and the warm-up period end condition according to the input time-domain airborne electromagnetic data to be inverted;

[0009] Step S2: Set the number of warm-up period Markov chains to 1, construct the initial state m0 of the warm-up period Markov chain based on the conductivity-depth imaging CDI results of the time-domain airborne electromagnetic data, and set the current sampling model for the first sampling to m c =m0;

[0010] Step S3: adopting the overall update method to perform cyclic sampling within the set model parameter space range;

[0011] Step S4: Initialize the model step size s2 after the warm-up period, and set the number of Markov chains after the warm-up period to N, where N>1;

[0012] Step S5: Construct the initial state m of N Markov chains t , set the current sampling model of the first sampling of the t-th Markov chain to m c '=m t , execute N Markov chains in parallel;

[0013] Step S6: using a component updating method to perform cyclic sampling within the set model parameter space range;

[0014] Step S7: Statistically analyze the model sampling results, analyze the uncertainty of the inversion parameters based on the statistical results, and estimate the inversion results of the time domain airborne electromagnetic data.

[0015] In order to better implement the present invention, further, the step S3 specifically includes the following steps:

[0016] Step S31: Randomly generate a suggested model m according to the model suggestion distribution p ;

[0017] Step S32: Calculate the likelihood function according to the acquired electromagnetic response, the forward modeled electromagnetic response, and the covariance matrix of the observed data;

[0018] Step S33: Calculate the acceptance probability according to the likelihood function, the prior probability distribution ratio, and the proposed distribution ratio;

[0019] Step S34: Determine the recommended model m based on the acceptance probability p Is it accepted? If the proposed model m p If accepted, update the current model m c is m p , predict the current model m according to the prediction network model obtained by training the one-dimensional convolutional neural network c The forward electromagnetic response of , and the average fitting error m of k consecutive accepted models is calculated avg If m p If it is not accepted, the current model remains unchanged;

[0020] Step S35: If the average fitting error m avg When the set preheating period end condition is reached, the current stage of cyclic sampling is ended and step S4 is executed.

[0021] In order to better implement the present invention, further, step S6 specifically includes the following steps:

[0022] Step S61: According to the probability of the basic state being executed during the sampling process, a basic state is randomly selected for execution, and a suggested model m is generated according to the model suggestion distribution. p ';

[0023] Step S62: Calculate the acceptance probability and determine the recommended model m based on the acceptance probability p 'Whether it is accepted, if the proposed model m p 'If accepted, update the current model m c ' is the recommended model m p ', predict the current model m based on the prediction network model obtained by training the one-dimensional convolutional neural network c ' and calculate the fitting error m of the currently accepted model c , according to mc The change of s2 determines whether the model sampling step s2 needs to be adjusted. If the sampling step s2 needs to be adjusted, the proportion p of the current sampling model fitting error reduction is calculated, and a new sampling step is randomly generated from the one-dimensional normal distribution N(s2, p×s2). If the sampling step s2 does not need to be adjusted, the sampling step s2 remains unchanged. If the recommended model m p 'If it is not accepted, keep the current model unchanged;

[0024] Step S63: Determine whether the current Markov chain sampling meets the convergence condition. If the convergence condition is met, stop sampling and output the model sampling result.

[0025] In order to better implement the present invention, further, the input of the one-dimensional convolutional neural network includes a one-dimensional matrix vector of length m composed of layered geoelectric model parameters, that is, m = [σ1, σ2, ..., t1, t2, ...] T , where σ1 and σ2 represent the conductivity of each layer, t1 and t2 represent the thickness of each layer, and a certain number of layered geoelectric models are generated by random generation. The training data set and test data set for network training are constructed according to the resistivity and thickness parameter combinations of these geoelectric models.

[0026] The output of the one-dimensional convolutional neural network includes a one-dimensional matrix vector H=[H1, H2, ..., H] with a length of n formed according to the time domain airborne electromagnetic response. n ] T .

[0027] In order to better implement the present invention, further, the variable-dimensional Bayesian inversion parameter range in step S1 includes the maximum and minimum values ​​of the depth inversion parameters of the layered model interface, the maximum and minimum values ​​of the conductivity inversion parameters, the maximum and minimum values ​​of the number of layers inversion parameters, the maximum and minimum values ​​of the transmitting coil height and the data error inversion parameters.

[0028] Based on the above-mentioned fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data, in order to better realize the present invention, further, a fast variable-dimensional Bayesian inversion system for time-domain airborne electromagnetic data is proposed, which is used to execute the above-mentioned fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data; it includes an initialization unit, an overall update unit, a component update, and an output unit;

[0029] The initialization unit is used to initialize the variable-dimensional Bayesian inversion parameter range, the warm-up period model step s1 and the warm-up period end condition according to the input time-domain airborne electromagnetic data to be inverted;

[0030] The overall update unit is used to set the number of warm-up period Markov chains to 1, construct the initial state m0 of the warm-up period Markov chain according to the conductivity-depth imaging CDI results of the time domain airborne electromagnetic data, and set the current sampling model of the first sampling to m c =m0; adopt the overall update method to perform cyclic sampling within the set model parameter space;

[0031] The component updating unit is used to initialize the model step size s2 after the warm-up period ends, set the number of Markov chains after the warm-up period ends to N, N>1; construct the initial state m of N Markov chains t , set the current sampling model of the first sampling of the t-th Markov chain to m c '=m t , execute N Markov chains in parallel; use the component update method to perform cyclic sampling within the set model parameter space;

[0032] The output unit is used to statistically analyze the model sampling results, analyze the uncertainty of the inversion parameters according to the statistical results, and estimate the inversion results of the time domain airborne electromagnetic data.

[0033] Based on the above-mentioned fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data, in order to better realize the present invention, further, an electronic device is proposed, characterized in that it includes a memory and a processor; a computer program is stored on the memory; when the computer program is executed on the processor, the above-mentioned fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data is realized.

[0034] Based on the above-mentioned fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data, in order to better realize the present invention, a computer-readable storage medium is further proposed, characterized in that computer instructions are stored on the computer-readable storage medium; when the computer program is executed on the above-mentioned electronic device, the above-mentioned fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data is realized.

[0035] The present invention has the following beneficial effects:

[0036] (1) The present invention introduces deep learning technology and uses a one-dimensional convolutional neural network to construct a mapping relationship between geoelectric model parameters and time-domain electromagnetic response data based on the deep learning network, thereby realizing rapid prediction of time-domain airborne electromagnetic response data in the process of variable-dimensional Bayesian inversion, and solving the problem of large amount of calculation caused by using conventional forward simulation methods to calculate time-domain airborne electromagnetic response data.

[0037] (2) During parallel computing, the present invention only runs one Markov chain during the warm-up period, and executes N Markov chains in parallel after the warm-up period. This parallel strategy can avoid executing N Markov chains in parallel from the beginning of the warm-up period, and the problem of poor initial state leading to non-convergence of Markov chain sampling and waste of computing time. The non-convergence of Markov chain sampling can be discovered in time during the sampling process of the warm-up period, thereby stopping the next step of operation and avoiding erroneous calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The present invention provides a schematic diagram of the overall process of the fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data.

[0039] Figure 2 A schematic diagram of the one-dimensional convolutional neural network structure provided by the present invention. DETAILED DESCRIPTION

[0040] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be understood that the described embodiments are only part of the embodiments of the present invention, not all of the embodiments, and therefore should not be regarded as limiting the scope of protection. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technical personnel in this field without making creative work are within the scope of protection of the present invention.

[0041] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "disposed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be an indirect connection through an intermediate medium, or it can be the internal communication of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0042] The professional terms and explanations involved in the embodiments of this application are as follows:

[0043] Warm-up period: Samples collected before the Markov chain reaches its steady-state distribution do not come from the posterior probability distribution of the model parameters and are usually discarded. This initial stage is called the "warm-up" period.

[0044] Embodiment 1:

[0045] This embodiment proposes a fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data, which specifically includes the following steps:

[0046] Step S1: Initialize the variable-dimensional Bayesian inversion parameter range, the warm-up period model step size s1 and the warm-up period end condition according to the input time-domain airborne electromagnetic data to be inverted.

[0047] The variable-dimensional Bayesian inversion parameter range in step S1 includes the maximum and minimum values ​​of the depth inversion parameter of the layered model interface, the maximum and minimum values ​​of the conductivity inversion parameter, the maximum and minimum values ​​of the layer number inversion parameter, the maximum and minimum values ​​of the transmitting coil height and the maximum and minimum values ​​of the data error inversion parameter.

[0048] Step S2: Set the number of warm-up period Markov chains to 1, construct the initial state m0 of the warm-up period Markov chain based on the conductivity-depth imaging CDI results of the time-domain airborne electromagnetic data, and set the current sampling model for the first sampling to m c =m0.

[0049] Step S3: adopt the overall update method to perform cyclic sampling within the set model parameter space.

[0050] Furthermore, the step S3 specifically includes the following steps:

[0051] Step S31: Randomly generate a suggested model m according to the model suggestion distribution p ;

[0052] Step S32: Calculate the likelihood function according to the acquired electromagnetic response, the forward modeled electromagnetic response, and the covariance matrix of the observed data;

[0053] Step S33: Calculate the acceptance probability according to the likelihood function, the prior probability distribution ratio, and the proposed distribution ratio;

[0054] Step S34: Determine the recommended model m based on the acceptance probability p Is it accepted? If the proposed model m p If accepted, update the current model m c is m p , predict the current model m according to the prediction network model obtained by training the one-dimensional convolutional neural network c The forward electromagnetic response of , and the average fitting error m of k consecutive accepted models is calculated avg If m p If it is not accepted, the current model remains unchanged;

[0055] Step S35: If the average fitting error m avg When the set preheating period end condition is reached, the current stage of cyclic sampling is ended and step S4 is executed.

[0056] Step S4: Initialize the model step size s2 after the warm-up period, and set the number of Markov chains after the warm-up period to N, where N>1.

[0057] Step S5: Construct the initial state m of N Markov chains t , set the current sampling model of the first sampling of the t-th Markov chain to m c '=m t , execute N Markov chains in parallel.

[0058] Step S6: Using the component updating method to perform cyclic sampling within the set model parameter space range.

[0059] Furthermore, the step S6 specifically includes the following steps:

[0060] Step S61: According to the probability of the basic state being executed during the sampling process, a basic state is randomly selected for execution, and a suggested model m is generated according to the model suggestion distribution. p ';

[0061] Step S62: Calculate the acceptance probability and determine the recommended model m based on the acceptance probability p 'Whether it is accepted, if the proposed model m p 'If accepted, update the current model m c ' is the recommended model m p ', predict the current model m based on the prediction network model obtained by training the one-dimensional convolutional neural network c ' and calculate the fitting error m of the currently accepted model c , according to m c The change of s2 determines whether the model sampling step s2 needs to be adjusted. If the sampling step s2 needs to be adjusted, the proportion p of the current sampling model fitting error reduction is calculated, and a new sampling step is randomly generated from the one-dimensional normal distribution N(s2, p×s2). If the sampling step s2 does not need to be adjusted, the sampling step s2 remains unchanged. If the recommended model m p 'If it is not accepted, keep the current model unchanged;

[0062] Step S63: Determine whether the current Markov chain sampling meets the convergence condition. If the convergence condition is met, stop sampling and output the model sampling result.

[0063] Step S7: Statistically analyze the model sampling results, analyze the uncertainty of the inversion parameters based on the statistical results, and estimate the inversion results of the time domain airborne electromagnetic data.

[0064] Furthermore, the input of the one-dimensional convolutional neural network includes a one-dimensional matrix vector of length m composed of layered geoelectric model parameters, that is, m = [σ1, σ2, ..., t1, t2, ...] T, where σ1 and σ2 represent the conductivity of each layer, t1 and t2 represent the thickness of each layer, and a certain number of layered geoelectric models are generated by random generation. The training data set and test data set for network training are constructed according to the resistivity and thickness parameter combinations of these geoelectric models.

[0065] The output of the one-dimensional convolutional neural network includes a one-dimensional matrix vector H=[H1, H2, ..., H] with a length of n formed according to the time domain airborne electromagnetic response. n ] T .

[0066] Working principle: This embodiment introduces deep learning technology and uses a one-dimensional convolutional neural network to construct a mapping relationship between geoelectric model parameters and time-domain electromagnetic response data based on the deep learning network, so as to realize the rapid prediction of time-domain airborne electromagnetic response data in the process of variable-dimensional Bayesian inversion; secondly, in parallel calculation, only one Markov chain is run in the warm-up period, and N Markov chains are executed in parallel after the warm-up period. This parallel strategy can avoid the problem of executing N Markov chains in parallel from the beginning of the warm-up period, and the initial state is poor, resulting in the non-convergence of Markov chain sampling and waste of calculation time. The non-convergence of Markov chain sampling can be discovered in time during the sampling process of the warm-up period, thereby stopping the next step of calculation and avoiding erroneous calculation.

[0067] Embodiment 2:

[0068] This embodiment is based on the above embodiment 1. Figure 1 , Figure 2 As shown, a specific embodiment is described in detail.

[0069] Step S1: input the time domain airborne electromagnetic data to be inverted; initialize the variable dimensional Bayesian inversion parameter range, including the maximum and minimum values ​​of inversion parameters such as the depth of the layered model interface, conductivity, number of layers, height of the transmitting coil and data error; initialize the variable dimensional Bayesian inversion parameter range, including the maximum and minimum values ​​of inversion parameters such as the depth of the layered model interface, conductivity, number of layers, height of the transmitting coil and data error.

[0070] Step S2: Initialize the warm-up period model sampling step s1 and the warm-up period end condition. Set the number of warm-up period Markov chains to 1, and use the conductivity-depth imaging (CDI) results of time-domain airborne electromagnetic data to construct the initial state m0 of the warm-up period Markov chain, and set the current sampling model for the first sampling to m c =m0.

[0071] Step S3: adopting the overall update method to perform cyclic sampling in the set model parameter space range, and the implementation of each cyclic sampling includes the following steps:

[0072] First, a proposed model m is randomly generated according to the model proposal distribution p .

[0073] Next, calculate the acceptance probability α, which is defined as:

[0074]

[0075] in, is the prior probability distribution ratio; is the recommended distribution ratio; is the likelihood function ratio, and the likelihood function P(d|m) is defined as:

[0076]

[0077] Where d is the observed electromagnetic response, F(m) is the forward electromagnetic response corresponding to the model parameter m, and C d is the covariance matrix of the observed data, that is, a diagonal matrix composed of the estimated errors of the electromagnetic response data of each time channel.

[0078] Then the recommended model m is determined according to the probability α p Is it accepted? If m p If accepted, update the current model to m p (i.e. m c =m p ), and at the same time, the network structure obtained by training a one-dimensional convolutional neural network is used to predict the current model m c The forward electromagnetic response of , and the average fitting error m of k consecutive accepted models is calculated avg ; if m p If it is not accepted, the current model remains unchanged.

[0079] When the overall update method is used for model sampling, the average fitting error m of the consecutive k accepted models is used. avg As the condition for judging the end of the warm-up period. Average fitting error m avg The definition form is:

[0080]

[0081] in, is the electromagnetic response of the jth time channel in the observed data, is the current sampling model m c The electromagnetic response of the jth time channel in the forward electromagnetic response.

[0082] Step S4: Initialize the model sampling step s2 after the warm-up period, and the probability p1,…,p5 of the five basic states being executed during the sampling process.

[0083] Set the number of Markov chains N (N>1) after the warm-up period and the maximum number of sampling times n for each Markov chain model max ;

[0084] Step S5: Randomly perturb p0% based on the parameters of the last accepted model at the end of the warm-up period to construct the initial state m of the N Markov chains t , where t is the serial number of the corresponding Markov chain, and the current sampling model of the first sampling of each Markov chain is set to m c '=m t .

[0085] Step S6: Execute the model sampling process of N Markov chains in a parallel computing manner, and perform cyclic sampling in the set model parameter space range using the component update method. The implementation process of each cyclic sampling of each Markov chain includes the following steps:

[0086] First, a basic state is randomly selected according to the probability p1, ..., p5, and a suggestion model m is generated according to the model suggestion distribution. p '.

[0087] Then calculate the acceptance probability α and judge the recommended model m based on α p 'Whether it is accepted. If m p 'If accepted, update the current model to m p 'That is m c '=m p ', and at the same time use the network structure obtained by training the one-dimensional convolutional neural network to predict the current model m c ' and calculate the fitting error m of the currently accepted model c , according to m c The change of s2 determines whether the model sampling step size s2 needs to be adjusted. If s2 needs to be adjusted, the proportion p of the current sampling model fitting error reduction is calculated, and a new sampling step size is randomly generated from the one-dimensional normal distribution N(s2, p×s2). If s2 does not need to be adjusted, the sampling step size s2 remains unchanged; if m p 'If not accepted, keep the current model unchanged.

[0088] Finally, determine whether the current Markov chain sampling meets the convergence conditions. If the convergence conditions are met, stop sampling and output the model sampling results.

[0089] Step S7: Perform statistical analysis on the model sampling results, analyze the uncertainty of the inversion parameters, and estimate the inversion results of the time domain airborne electromagnetic data.

[0090] Working principle: This embodiment is different from the conventional Bayesian inversion in which N Markov chains are run in parallel from the beginning of the warm-up period. This method sets the number of Markov chains in the warm-up period to 1, and sets the number of Markov chains after the warm-up period to N, and N>1. The parallel computing method is used to execute N Markov chains in parallel after the warm-up period. If N Markov chains are run in parallel for sampling during the warm-up period, when the initial state is not set well, or the model sampling deviates far from the target parameter space, the sampling of some chains will not converge, and the sampling time of the warm-up period may also be too long. These situations will affect the overall execution efficiency of the Bayesian algorithm. The setting strategy of the number of Markov chains can promptly detect the situation where the Markov chain sampling does not converge, avoid invalid sampling, and improve the overall efficiency of the Bayesian inversion calculation.

[0091] This embodiment uses a network structure obtained by training a one-dimensional convolutional neural network to predict the current model m c The forward electromagnetic response of the CNN is as follows: Figure 2 The input of the deep learning network structure is a one-dimensional matrix vector of length m composed of the parameters of the layered geoelectric model, including the conductivity (σ) and thickness (t) of each layer, that is, m = [σ1, σ2, …, t1, t2, …] T A certain number of layered geoelectric models, for example, 1 million, are generated by random generation, and the conductivity and thickness parameters of these geoelectric models are combined to construct training and test data sets for network training. The output of the deep learning network structure is a one-dimensional matrix vector H = [H1, H2, ..., H n ] T , obtained by simulating and calculating the randomly generated geoelectric model through conventional forward simulation method of time domain airborne electromagnetic method. The internal structure of the network consists of multiple convolution and pooling layers in the left and right parts. The number of filters in the next layer of the convolution part on the left is twice that of the previous layer, and the number of filters in the next layer of the convolution part on the right is half of the previous layer. The specific number of convolution and pooling layers can be adjusted according to the size of time input and output, and the activation function uses Leaky ReLU function. The root mean square error is used to evaluate the performance of the network structure training, and its calculation formula is:

[0092]

[0093] Where M is the number of samples, and The electromagnetic responses predicted by the deep learning network structure and calculated by the conventional forward modeling method, respectively.

[0094] The significant feature of the one-dimensional convolutional neural network deep learning method is that the training calculation of the network structure takes a long time, but once the network structure is trained, it can achieve rapid prediction of new input data. Usually, the training of the network structure is completed in advance, so the calculation of the training part does not affect the prediction in actual application, which can greatly improve the prediction efficiency in actual application.

[0095] Through Figure 2 The network structure shown in the figure can be trained and optimized to construct an implicit mapping relationship H=f(M) between the parameters of the layered geoelectric model and the time-domain airborne electromagnetic response data. Once the mapping relationship between the two is established, the rapid prediction of the time-domain airborne electromagnetic response data in the variable-dimensional Bayesian inversion process can be achieved. For example, it takes about 400 seconds to calculate the time-domain airborne electromagnetic response of 100,000 layered geoelectric models using conventional forward simulation methods on an ordinary personal PC, while it only takes about 30 seconds to predict the same amount of electromagnetic response data using a one-dimensional convolutional neural network method. Variable-dimensional Bayesian inversion usually requires 500,000 to 1 million samples per Markov chain per measuring point, which means that the electromagnetic response of the geoelectric model needs to be calculated 50 to 100 times, and a time-domain airborne electromagnetic exploration area usually has 10 4 -10 7 The data of each observation point needs to be inverted. In short, the conventional variable-dimensional Bayesian inversion of time-domain airborne electromagnetic data will result in a very large amount of overall calculation due to the large number of forward simulation calculations during the sampling process. Therefore, using a one-dimensional convolutional neural network method instead of the conventional forward simulation method to calculate the time-domain airborne electromagnetic response data can effectively solve the problem of large amount of calculation of variable-dimensional Bayesian inversion.

[0096] The other parts of this embodiment are the same as those of the above-mentioned embodiment 1, and thus will not be described in detail.

[0097] Embodiment 3:

[0098] Based on any one of the above-mentioned embodiments 1-2, this embodiment proposes a fast time-domain airborne electromagnetic data variable-dimensional Bayesian inversion system, which is used to execute the above-mentioned fast time-domain airborne electromagnetic data variable-dimensional Bayesian inversion method; including an initialization unit, an overall update unit, a component update, and an output unit;

[0099] The initialization unit is used to initialize the variable-dimensional Bayesian inversion parameter range, the warm-up period model step s1 and the warm-up period end condition according to the input time-domain airborne electromagnetic data to be inverted;

[0100] The overall update unit is used to set the number of warm-up period Markov chains to 1, construct the initial state m0 of the warm-up period Markov chain according to the conductivity-depth imaging CDI results of the time domain airborne electromagnetic data, and set the current sampling model of the first sampling to m c =m0; adopt the overall update method to perform cyclic sampling within the set model parameter space;

[0101] The component updating unit is used to initialize the model step size s2 after the warm-up period ends, set the number of Markov chains after the warm-up period ends to N, N>1; construct the initial state m of N Markov chains t , set the current sampling model of the first sampling of the t-th Markov chain to m c '=m t , execute N Markov chains in parallel; use the component update method to perform cyclic sampling within the set model parameter space;

[0102] The output unit is used to statistically analyze the model sampling results, analyze the uncertainty of the inversion parameters according to the statistical results, and estimate the inversion results of the time domain airborne electromagnetic data.

[0103] This embodiment further proposes an electronic device, characterized in that it includes a memory and a processor; a computer program is stored in the memory; when the computer program is executed on the processor, the above-mentioned fast variable-dimensional Bayesian inversion method of time-domain airborne electromagnetic data is implemented.

[0104] This embodiment further proposes a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions; when the computer program is executed on the above-mentioned electronic device, the above-mentioned fast variable-dimensional Bayesian inversion method of time-domain airborne electromagnetic data is implemented.

[0105] The other parts of this embodiment are the same as any one of the above-mentioned embodiments 1-2, so they will not be repeated here.

[0106] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. Any simple modification or equivalent change made to the above embodiment based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data, characterized in that: The specific steps include: Step S1: Initialize the variable-dimensional Bayesian inversion parameter range, the warm-up period model step length s1 and the warm-up period end condition according to the input time-domain airborne electromagnetic data to be inverted; Step S2: Set the number of warm-up period Markov chains to 1, construct the initial state m0 of the warm-up period Markov chain based on the conductivity-depth imaging CDI results of the time-domain airborne electromagnetic data, and set the current sampling model for the first sampling to m c =m0; Step S3: adopting the overall update method to perform cyclic sampling within the set model parameter space range; Step S4: Initialize the model step size s2 after the warm-up period, and set the number of Markov chains after the warm-up period to N, where N>1; Step S5: Construct the initial state m of N Markov chains t , set the current sampling model of the first sampling of the t-th Markov chain to m c '=m t , execute N Markov chains in parallel; Step S6: using a component updating method to perform cyclic sampling within the set model parameter space range; Step S7: Statistically analyze the model sampling results, analyze the uncertainty of the inversion parameters based on the statistical results, and estimate the inversion results of the time domain airborne electromagnetic data.

2. A fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data according to claim 1, characterized in that: The step S3 specifically comprises the following steps: Step S31: Randomly generate a suggested model m according to the model suggestion distribution p ; Step S32: Calculate the likelihood function according to the acquired electromagnetic response, the forward modeled electromagnetic response, and the covariance matrix of the observed data; Step S33: Calculate the acceptance probability according to the likelihood function, the prior probability distribution ratio, and the proposed distribution ratio; Step S34: Determine the recommended model m based on the acceptance probability p Is it accepted? If the proposed model m p If accepted, update the current model m c is m p , predict the current model m based on the prediction network model obtained by training the one-dimensional convolutional neural network c The forward electromagnetic response of the model is calculated, and the average fitting error m of the k consecutive accepted models is calculated. avg If m p If it is not accepted, the current model remains unchanged; Step S35: If the average fitting error m avg When the set preheating period end condition is reached, the current stage of cyclic sampling is ended and step S4 is executed.

3. A fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data according to claim 1, characterized in that: The step S6 specifically comprises the following steps: Step S61: According to the probability of the basic state being executed during the sampling process, a basic state is randomly selected for execution, and a suggested model m is generated according to the model suggestion distribution. p '; Step S62: Calculate the acceptance probability and determine the recommended model m based on the acceptance probability p 'Whether it is accepted, if the proposed model m p 'If accepted, update the current model m c ' is the recommended model m p ', predict the current model m based on the prediction network model obtained by training the one-dimensional convolutional neural network c ' and calculate the fitting error m of the currently accepted model c , according to m c The change of s2 determines whether the model sampling step s2 needs to be adjusted. If the sampling step s2 needs to be adjusted, the proportion p of the current sampling model fitting error reduction is calculated, and a new sampling step is randomly generated from the one-dimensional normal distribution N(s2, p×s2). If the sampling step s2 does not need to be adjusted, the sampling step s2 remains unchanged. If the recommended model m p 'If it is not accepted, keep the current model unchanged; Step S63: Determine whether the current Markov chain sampling meets the convergence condition. If the convergence condition is met, stop sampling and output the model sampling result.

4. A fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data according to claim 2 or 3, characterized in that: The input of the one-dimensional convolutional neural network includes a one-dimensional matrix vector of length m composed of layered geoelectric model parameters, that is, m = [σ1, σ2, ..., t1, t2, ...] T , where σ1 and σ2 represent the conductivity of each layer, t1 and t2 represent the thickness of each layer, and a certain number of layered geoelectric models are generated by random generation. Training data sets and test data sets for network training are constructed based on the conductivity and thickness parameter combinations of these geoelectric models. The output of the one-dimensional convolutional neural network includes a one-dimensional matrix vector H=[H1, H2, ..., H] with a length of n formed according to the time domain airborne electromagnetic response. n ] T .

5. The fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data according to claim 1, characterized in that: The variable-dimensional Bayesian inversion parameter range in step S1 includes the maximum and minimum values ​​of the depth inversion parameter of the layered model interface, the maximum and minimum values ​​of the conductivity inversion parameter, the maximum and minimum values ​​of the layer number inversion parameter, the maximum and minimum values ​​of the transmitting coil height and the maximum and minimum values ​​of the data error inversion parameter.

6. A fast time-domain airborne electromagnetic data variable-dimensional Bayesian inversion system, used to execute the fast time-domain airborne electromagnetic data variable-dimensional Bayesian inversion method as claimed in claim 1; characterized in that: It includes an initialization unit, an overall update unit, a component update unit, and an output unit; The initialization unit is used to initialize the variable-dimensional Bayesian inversion parameter range, the warm-up period model step s1 and the warm-up period end condition according to the input time-domain airborne electromagnetic data to be inverted; The overall update unit is used to set the number of warm-up period Markov chains to 1, construct the initial state m0 of the warm-up period Markov chain according to the conductivity-depth imaging CDI results of the time domain airborne electromagnetic data, and set the current sampling model of the first sampling to m c =m0; adopt the overall update method to perform cyclic sampling within the set model parameter space; The component updating unit is used to initialize the model step size s2 after the warm-up period ends, set the number of Markov chains after the warm-up period ends to N, N>1; construct the initial state m of N Markov chains t , set the current sampling model of the first sampling of the t-th Markov chain to m c '=m t , execute N Markov chains in parallel; The component updating method is used to perform cyclic sampling within the set model parameter space; The output unit is used to statistically analyze the model sampling results, analyze the uncertainty of the inversion parameters according to the statistical results, and estimate the inversion results of the time domain airborne electromagnetic data.

7. An electronic device, characterized in that: The device comprises a memory and a processor; the memory stores a computer program; When the computer program is executed on the processor, the fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data according to any one of claims 1 to 5 is implemented.

8. A computer-readable storage medium, characterized in that: The computer readable storage medium has computer instructions stored thereon; When the computer program is executed on the electronic device according to claim 7, the fast variable-dimensional Bayesian inversion method for time-domain airborne electromagnetic data according to any one of claims 1 to 5 is implemented.

Images