Active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method

Abstract

The invention relates to an active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method, and provides an active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method aiming at induction response caused by conductivity, polarization response caused by chargeable property and magnetization response caused by magnetic complex physical property in polymetallic ore detection. When the emission current is in a flat-topped stage, based on an active magnetic method, magnetic susceptibility extraction is carried out by using an Oxim method; when the emission current is in a turn-off stage, resistivity and polarization extraction is carried out by using a depth Q network method based on a time domain electromagnetic method by taking the extracted magnetic susceptibility as a constraint condition; and finally, multi-parameter extraction of resistivity, polarizability and magnetic susceptibility of the active source magnetization-time domain electromagnetic induction polarization method is realized.

Description

Technical Field

[0001] The present invention relates to an active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method in the field of geophysical exploration, and in particular to a method for extracting magnetic susceptibility by utilizing a trapezoidal emission waveform in a flat-top phase and resistivity and polarizability by utilizing an off phase. Background Art

[0002] In geophysical exploration, magnetic susceptibility is mostly extracted from static magnetic data. The total magnetic field intensity usually measured is composed of the induced magnetization and residual magnetization vectors of the Earth's main magnetic field and magnetic anomalies. However, it is difficult to distinguish between induced magnetization and residual magnetization in traditional static magnetic fields. Residual magnetization seriously interferes with the extraction of magnetic susceptibility from static magnetic detection. On the other hand, the vertical component of the Earth's magnetic field is extremely weak or even close to zero in the Earth's magnetic equator. The induced magnetization data stimulated by this is abnormal, making it difficult to estimate the number of magnetic anomalies and their susceptibility. Active magnetic method can effectively make up for the shortcomings of static magnetic method. It uses magnetic field signals stimulated by active sources to detect underground magnetic media, improving detection resolution and signal-to-noise ratio.

[0003] In the detection of polymetallic minerals, the time-domain electromagnetic method has inductive responses caused by conductivity, polarization responses caused by chargeability, and magnetization responses caused by complex magnetic properties. It is impossible to extract all parameters using a single parameter extraction method. Therefore, a method for synchronous extraction of multiple parameters of resistivity, polarizability, and magnetic susceptibility is urgently needed. In addition, using only time-domain electromagnetic method data for multi-parameter extraction of resistivity, polarizability, and magnetic susceptibility is more prone to multiple solutions. Therefore, how to perform synchronous and high-precision extraction of resistivity, polarizability, and magnetic susceptibility is a technical problem that needs to be solved urgently by those skilled in the art. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide an active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method. When the emission current is in the flat-top stage, the magnetic susceptibility is extracted using the Occam method based on the active magnetic method; when the emission current is in the off stage, the resistivity and polarization are extracted using the deep Q network method based on the time domain electromagnetic method with the extracted magnetic susceptibility as a constraint condition.

[0005] The present invention is implemented as follows: a method for extracting multiple parameters of induced polarization by active source magnetization-time domain electromagnetic detection comprises:

[0006] 1) Split the flat-top phase data and the off-time data;

[0007] 2) Active magnetic method is used to extract magnetic susceptibility parameters from the flat-top stage data;

[0008] 3) Use the extracted magnetic susceptibility as a constraint in the time-domain electromagnetic method to extract resistivity and polarizability parameters;

[0009] 4) Obtain the resistivity, polarizability and magnetic susceptibility parameters.

[0010] Furthermore, in step 2, a parameter extraction model f = Gm is constructed, where m is the magnetic susceptibility, f is the response data, and G is the numerical simulation model operator; it is introduced into the Occam algorithm Where m ref is the magnetic susceptibility of the reference model, C is the roughness matrix composed of second-order differential operators, W z is a diagonal matrix, D is the diagonal weighting matrix of the reference model, d obs is the observed data, F(m) is the predicted response of the magnetic susceptibility model, λ and μ are trade-off parameters; the Occam parameter extraction model update formula is

[0011] The magnetic susceptibility χ is extracted by cyclic iteration;

[0012] Furthermore, in step 3, the extracted magnetic susceptibility χ is brought into the time domain electromagnetic numerical simulation as a constraint to correct the response curve; and a DQN parameter extraction model is constructed:

[0013] First, construct the objective function

[0014]

[0015] Where (σ,η) represents the resistivity σ and polarizability η in the current state, (σ′,η′) represents the resistivity σ′ and polarizability η′ in the next state, (Δσ,Δη) represents the current action perturbation value, (Δσ′,Δη′) represents the next action perturbation value, Q′ represents the updated state-action value, and the action reward r is obtained and evaluated in the state (σ,η). γ is the discount factor. The larger the γ, the more attention is paid to long-term rewards, and the smaller the γ, the more attention is paid to short-term benefits. θ is the neural network parameter; construct the target state-action value function y′=r+γmaxQ[(σ,η),(Δσ,Δη),θ], where y′ represents the target Q value; construct the loss function L(θ)=E{{y′-Q[(σ,η),(Δσ,Δη),θ]} 2}, where θ is the weight parameter trained in the neural network structure model.

[0016] Secondly, the Q value network is constructed. DQN evaluates the current state-action value function through the target network and the estimation network. The target function obtains the target Q value based on the neural network, and uses the target Q value to estimate the Q value at the next moment. The estimation network uses the stochastic gradient descent method to update the network weight Δθ, where is the gradient operator, responsible for adjusting the parameters along the negative gradient direction. The gradient descent algorithm is

[0017]

[0018] Finally, an experience replay pool is constructed. When the intelligent agent interacts with the environment, a sample database can be obtained. The sample database is stored in the established experience pool, and a portion of data is randomly extracted from the experience pool for training samples, which are then sent to the neural network for training.

[0019] Compared with the prior art, the present invention has the following beneficial effects: for the problem of multi-parameter extraction of resistivity, polarizability and magnetic susceptibility of polymetallic ores, the data is segmented, the magnetic susceptibility is extracted in the flat-top stage, constraints are provided for parameter extraction in the shut-off stage, the difficulty of parameter extraction is reduced, and the problem of multiple solutions is weakened. In the flat-top stage of the emission current, there is a linear relationship between the magnetic susceptibility parameter and the response of the active magnetic method, and the use of the Occam algorithm for parameter extraction can effectively improve the efficiency of parameter extraction. In the shut-off stage of the emission current, the magnetic susceptibility is used as a constraint condition to reduce the difficulty of parameter extraction, and a deep Q network method is used for parameter extraction, which integrates the perception ability of deep learning and the decision-making ability of reinforcement learning, without the need to build a huge data set, and is suitable for processing complex multi-effect multi-parameter extraction problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic diagram of the active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method;

[0021] Figure 2 This is a flowchart of magnetization parameter extraction based on Occam algorithm;

[0022] Figure 3 It is a flow chart of induction-polarization parameter extraction based on deep Q-network algorithm. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and specific implementation methods. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0024] Example

[0025] Combination Figure 1 As shown, a method for extracting multi-parameters of induced polarization by active source magnetization-time domain electromagnetic detection includes:

[0026] 1) Split the flat-top phase data and the off-time data;

[0027] 2) Active magnetic method is used to extract magnetic susceptibility parameters from the flat-top stage data;

[0028] 3) Use the extracted magnetic susceptibility as a constraint in the time-domain electromagnetic method to extract resistivity and polarizability parameters;

[0029] 4) Obtain the resistivity, polarizability and magnetic susceptibility parameters.

[0030] Furthermore, in step 2, combined Figure 2 As shown, the parameter extraction model f = Gm is constructed, where m is the magnetic susceptibility, f is the response data, and G is the numerical simulation model operator; it is brought into the Occam algorithm Where m ref is the magnetic susceptibility of the reference model, C is the roughness matrix composed of second-order differential operators, W z is a diagonal matrix, D is the diagonal weighting matrix of the reference model, d obs is the observed data, F(m) is the predicted response of the magnetic susceptibility model, λ and μ are trade-off parameters; the Occam parameter extraction model update formula is

[0031] The magnetic susceptibility χ is extracted by cyclic iteration;

[0032] Furthermore, in step 3, combined Figure 3 As shown in the figure, the extracted magnetic susceptibility χ is brought into the time-domain electromagnetic numerical simulation as a constraint to correct the response curve; the DQN parameter extraction model is constructed:

[0033] First, construct the objective function

[0034]

[0035] Where (σ,η) represents the resistivity σ and polarizability η in the current state, (σ′,η′) represents the resistivity σ′ and polarizability η′ in the next state, (Δσ,Δη) represents the current action perturbation value, (Δσ′,Δη′) represents the next action perturbation value, Q′ represents the updated state-action value, and the action reward r is obtained and evaluated in the state (σ,η). γ is the discount factor. The larger the γ, the more attention is paid to long-term rewards, and the smaller the γ, the more attention is paid to short-term benefits. θ is the neural network parameter; construct the target state-action value function y′=r+γmaxQ[(σ,η),(Δσ,Δη),θ], where y′ represents the target Q value; construct the loss function L(θ)=E{{y′-Q[(σ,η),(Δσ,Δη),θ]} 2}, where θ is the weight parameter trained in the neural network structure model.

[0036] Secondly, the Q value network is constructed. DQN evaluates the current state-action value function through the target network and the estimation network. The target function obtains the target Q value based on the neural network, and uses the target Q value to estimate the Q value at the next moment. The estimation network uses the stochastic gradient descent method to update the network weight Δθ, where is the gradient operator, responsible for adjusting the parameters along the negative gradient direction. The gradient descent algorithm is

[0037]

[0038] Finally, an experience replay pool is constructed. When the intelligent agent interacts with the environment, a sample database can be obtained. The sample database is stored in the established experience pool, and a portion of data is randomly extracted from the experience pool for training samples, which are then sent to the neural network for training.

[0039] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method, characterized in that: The steps include: 1) Split the flat-top phase data and the off-time data; 2) Active magnetic method is used to extract magnetic susceptibility parameters from the flat-top stage data; 3) Use the extracted magnetic susceptibility as a constraint in the time-domain electromagnetic method to extract resistivity and polarizability parameters; 4) Obtain the resistivity, polarizability and magnetic susceptibility parameters.

2. The active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method according to claim 1 is characterized in that: In step 2, a parameter extraction model f = Gm is constructed, where m is the magnetic susceptibility, f is the response data, and G is the numerical simulation model operator; it is introduced into the Occam algorithm Where m ref is the magnetic susceptibility of the reference model, C is the roughness matrix composed of second-order differential operators, W z is a diagonal matrix, D is the diagonal weighting matrix of the reference model, d obs is the observed data, F(m) is the predicted response of the magnetic susceptibility model, λ and μ are trade-off parameters; the Occam parameter extraction model update formula is The magnetic susceptibility χ is extracted through a loop iteration.

3. The active source magnetization-time domain electromagnetic induction polarization multi-parameter extraction method according to claim 1 is characterized in that: In step 3, the extracted magnetic susceptibility χ is used as a constraint condition to correct the response curve in the time domain electromagnetic numerical simulation; and a DQN parameter extraction model is constructed: First, construct the objective function Q′[(σ,η),(Δσ,Δη),θ]←Q[(σ,η),(Δσ,Δη),θ] +α{r+γmaxQ[(σ′,η′),(Δσ′,Δη′),θ]-Q[(σ,η),(Δσ,Δη),θ]}, Where (σ, η) represents the resistivity σ and polarizability η in the current state, (σ′, η′) represents the resistivity σ′ and polarizability η′ in the next state, (Δσ, Δη) represents the current action perturbation value, (Δσ′, Δη′) represents the next action perturbation value, Q′ represents the updated state-action value, and the action reward r is obtained under the state (σ, η) and evaluated; γ is the discount factor, the larger the γ, the more attention is paid to long-term rewards, and the smaller the γ, the more attention is paid to short-term benefits, and θ is the neural network parameter; construct the target state-action value function y′=r+γmaxQ[(σ, η),(Δσ, Δη),θ], where y′ represents the target Q value; construct the loss function L(θ)=E{{y′-Q[(σ, η),(Δσ, Δη),θ]} 2 }, where θ is the weight parameter trained in the neural network structure model; Secondly, construct the Q value network. DQN evaluates the current state-action value function through the target network and the estimation network. The objective function obtains the target Q value based on the neural network, and uses the target Q value to estimate the Q value at the next moment; The estimation network uses stochastic gradient descent to update the network weights Δθ, where ▽ θ is the gradient operator, responsible for adjusting parameters along the negative gradient direction; the gradient descent algorithm is Δθ=E{y′-Q[(σ,η),(Δσ,Δη),θ]}▽ θ Q[(σ,η),(Δσ,Δη),θ]; Finally, an experience replay pool is constructed. When the intelligent agent interacts with the environment, a sample database can be obtained. The sample database is stored in the established experience pool, and a portion of data is randomly extracted from the experience pool for training samples, which are then sent to the neural network for training.

Images