阅读说明：本技术 一种基于监督下降法的时间域电磁数据目标反演方法 (A kind of time-domain electromagnetic data object inversion method based on supervision descent method ) 是由 张晓娟 谢吴鹏 于 2019-08-02 设计创作，主要内容包括：本发明提供一种基于监督下降法的时间域电磁数据目标反演方法,过程为：训练过程：设置不同的学习样本,获取其对应的二次场响应作为观测矩阵；针对每一学习样本,进行迭代训练,获得整个训练样本的梯度下降矩阵；反演过程：在异常区域附近进行反演目标二次场响应测量,获得反演目标的观测向量；设定反演目标的初始参数,利用整个训练样本的梯度下降矩阵及反演目标的观测向量进行反演,获得反演目标的真实参数。该方法能够准确获得地下目标的目标参数。(The present invention provides a kind of time-domain electromagnetic data object inversion method based on supervision descent method, and process are as follows: training process: being arranged different learning samples, obtains its corresponding secondary field response as observing matrix；For each learning sample, it is iterated training, obtains the gradient decline matrix of entire training sample；Refutation process: carrying out inverting target secondary field response measurement near abnormal area, obtains the observation vector of inverting target；The initial parameter for setting inverting target carries out inverting using the observation vector that the gradient of entire training sample declines matrix and inverting target, obtains the actual parameter of inverting target.This method can accurately obtain the target component of buried target.)

1. a kind of time-domain electromagnetic data object inversion method based on supervision descent method, which is characterized in that detailed process are as follows:

Training process:

Different learning samples is set, obtains its corresponding secondary field response as observing matrix；

For each learning sample, it is iterated training, obtains the gradient decline matrix of entire training sample；

Refutation process:

Inverting target secondary field response measurement is carried out near abnormal area, obtains the observation vector of inverting target；

The initial parameter for setting inverting target declines the observation vector of matrix and inverting target using the gradient of entire training sample Inverting is carried out, the actual parameter of inverting target is obtained.

2. the time-domain electromagnetic data object inversion method according to claim 1 based on supervision descent method, which is characterized in that The process of the repetitive exercise are as follows:

Set the initial parameter m of training sample₀, calculate the forward response d under the initial value₀, enable initial value=0 of i；

Calculate Δ d_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of sample, Δ m_i=m_obs-m_i, m_obsIndicate the true ginseng of sample Number；

It utilizesTo K_iIt is iterated calculating；

Reach training maximum step number M or meets the setting condition of convergenceWhen, update i =i+1 re-starts iterative calculation, RMS₀For the convergence threshold of setting；

K corresponding to all samples_iForm the gradient decline matrix K of entire training sample.

3. the time-domain electromagnetic data object inversion method according to claim 2 based on supervision descent method, which is characterized in that The gradient using entire training sample declines matrix and the observation vector of inverting target carries out inverting, obtains inverting target Actual parameter are as follows:

If the initial value of inverting target component is m₀, calculate the forward response d under the initial value₀, and enable initial value=0 of i；

Calculate Δ d_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of inverting target；

According to sample K obtained each in training process_i, update m_i=m_i+Δd_iK_i；

Judge whether that at least one meets the condition of convergenceOr reach greatest iteration Step number M, RMS₀For the convergence threshold of setting, if so, by minimum RMS_iCorresponding sample parameter is as the parameter for being really target m_f, i=i+1 is otherwise enabled, update judgement is re-started.

4. according to claim 1 or the 3 time-domain electromagnetic data object inversion methods based on supervision descent method, feature exist In the sample parameter and target component representation are m:

M=[x₀,y₀,z₀,α,β,k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃], wherein x₀,y₀,z₀Respectively represent the three-dimensional of target Coordinate, α, β respectively represent the land burial inclination angle of target, with three-dimensional doublet forward model approximate representation target, three-dimensional dipole Subparameter is respectively k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃。

5. the time-domain electromagnetic data object inversion method based on supervision descent method according to Claims 2 or 3, maximum to change Ride instead of walk several M=10.

Technical field

The present invention provides a kind of time-domain electromagnetic data object inversion method based on supervision descent method, belongs to buried target Detection technique field.

Background technique

Time domain electromagnetic method is a kind of artificial source's lossless detection method established on the basis of electromagnetic induction principle, it is utilized Earth-free loop line (magnetic source) or ground connection line source (Electric Dipole) are to underground transmitting primary field, under its excitation, in sub-surface conductors target The induction field that the inductive loop motivated changes over time generation.Compared with frequency domain electromagnetic methods, time domain electromagnetic method has Do not influenced by primary field, the advantages that signal-to-noise ratio is high, there is higher detection and resolution capability, thus time domain electromagnetic method at For one of the important method of shallow underground target acquisition.Due to very low emission signal frequency, time domain electromagnetic method satisfaction is The diffusion equation of electromagnetic wave rather than wave equation cause to utilize since the bulk effect in diffusion field makes its resolution ratio very low Time domain electromagnetic method can not carry out direct imaging to buried target, can only carry out inverting to buried target by forward model and ask Solution, but the response of the secondary field of general finite conductor can not acquire analytic solutions, according to the numerical value meter of finite element or finite difference Calculation method, then calculation amount is excessively huge, therefore is changed come approximatively equivalent sub-surface conductors target by inverting using dipole model In generation, obtains the equivalent dipole intensity of target, and further judgement obtains the information of target.In inverting iterative process, initial value It chooses particularly important, directly affects final inversion result.Currently used method is to provide buried target using actual measurement response General plan-position, but target buried depth, inclination angle and dipole strength can not be obtained by actual measurement response, can only provide conjecture Initial value, and requirement of the local optimum LM algorithm generally used to initial value is very high, if initial value is inaccurate, what is finally solved is anti- It drills result and differs quite big with practical.And using global optimization DE algorithm, time-consuming, and global search and the local search of DE algorithm Ability be it is contradictory, equally exist the problem of cannot converging to optimal solution.

To sum up, there are following technological deficiencies for currently existing technology:

Due to very low electromagnetic induction signal frequency (usually tens to several hundred kHz), time domain electromagnetic method satisfaction is The diffusion equation of electromagnetic wave rather than wave equation cause to utilize since the bulk effect in diffusion field makes its resolution ratio very low Time domain electromagnetic method can not carry out direct imaging to buried target.Usually inverting is carried out to buried target by forward model to ask Solution, but the response of the secondary field of general finite conductor can not acquire analytic solutions；According to the numerical value meter of finite element or finite difference Calculation method, then calculation amount is excessively huge, is unfavorable for the real-time processing and analysis of data；Come according to forward models such as dipoles close As equivalent sub-surface conductors target, the equivalent dipole intensity of target is obtained to judge target signature by inverting iteration, The quality of inversion result usually has with the initial value of selection compared with Important Relations.In inverting iterative process, the selection of initial value is particularly important, Directly affect final inversion result.Currently used method is to provide the general plane of buried target using actual observation response Position, but target buried depth, inclination angle and dipole strength can not be obtained by actual measurement response, can only provide the initial value of conjecture, and The inverting optimization algorithm used all has higher dependence to initial value to a certain extent, if initial value is inaccurate, what is finally solved is anti- It drills result and differs quite big with practical.

Summary of the invention

In view of this, the present invention provides a kind of time-domain electromagnetic data object inversion method based on supervision descent method, it should Method can accurately obtain the target component of buried target.

Realize that technical scheme is as follows:

A kind of time-domain electromagnetic data object inversion method based on supervision descent method, detailed process are as follows:

Training process:

Different learning samples is set, obtains its corresponding secondary field response as observing matrix；

For each learning sample, it is iterated training, obtains the gradient decline matrix of entire training sample；

Refutation process:

Inverting target secondary field response measurement is carried out near abnormal area, obtains the observation vector of inverting target；

The initial parameter for setting inverting target declines the observation of matrix and inverting target using the gradient of entire training sample Vector carries out inverting, obtains the actual parameter of inverting target.

Further, the process of repetitive exercise of the present invention are as follows:

Set the initial parameter m of training sample₀, calculate the forward response d under the initial value₀, enable initial value=0 of i；

Calculate Δ d_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of sample, Δ m_i=m_obs-m_i, m_obsIndicate sample Actual parameter；

It utilizesTo K_iIt is iterated calculating；

Reach training maximum step number M or meets the setting condition of convergenceWhen, I=i+1 is updated, iterative calculation, RMS are re-started₀For the convergence threshold of setting；

K corresponding to all samples_iForm the gradient decline matrix K of entire training sample.

Further, the gradient of the present invention using entire training sample declines the observation vector of matrix and inverting target Inverting is carried out, the actual parameter of inverting target is obtained are as follows:

If the initial value of inverting target component is m₀, calculate the forward response d under the initial value₀, and enable initial value=0 of i；

Calculate Δ d_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of inverting target；

According to sample K obtained each in training process_i, update m_i=m_i+Δd_iK_i；

Judge whether that at least one meets the condition of convergenceOr reach maximum Iterative steps M, RMS₀For the convergence threshold of setting, if so, by minimum RMS_iGinseng of the corresponding sample parameter as really target Number m_f, i=i+1 is otherwise enabled, update judgement is re-started.

Further, sample parameter of the present invention and target component representation are m:

M=[x₀,y₀,z₀,α,β,k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃], wherein x₀,y₀,z₀Respectively represent target Three-dimensional coordinate, α, β respectively represent the land burial inclination angle of target, three-dimensional with three-dimensional doublet forward model approximate representation target Current dipole parameters are respectively k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃.；

Further, the maximum iterative steps M=10 of the present invention.

Beneficial effect

(1) the method for the present invention is directed to the characteristics of traditional optimal method is vulnerable to initial value affecting, proposes with supervision descent method Inverting is carried out to buried target, constantly updates iteration by the parameter that training set obtains under the premise of not needing prior information, Until being optimal solution, the parameter of accurate object to be measured is obtained.

(2) the method for the present invention directly obtains global downward gradient from training data, avoids in optimization in inverting Local minimum, do not need extra computation Jacob ian matrix and Hess i an matrix, greatly save operation time.

Detailed description of the invention

Fig. 1 is the flow chart of the method for the present invention；

Fig. 2 is that secondary field response measurement records coordinate schematic diagram in this example；

Fig. 3 is the learning process convergence curve of a sample in example；

Fig. 4 is the convergence curve of renewal process in example；

Fig. 5 is supervision descent method ((a): supervision compared with the dipole inversion result of tradition LM optimization algorithm in example Descent method (b): traditional LM optimization algorithm)；

Fig. 6 is the Comparative result (a) that descent method and tradition LM optimal method are supervised in example: being positive respectively from left to right Area's figure, the two residual plot (b) are surveyed in the inverting that the area Yan Ce figure, supervision descent method obtain: being from left to right respectively that forward modeling is surveyed area's figure, passed Area's figure, the two residual plot are surveyed in the inverting that LM optimal method of uniting obtains).

Specific embodiment

In order to make the object, technical scheme and advantages of the embodiment of the invention clearer, below in conjunction with the embodiment of the present invention In attached drawing, the technical scheme in the embodiment of the invention is clearly and completely described.

A kind of time-domain electromagnetic data object inversion method based on supervision descent method of the present embodiment, as shown in Figure 1, specifically Process are as follows:

Sample training:

(1) it is surveying in area, is placing different targets with different inclination angles in different coordinates respectively, generating a large amount of study Sample；

(2) secondary field that one-point measurement obtains sample (in the area Ji Ce) near abnormal area is generated in target to respond, it will be every Secondary field response splicing one row vector of composition for several measuring points that secondary measurement obtains is as observation vector, each learning sample A corresponding row vector is obtained, the row vector composition observation matrix of multiple samples is as training data；

For the corresponding row vector of each sample, following process is executed by the way of concurrent operation:

(3) target component to be solved is set as m=[x₀,y₀,z₀,α,β,k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃], wherein x₀,y₀,z₀The three-dimensional coordinate of target is respectively represented, α, β respectively represent the land burial inclination angle of target, with three-dimensional doublet forward modeling Model approximate representation target, three-dimensional doublet parameter are respectively k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃, only with target sheet Body property is related, such as size, shape and material etc..Select any initial value m₀, calculate the forward response d under the initial value₀, enable i Initial value=0.

(4) Δ d is calculated_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of sample, Δ m_i=m_obs-m_i, m_obsIndicate sample Actual parameter；

(5) the Δ d calculated according to step (4)_iWith Δ m_iTo K_iIt is iterated calculating, iterative formula is

(6) K is acquired_iAfterwards to m_iIt is updated m_i=m_i+Δd_iK_i, maximum step number M or meet the condition of convergence when reaching trainingWhen, stop iteration, wherein m_iFor the target component of current step number, i=i is otherwise enabled + 1, calculate current m_iUnder forward response d_i, then return step (4)；

(7) gradient for obtaining entire learning sample declines matrix K, includes each learning sample in the matrix in each solution The K that step number i is solved_i。

Refutation process:

(8) the secondary field response that one-point measurement inverting target near abnormal area is generated in target, obtains inverting target Observation vector；

Set the initial value m of inverting target component₀, calculate the forward response d under the initial value₀, and enable initial value=0 of i；

(9) Δ d is calculated_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of inverting target；

According to sample K obtained each in training process_i, all it is updated m_i=m_i+Δd_iK_i；

Judge whether that at least one meets the condition of convergenceOr reaches maximum and change Ride instead of walk several M, m_obsThe actual parameter for indicating sample, if so, by minimum RMS_iParameter m of the corresponding sample parameter as target_f, Otherwise i=i+1 is enabled, update judgement is re-started.

Example

In this example, by the time-domain electromagnetic data inversion method based on supervision descent method illustrated, to buried target Quick and precisely identified.

Sample training process are as follows:

(1) survey area be 0.5m*0.5m, establish observation coordinate system, by 20*5*5 (cm) steel drum, 20*10*10 (cm) steel drum and 20*20*10 (cm) steel disk is placed at different coordinates with different angle respectively, generates a large amount of learning sample, sample number at any time Amount is 2880, is a sample for example, 20*5*5 (cm) steel drum to be arranged in the position A for surveying area with certain inclination angle.

(2) in the secondary field response for surveying detection system interval 25cm one-point measurement sample object in area, referring to fig. 2, r is mesh Mark position, r_RFor ceoncentrically wound coil position, the secondary field response splicing of several measuring points in each measurement is formed one A row vector carries out the measurement of secondary field response for each sample standard deviation, obtains different row vector composition observation matrixes and makees For training data.

Following step (3)-(6) process is executed by the way of concurrent operation for each sample:

(3) target component to be solved is set as m=[x₀,y₀,z₀,α,β,k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃], In, x₀,y₀,z₀The three-dimensional coordinate of target is respectively represented, α, β respectively represent the land burial inclination angle of target, just with three-dimensional doublet Model approximate representation target is drilled, three-dimensional doublet parameter is respectively k₁,β₁,γ₁,k₂,β₂,γ₂,k₃,β₃,γ₃, only and target Nature is related, such as size, shape and material etc..Select any initial value m₀, calculate the forward response d under the initial value₀, Enable initial value=0 of i.

(4) Δ d is calculated_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of sample, Δ m_i=m_obs-m_i, m_obsIndicate sample Actual parameter；

(5) the Δ d calculated according to step (4)_iWith Δ m_iTo K_iIt is iterated calculating, iterative formula is

(6) K is acquired_iAfterwards to m_iIt is updated m_i=m_i+Δd_iK_i, judge whether to reach training maximum step number M or meet to receive Hold back conditionWherein M is the maximum step number 10 of training, if being off repeatedly Generation, m_iFor the target component of current step number, otherwise the learning process convergence curve of one of sample enables i=i+1 referring to Fig. 3, Calculate current m_iUnder forward response d_i, then return step (4).

(7) iteration that (6) reach greatest iteration step number is executed the step, the gradient decline matrix of entire learning sample is obtained K, dimension are P × Q, and wherein P is the length of row vector in step (2), and Q is the length of vector m to be solved；Entire training sample Shi Changwei 24.15s.

Refutation process are as follows:

(8) a large amount of test sample data, test sample 3840 are generated according to the method in (1).Choose test sample Any test data concentrated carries out inverting, chooses 20*5*5 (cm) steel drum herein with α=0 °, β=50 ° (0.21, 0.3, -0.81) the secondary field response of m, selectes initial value m₀, calculate the forward response d under the initial value₀, enable initial value=0 of i.

(9) Δ d is calculated_i=d_i-d_obs, d_obsIndicate the corresponding observation vector of inverting target；

(10) m is updated_i=m_i+Δd_iK_iRe-start iteration update；

Meet the m of the condition of convergence for first in iterative process_iPosition as target；

Do not meet the condition of convergence also when all training samples all reach maximum train epochs 10, at this time by 2880 training It is the smallest in sampleCorresponding m_iPosition as target；

The process of above-mentioned steps (9) (10) is to execute parallel to all learning samples (quantity 2880), such as first A learning sample executes, at this time K_iFor the K obtained when first training sample training_i。

This example stops iteration when reaching maximum train epochs 10, convergence curve referring to fig. 4, when inverting a length of 7.5s, mesh Mark final argument m_fIn position and inclination angle result and tradition LM optimal method inversion result comparison referring to table 1, target sheet Body dipole decay characteristics curve and tradition LM optimal method inversion result are compared referring to Fig. 5, as seen from the figure, under supervision The dipole decay characteristics curve of drop method is consistent with target signature, hence it is evident that better than the inversion result of traditional LM optimal method.Just The area Yan Ce figure surveys area's figure and residual plot referring to Fig. 6, by scheming with the inverting that supervision descent method and tradition LM optimal method obtain It can obtain, it is of slight difference to survey area and forward modeling survey area with the inverting of supervision descent method, and uses the inverting of tradition LM optimal method It surveys area and differs quite big with forward modeling survey area's result.

1 SDM of table and traditional inversion result compare

Fig. 5 is supervision descent method ((a): supervision compared with the dipole inversion result of tradition LM optimization algorithm in example Descent method (b): traditional LM optimization algorithm).

Fig. 6 is the Comparative result (a) that descent method and tradition LM optimal method are supervised in example: being positive respectively from left to right Area's figure, the two residual plot (b) are surveyed in the inverting that the area Yan Ce figure, supervision descent method obtain: being from left to right respectively that forward modeling is surveyed area's figure, passed Area's figure, the two residual plot are surveyed in the inverting that LM optimal method of uniting obtains).

In conclusion the above is merely preferred embodiments of the present invention, being not intended to limit the scope of the present invention. All within the spirits and principles of the present invention, any modification, equivalent replacement, improvement and so on should be included in of the invention Within protection scope.