A transient electromagnetic inversion method based on improved PSO-FADBN network

Abstract

The invention discloses a transient electromagnetic inversion method based on an improved PSO-FADBN network, comprising the steps of: S1, establishing a layered geoelectric model sample according to equal-interval sampling of resistivity index; S2, reducing the dimension of the layered geoelectric model sample by using a factor analysis algorithm, and using the extracted main signal components as input data of DBN to construct a FADBN network; S3, based on the FADBN network, pre-training the weights and biases in the DBN by using an improved PSO algorithm to construct an improved PSO-FADBN network; S4, training the improved PSO-FADBN network model, and using it to predict the geoelectric structure of an unknown earth. The invention can use TEM observation data to quickly and accurately predict complex underground layered electrical structures.

Description

Technical Field

[0001] The invention relates to the technical field of geophysical transient electromagnetic method, and in particular to a transient electromagnetic inversion method based on an improved PSO-FADBN network. Background Art

[0002] Transient electromagnetic method (TEM) is a method of detecting the resistivity of a medium by using an ungrounded loop or grounded line source to emit a pulse magnetic field underground, and using a coil or grounding electrode to observe the secondary induced eddy current field caused in the underground medium during the interval of the pulse magnetic field. Its basic working method is: a transmitting coil with a certain waveform current is set on the ground or in the air, so as to generate an electromagnetic field in the surrounding space and in the underground conductive rock ore body. After the power is cut off, the induced current decays over time due to heat loss.

[0003] TEM inversion is to infer the underground geoelectric structure based on the acquired time-domain induced magnetic field signal. The traditional gradient-based iterative optimization method takes the underground geoelectric structure as the optimization object to minimize the L2 objective function of the predicted induced magnetic field signal and the acquired signal. This method is highly dependent on the initial model, and the iterative process is prone to fall into extreme values. In order to avoid this problem, nonlinear inversion methods such as particle swarm optimization, genetic algorithm, simulated annealing and artificial neural network have been proposed. Among them, with the formation of machine learning and artificial intelligence related research, the theory and corresponding algorithms related to artificial neural networks have developed most rapidly, not only in transient electromagnetic inversion, but also in other electromagnetic detection problems.

[0004] In order to improve the training efficiency of neural networks and improve the local optimal problem, the Deep Belief Network (DBN) was proposed. The network consists of a multi-layer restricted Boltzmann machine (RBM) and a backpropagation neural network (BP), which is widely used in regression prediction and classification. However, the high dimensionality and ill-posedness of TEM samples are still obstacles to the training efficiency of neural networks. Commonly used dimensionality reduction methods include principal component analysis (PCA), kernel PCA, and dictionary learning. It is assumed that there is multicollinearity between the standardized original variables, that is, there is non-negligible information overlap between the original variables, but PCA cannot effectively eliminate this overlap.

[0005] Pre-training neural network parameters through heuristic algorithms has become a common solution to improve the quality of neural network training. In the evolution process of a single heuristic algorithm, it is difficult to balance the evolution speed and direction of the individual itself and the social group. The corresponding optimization problem is a debate between global optimization and local optimization. It is difficult to effectively maintain the diversity of the population, which makes it easy for the optimization process to converge prematurely and fall into the local optimum. In recent years, memetic algorithms have successfully solved many complex heuristic optimization problems, and their performance is better than that obtained by using global optimization methods alone. Particle swarm optimization (PSO), originated from the study of bird flock foraging behavior, realizes collective information sharing through collaboration between individuals to find the optimal solution. It is a typical heuristic evolutionary algorithm and can be used as a basic global optimization algorithm.

[0006] The initial population of the particle swarm algorithm is randomly generated, which affects the convergence speed of the algorithm and the accuracy of the final solution. The inertia weight w can limit the search range of the particles, which allows the particles to maintain the inertia of movement and search in new areas, which means that this new evolution contains the old evolutionary habits and experience. When the inertia weight is relatively small, the new evolution can get rid of the influence of the previous evolutionary experience, which is conducive to expanding the search field, but it is easy to slow down the convergence speed in the optimization; when the inertia weight is relatively large, the particles maintain the evolutionary direction according to the previous experience. When the evolutionary direction is correct, it will help to speed up the convergence speed of global optimization, but it is easy to make the optimization fall into local extremes. In addition, in the early stages of the optimization process, particles need strong self-cognition and weak social cognition. At this time, particles can traverse as many local extremes as possible in the search space; in the later stages of optimization, particles must have strong social cognition and weak self-awareness to avoid falling into local extremes in optimization. The value of the cognitive attraction coefficient reflects the degree of influence of information exchange on particles. Information exchange includes self-experience information and global optimal information. Setting the learning factor too large or too small is not conducive to particle optimization.

[0007] In order to more effectively apply deep learning technology to the field of geophysical transient electromagnetic inversion imaging, it is necessary to invent a factor analysis deep belief network transient electromagnetic inversion technology that improves PSO, so that complex underground layered electrical structures can be quickly and accurately predicted using TEM observation data. Summary of the invention

[0008] The present invention aims to provide a transient electromagnetic inversion method based on an improved PSO-FADBN network to improve the training efficiency of the TEM inversion neural network, enhance the generalization of the network, improve the problem of insufficient inversion ability of deep neural networks for TEM data, and ultimately improve the prediction accuracy of the TEM inversion neural network.

[0009] In order to achieve the above object, the present invention provides the following technical solutions:

[0010] A transient electromagnetic inversion method based on an improved PSO-FADBN network comprises the following steps:

[0011] S1. Establish layered geoelectric model samples based on equal interval sampling of resistivity index;

[0012] S2, using factor analysis algorithm to reduce the dimension of layered geoelectric model samples, and using the extracted main signal components as DBN input data to build the FADBN network;

[0013] S3. Based on the FADBN network, the improved PSO algorithm is used to pre-train the weights and biases in the DBN to construct an improved PSO-FADBN network;

[0014] S4, train the improved PSO-FADBN network model and use it to predict the geoelectric structure of unknown earth;

[0015] The improved PSO algorithm comprises:

[0016] Use the OBL strategy to generate a random initial population and a counter-population. After merging the two, select particles with smaller fitness to form a new initial population.

[0017] Use sine mapping to adjust the inertia weight ω of the PSO algorithm;

[0018] Adjust the learning factor c of the PSO algorithm using sine-cosine mapping 1 and c 2 ;

[0019] The mutation operator in the genetic algorithm is introduced into the PSO algorithm to randomly perturb the global extreme value.

[0020] Furthermore, the layered geoelectric model sample includes M geoelectric models {m i , i=1,2,…,M}, the geoelectric model contains N resistivity parameters and N-1 layer thickness parameters;

[0021] The vertical component B of the vertical induced magnetic field generated by the step current in the frequency domain z for:

[0022]

[0023] The magnetic field response in the time domain is obtained by the following Fourier transform:

[0024]

[0025] B z The derivative with respect to time is expressed as:

[0026]

[0027] The time domain response of the above integral is obtained through the second derivative property of the function as follows:

[0028]

[0029] Where I is the excitation current, a is the coil radius, h is the vertical height of the transmitting coil, z is the height of the receiving coil from the ground, and J 1 (*) is the first-order Bessel function, λ is the wavelength, μ 0 is the free space permeability, r TE is the reflection coefficient and ω is the angular frequency.

[0030] Further, the inertia weight ω of the PSO algorithm is adjusted using the sine mapping, which is expressed as:

[0031]

[0032] Here, q ranges from 0 to 4.

[0033] Further, the sine-cosine mapping is used to adjust the learning factor c of the PSO algorithm. 1 and c 2 , expressed as:

[0034]

[0035]

[0036] Here, the constants α and δ are 2 and 0.5 respectively.

[0037] Furthermore, the mutation operator in the genetic algorithm is introduced into the PSO algorithm to randomly perturb the global extreme value, and the particle change is expressed as:

[0038] x id (t) = r × (U x -L x ) / n+(U x -L x ) / 2 (22)

[0039] Among them, r is uniformly distributed in the range of [-1,1], U x and L x are the upper and lower limits of a given position, n is 4, the mutation condition is random mutation, and the mutation probability is 10%.

[0040] The beneficial effects of the present invention are:

[0041] The present invention constructs an improved PSO-FADBN network, uses factor analysis to reduce the dimension of training samples, and takes the extracted main signal components as the input data of the DBN to perform regression fitting of the transient electromagnetic inverse problem by the DBN, thereby improving the training efficiency; uses the method of Opposition-based learning (OBL) to replace the random initial population position as the new initial population strategy, effectively increasing the diversity of the initial population and increasing the chance of reaching the global optimal solution; uses a sine mapping to adjust the inertia weight ω of the PSO algorithm, which can not only enhance the population diversity in the search process, but also enhance the ability to converge to the global optimum; to enhance the ability to jump out of local extrema, a mutation operator in the genetic algorithm is introduced into the PSO algorithm to randomly perturb the global extreme value, which can expand the search space of the particles themselves, enhance the diversity of the population, and further increase the possibility of finding the optimal solution. The present invention enhances the inversion ability of the deep belief neural network, improves the accuracy of inversion imaging, can be widely applied to the field of transient electromagnetic fast imaging, and has good practical value and application prospects for quickly and accurately predicting the underground electrical structure. Description of the Drawings

[0042] Figure 1 It is a flowchart of a transient electromagnetic inversion method based on an improved PSO-FADBN network provided by an embodiment of the present invention;

[0043] Figure 2 It is a Hankel transformation coefficient table;

[0044] Figure 3 It is a nine-layer geoelectric model predicted by different DBN networks and its error table;

[0045] Figure 4 It is an FADBN network architecture diagram;

[0046] Figure 5 It is an improved PSO flowchart;

[0047] Figure 6 It is a nine-layer geoelectric model and its TEM response schematic diagram; where, (a) represents the geoelectric model, and (b) represents the –dB z / dt fitting curve. Detailed Embodiments

[0048] The present invention will be further described in detail below in conjunction with the drawings and embodiments:

[0049] As Figure 1 shown, a transient electromagnetic inversion method based on an improved PSO-FADBN network includes: (1) Sample establishment

[0050] Establish a layered geoelectric model sample. Determine the number of layers N in the geoelectric model, then the geoelectric model contains N resistivity parameters and N-1 layer thickness parameters, a total of 2N-1 parameters, and construct M geoelectric models {m i ,i=1,2,…,M}.

[0051] In the layered geoelectric model, the initial field excited by the horizontal coil is incident vertically from the surface to the underground. Through propagation in the medium, transmission and reflection at the electrical interface, a secondary field signal is generated on the surface. In order to reduce the influence of lateral inhomogeneity, we often measure the vertical component of the magnetic field at the center of the coil. The vertical magnetic field signal generated by the step current can be expressed as:

[0052]

[0053] Where I is the excitation current, a is the coil radius, h is the vertical height of the transmitting coil, z is the height of the receiving coil from the ground, and J 1 (*) is a first-order Bessel function. TE is the reflection coefficient, which can be calculated from the mutual impedance ^γ and the self-impedance γ:

[0054]

[0055] For the i-th layer, the self-impedance can be expressed as:

[0056] γ i =u i / iωμ 0 (3)

[0057] The mutual impedance can be expressed as:

[0058]

[0059] The conductivity of layer i is σ i , layer thickness is h i , electrical coefficient u i =(λ 2 +iωμ 0 σ i ) 1 / 2 , under quasi-static conditions, wavelength λ=u 0 , μ 0 is the free space permeability, ω is the angular frequency. i=0 represents the air medium.

[0060] Formula (1) can obtain the vertical component of the vertical induced magnetic field in the frequency domain, and the magnetic field response in the time domain can be obtained by the following Fourier transform:

[0061]

[0062] The real part of the magnetic field and the cosine function are even functions, and the imaginary part and the sine function are odd functions. Therefore, in the above expression of the time domain field, the real part is an even function, the imaginary part is an odd function, and the integral of the imaginary part is zero. After simplification, when t>0, using the symmetry of the even function, the derivative of the magnetic field with respect to time can be obtained:

[0063]

[0064] The discrete calculation of the above integral can be calculated by the broken line approximation method, and the time domain response can be obtained by the second-order derivative property of the function, as shown below:

[0065]

[0066] Example:

[0067] Constructing a resistivity data set D based on an exponentially equally spaced series rho ={10 0:0.2:4}, layer thickness dataset D h ={10 0:0.2:3 For a five-layer geoelectric model, N can be set to 5. Each electrical model has 5 resistivity parameters and 4 layer thickness parameters, a total of 9 parameters.

[0068] Traverse the resistivity data set D rho and layer thickness dataset D h All parameters in the layered model, the total number of samples M = 20 5 ×15 4 .

[0069] The frequency domain response corresponding to each geoelectric model is calculated according to formula (1), which can be simplified as follows:

[0070]

[0071] Among them, the first-order Bessel function J 1 The discrete calculation of (*) is calculated by Hankel transformation. The above form can be changed to:

[0072]

[0073] Among them, λ i is the sampling point, W i is the weight, i=1,2,…,n, and n=140 discrete sampling points. The weight coefficient is as follows Figure 2 As shown, the sampling point is calculated by the following formula:

[0074]

[0075] Among them, a=-7.91001919, s=8.79671439570e-02.

[0076] The magnetic field response in the time domain can be obtained by Fourier transform. In the discrete form of formula (6), the discrete frequency sequence is ω = 10 -5:5.E-04:15 .

[0077] (2) Construction of FADBN

[0078] The specific structure of FADBN is as follows Figure 4 TEM induced magnetic signal - dB z / dt is used as input data x, and the main components are extracted through factor analysis. After dimensionality reduction, they can be built into the feature space. The feature space is input into a DBN consisting of multiple RBM layers and a BP layer. The BP output data y constitutes a geoelectric model, including resistivity values ​​and layer thickness values. The relationship between the output y and input x of FADBN is x=F[y], and F[*] is the TEM forward operator, which simulates the time domain Maxwell problem. The entire FADBN can be regarded as a regression model that fits the TEM inverse problem.

[0079] For a Ψ=σ 2 The factor model of I, the columns of W are the scaled eigenvectors of the sample covariance, σ 2 is the average of the discarded eigenvalues. When the eigenvectors that make up W are selected as the main eigenvectors, the log-likelihood function takes the maximum value. Considering that the solution of μ is the average value of the data set, the log-likelihood function can be expressed as:

[0080]

[0081] Where S is the covariance matrix of x and Tr is the trace of the matrix. The expected value of the posterior probability distribution of the latent variable is:

[0082]

[0083] The correlation matrix W and covariance matrix Ψ are updated as:

[0084]

[0085] and

[0086]

[0087] Example:

[0088] The pseudo code flow of factor analysis is as follows:

[0089]

[0090] (3) Establishing an improved PSO algorithm

[0091] The improved PSO algorithm is an optimization algorithm that uses the information sharing strategy between individuals and groups to improve the PSO optimization ability. The improved strategies in this invention include: 1) adversarial learning strategy, 2) dynamic inertia weight, and 3) evolution acceleration. Combined with the basic framework of PSO, such as Figure 5 As shown, including:

[0092] 1) Basic framework of PSO

[0093] The PSO algorithm will make the population evolve more intelligently after each iteration and can accumulate search knowledge, which is called an evolutionary algorithm. The PSO algorithm does not use the survival of the fittest, but uses the mechanism of each individual in the population to compete with each other to generate the global optimal solution. It generates the best solution through information sharing and the cooperation mechanism between each individual in the group.

[0094] Assume that PSO consists of multiple particles. In the D-dimensional search space, the particle swarm contains n particles. The position of the nth particle in the D-dimensional space is defined as x i :

[0095] x i =(x i1 ,x i2 ,L,x iD ),i=1,2,L,n (12)

[0096] Assume that particle x i The current speed and its single best historical position are v i and p i :

[0097] For particle xi, we can update the position x in dimension d according to the velocity id :

[0098] v i =(v i1 ,v i2 ,L,v iD ) (13)

[0099] p i =(p i1 ,p i2 ,L,p iD ) (14)

[0100] Then, for the entire particle swarm, the global optimal position is P g :

[0101] p g =(p g1 ,p g2 ,L,p gD ) (15)

[0102] At time t, particle xi The speed update formula of the d-dimensional is:

[0103] v id (t+1)=ωv id (t)+c 1 r 1 (p id (t)-x id (t))+c 2 r 2 (p gi (t)-x id (t))(16)

[0104] Where w is the inertial weight, which belongs to [0, 1]. c1 and c2 are acceleration coefficients, also known as learning factors. r1 and r2 are random coefficients, both belong to [0, 1], which determine the motion of semi-random particles affected by the single and global optimal solutions.

[0105] The particle velocity update mainly consists of three parts: the current initial velocity part, the self-motion trajectory part, and the self-trajectory correction part. The influence of the current velocity on the particle update velocity can be adjusted by the inertia weight. The influence of the particle's own trajectory on the particle update velocity can be adjusted by the acceleration coefficient and the random coefficient. When the own trajectory is inaccurate, it is necessary to use global optimization to correct the trajectory.

[0106] For particle x i , we can update the position x in dimension d according to the velocity id :

[0107] x id (t+1)=x id (t)+v id (t+1) (17)

[0108] 2) Oppositional learning strategy

[0109] We use the OBL strategy to generate the initial population, which includes two types of populations: random initial population and anti-population. id ,d=1,2,..,D} is randomly generated according to the form shown in formula (12). Assume that the inverse population is {x' id ,d=1,2,..,D}, then x' id It can be expressed as:

[0110] x'id=xmax,d+xmin,d-xid (18)

[0111] Among them, x max,d and x min,d , is the particle x in the d-dimensional search space iAfter merging the random initial population and the anti-population, we select the particle with the smallest fitness to form a new initial population.

[0112] 3) Dynamic inertia weight

[0113] Since the slope of the inertia weight change is constant and the speed change is always kept at the same level, if the initial iteration does not produce a better point, the accumulation of iterations and the rapid decay of speed may cause the final result to be a local optimal value. Therefore, we use a nonlinear strategy sine mapping with ergodicity, non-repetitiveness and irregularity to adjust the inertia weight ω of PSO. This strategy can not only enhance the population diversity during the search process, but also enhance the ability to converge to the global optimum. The dynamic inertia weight based on sine mapping can be expressed as:

[0114]

[0115] Here, q ranges from 0 to 4.

[0116] 3) Accelerated Evolution Strategy

[0117] The acceleration coefficients c1 and c2 are the cognitive attraction coefficients during the optimization process, which are controlled by the learning situation. During the optimization process, the enhancement effect of sine cosine mapping on population diversity and convergence can be used to improve this linear asynchronous strategy. We use sine cosine acceleration coefficients (SCAC) to adjust the balance between individual cognition and social cognition. The cognitive attraction coefficient can be expressed as:

[0118]

[0119]

[0120] Here, the constants α and δ are 2 and 0.5 respectively.

[0121] 4) Population variation

[0122] In order to enhance the ability to jump out of local extreme values, we introduced the mutation operator in the genetic algorithm into PSO, which can expand the search space of the particles themselves, enhance the diversity of the population, and further improve the possibility of finding the optimal solution. After each evolution, the particles will be reset with a certain probability. When the mutation condition is met, the mutation jumps out of the current position, otherwise the original position remains unchanged. The particle change can be expressed as:

[0123] x id (t) = r × (U x -L x ) / n+(U x -Lx ) / 2 (22)

[0124] Among them, r is uniformly distributed in the range of [-1,1], U x and L x are the upper and lower limits of a given position, and n is 4. The mutation condition is random mutation, and the mutation probability is 10%.

[0125] 5) Fitness

[0126] The fitness in the optimization process refers to the objective function setting in TEM inversion. The L2 norm is used to define the mismatch between the MT observed response data and the predicted response data. The fitness can be expressed as:

[0127]

[0128] Among them, the overall fitness is composed of apparent resistivity fitness and phase fitness, and their weight coefficients are c rho , c phi Since the apparent resistivity and phase are variables in different units, in order to convert the apparent resistivity and phase fitness into a unified dimension, the response data and the predicted data are normalized.

[0129] (4) Training DBN learning model

[0130] 90% of the learning samples are taken as the training group, and the remaining 10% are taken as the validation group. The training group is input into the learning model for training. During the training, the validation group is used to verify the training results, and the validation results are evaluated by comparing the predicted geoelectric model with the geoelectric model in the validation group.

[0131] For a binary RBM, when a neuron in the hidden layer is activated, the probability P(h k =1|v) can be expressed as

[0132]

[0133] Combined with the joint probability, the activity probability of the kth neuron can be expressed as

[0134]

[0135] After simplifying the above formula, the formula of the sigmoid function is as follows:

[0136]

[0137] where h -k is the hidden layer state h,h -k =(h 1 ,h 2 ,...,hk-1 ,h k+1 ,...,h Nk ) k Similarly, the probability P(v k =1|h) can be obtained by the following formula

[0138]

[0139] We use an iterative gradient ascent method to maximize the target log-likelihood function. The log-likelihood function is the probability distribution P of the observed data. θ (v). After logarithmic transformation, the function can be expressed as:

[0140]

[0141] The iterative process of weights and bias coefficients is:

[0142]

[0143] Where η is the learning rate, which controls the step size of each iteration. The derivative of the log-likelihood function with respect to the RBM parameters during training can be expressed as:

[0144]

[0145] In order to reduce the computational complexity of this method, the k-step contrastive divergence (CD-k) algorithm is used to approximate the gradient as:

[0146]

[0147] Among them, h of the t-th iteration step is calculated in sequence (t-1) ~P(h (t-1) |v (t-1) ), v t ~P(v t |h (t-1) ), h t ~P(h t |v t )(t=1:k). “~” indicates the process of sampling from the calculated probability distribution P.

[0148] Example:

[0149] The training process of DBN training is as follows:

[0150]

[0151]

[0152] (5) Example Analysis

[0153] Assuming a nine-layer TEM geoelectric model ( Figure 6 a), the resistivity of each layer is 200Ω·m, 50Ω·m, 500Ω·m, 100Ω·m, 1000Ω·m, 300Ω·m, 10Ω·m, 200Ω·m and 500Ω·m respectively; the thickness of each layer is 15m, 15m, 20m, 20m, 10m, 310m, 10m, 20m and 40m respectively. The induced magnetic response of the five-layer geoelectric model ( Figure 6 b) is the hypothetical response signal used to predict the assumed parameters of the geoelectric model.

[0154] The geoelectric models predicted by the four inversion methods were compared. The four models can effectively reflect the assumed five-layer geoelectric model, but the prediction accuracy is different ( Figure 6 a). The geoelectric model of DBN has the largest error starting from the resistivity and thickness of the first layer, and the prediction deviates significantly, with the largest deviation when predicting the thickness of the fourth layer. After pre-optimizing the parameters in DBN using the PSO method, the prediction accuracy of PSO-DBN is slightly improved. The predicted geoelectric structure obtained after adding factor analysis shows that reducing the dimensionality of the signal using factor analysis can significantly improve the prediction ability of DBN, because the compression of the feature dimension improves the training efficiency of DBN. On this basis, the prediction accuracy of DBN can be further improved by improving the optimization efficiency of PSO using a hybrid strategy.

[0155] The corresponding TEM responses of the four predicted geoelectric models are shown in Figure 6 As shown in b, the predicted signal has a high fit with the hypothesized signal. The fit is about 1e -5 The fitting accuracy of the response signal is consistent with the changes in the fitting accuracy of the geoelectric model predicted by different methods. Figure 3 The geoelectric models and errors predicted by the five schemes are shown. From the final absolute percent error (APE), the prediction accuracy of single DBN is the worst (the error is higher than 15%). After adding PSO and factor analysis strategies, the errors of PSO-DBN and PSO-FADBN are about 4.6% and 8% respectively. The improved PSO-FADBN has the highest prediction accuracy, which can control the error within 2.8%.

[0156] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A transient electromagnetic inversion method based on improved PSO-FADBN network, It is characterized in that Includes steps: S1. Establish layered geoelectric model samples based on equal interval sampling of resistivity index; S2, using factor analysis algorithm to reduce the dimension of layered geoelectric model samples, and using the extracted main signal components as DBN input data to build the FADBN network; S3. Based on the FADBN network, the improved PSO algorithm is used to pre-train the weights and biases in the DBN to construct an improved PSO-FADBN network; S4, train the improved PSO-FADBN network model and use it to predict the geoelectric structure of unknown earth; The improved PSO algorithm comprises: Use the OBL strategy to generate a random initial population and a counter-population. After merging the two, select particles with smaller fitness to form a new initial population. Use sine mapping to adjust the inertia weight ω of the PSO algorithm; Adjust the learning factor c of the PSO algorithm using sine-cosine mapping 1 and c 2 ; The mutation operator in the genetic algorithm is introduced into the PSO algorithm to randomly perturb the global extreme value.

2. According to claim 1, a transient electromagnetic inversion method based on an improved PSO-FADBN network, Features: The layered geoelectric model sample includes M geoelectric models {m i , i=1,2,…,M}, the geoelectric model contains N resistivity parameters and N-1 layer thickness parameters; The vertical component B of the vertical induced magnetic field generated by the step current in the frequency domain z for: The magnetic field response in the time domain is obtained by Fourier transform as follows: B z The derivative with respect to time is expressed as: The time domain response of the above integral is obtained through the second derivative property of the function as follows: Where I is the excitation current, a is the coil radius, h is the vertical height of the transmitting coil, z is the height of the receiving coil from the ground, and J 1 (*) is the first-order Bessel function, λ is the wavelength, μ 0 is the free space permeability, r TE is the reflection coefficient and ω is the angular frequency.

3. According to claim 1, a transient electromagnetic inversion method based on an improved PSO-FADBN network, Features: The inertia weight ω of the PSO algorithm is adjusted using a sine mapping, which is expressed as: Here, q ranges from 0 to 4.

4. According to claim 1, a transient electromagnetic inversion method based on an improved PSO-FADBN network, Features: The sine-cosine mapping is used to adjust the learning factor c of the PSO algorithm 1 and c 2 , expressed as: Here, the constants α and δ are 2 and 0.5 respectively.

5. According to claim 1, a transient electromagnetic inversion method based on an improved PSO-FADBN network, Features: The mutation operator in the genetic algorithm is introduced into the PSO algorithm to randomly perturb the global extreme value. The particle change is expressed as: x id (t)=r×(U x -L x ) / n+(U x -L x ) / 2 (22) Among them, r is uniformly distributed in the range of [-1,1], U x and L x are the upper and lower limits of a given position, n is 4, the mutation condition is random mutation, and the mutation probability is 10%.

Images