Resistivity method data inversion method based on BP neural network

Abstract

The invention discloses a resistivity method data inversion method based on a BP neural network, and the method comprises the steps: constructing a horizontal stratified model, enabling model parameters to be the thickness of each layer and the resistivity of each layer, carrying out the forward modeling of the horizontal stratified model, obtaining the apparent resistivity under different polar distances, enabling the apparent resistivity after the forward modeling to serve as the input of the BP neural network, and obtaining the resistivity of the horizontal stratified model. The geoelectric model parameters are output, and the one-dimensional resistivity sounding geoelectric model parameters are obtained through training inversion of the BP neural network. Compared with a traditional linear inversion method, the degree of dependence on an initial model is reduced, and a Jacobian matrix does not need to be calculated in the inversion process. Compared with the inversion result of the traditional least square method, the inversion method provided by the invention can accurately approach a real model, and is feasible and effective.

Description

Technical Field

[0001] The invention relates to the technical field of resistivity imaging in the geophysical exploration industry, and in particular to a resistivity method data inversion method based on BP neural network. Background Art

[0002] The resistivity method is a branch of electrical exploration that uses the difference in conductivity between rocks and ores in the earth's crust as its material basis. It achieves the purpose of prospecting and solving other geological problems by observing and studying the distribution law of artificially established ground current fields (stable fields or alternating fields). The resistivity depth sounding method is to understand the vertical changes in the geological conditions below the measuring point by analyzing the geoelectric section reflected by the electrical depth sounding curve. At present, most of the inversion and interpretation work on georesistivity data is still limited to the traditional linear inversion method. Although linear inversion is relatively stable, its multiple solutions and easy to fall into local optimal problems have not been well solved.

[0003] In the prior art, artificial neural networks were used to invert one-dimensional electrical depth data and seismic wave data. The geoelectric model used for network training was extracted from a set of random models within some pre-specified parameter limits. Data simulation shows that the neural network can achieve parameter estimation of the one-dimensional model more accurately, but the experiments conducted cannot reasonably assign weights. Although the neural network inversion of one-dimensional and two-dimensional electrical depth data has been studied, the construction method of the neural network structure has been given, and the performance of different learning algorithms has been compared, it has only been applied to a one-dimensional six-layer model. The use of artificial neural networks to invert the apparent resistivity data of high resistivity contrast areas proves that the trained network can invert two-dimensional resistivity data, and the vertical structure has obtained relatively satisfactory results for deep electrical information, but the inversion results for the near-surface part are not ideal. Therefore, a resistivity method data inversion method based on BP neural network is needed to adapt to the inversion of various types of geoelectric model data, and at the same time, it can more finely characterize the near-surface resistivity structure. Summary of the invention

[0004] The purpose of the present invention is to provide a resistivity data inversion method based on BP neural network.

[0005] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0006] The present invention comprises the following steps:

[0007] Step S1 constructs a horizontal layered model based on the preprocessed geoelectric model parameters, and obtains an apparent resistivity data set by forward modeling the geoelectric model parameters;

[0008] Step S2 builds and trains a BP neural network model according to the constructed data set, and optimizes the BP neural network model through an optimization algorithm;

[0009] Step S3: input the test sample, predict the geoelectric model parameters through the BP neural network, and output the prediction result;

[0010] Step S4 analyzes and evaluates the prediction results.

[0011] Furthermore, the preprocessing method includes removing duplicate data, deleting missing values, processing outliers, data conversion, data normalization and data reduction.

[0012] Furthermore, a two-layer and three-layer horizontal layered model was constructed. The layer thickness ranged from 10 to 100 m with an interval of 10 m, and the resistivity ranged from 100 to 1000 Ω·m with an interval of 100 Ω·m. 21 groups of different pole distances were set. Through forward modeling, the apparent resistivity data corresponding to each geoelectric model parameter were obtained. The two-layer horizontal layered model contains 10 3 The combination includes all the models of type G and type D of the two-layer horizontal layer model, and the three-layer horizontal layer model includes 10 5 A combination of types, including all models of A, Q, H and K types of three-layer horizontal layered models.

[0013] Furthermore, the apparent resistivity is obtained by forward modeling different types of geoelectric models to form a data set.

[0014] Furthermore, the steps of constructing the BP neural network are:

[0015] Step S2.1: Input apparent resistivity data, divide it into training set and test set, and normalize the data;

[0016] Step S2.2: Specify the number of hidden layers, use the training set data to train the BP neural network, build a BP neural network model, and use the Levenberg-Marquardt algorithm to train the samples; the input and output characteristics of the BP neural network are expressed by the following equations:

[0017] The input to this neuron is: The output of this neuron is:

[0018] Where w i is the connection weight between the ith neuron and this neuron; x i is the input of the i-th neuron; f is the activation function; θ is the threshold of the hidden layer neural node;

[0019] The number of hidden layer nodes is related to the number of input and output parameters. The value range of the hidden layer can be determined according to the empirical formula:

[0020]

[0021] Among them, l is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and the value range of a is 1-10.

[0022] Step S2.3: Set the error function; select random numbers in [-1,1] to assign connection weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer; set the number of training times, error threshold, and learning rate training parameters.

[0023] The apparent resistivity data is used as the input set of the BP neural network, and the model parameters corresponding to the apparent resistivity are used as the output set. 80% of the apparent resistivity data of the classic model is used for training the BP neural network, and multiple inversion predictions are performed on 20% of the model data and the apparent resistivity data obtained by actual measurement.

[0024] Assuming the number of AB pole distances is m, the number of layered model layers is n, then the number of model parameters is 2n-1, and the input of the BP neural network is the apparent resistivity at different AB pole distances;

[0025] Step S2.4: Correct the connection weights and thresholds between the hidden layer and the output layer, and between the input layer and the hidden layer according to the network output;

[0026] Step S2.5: perform global error calculation to determine whether the requirements are met;

[0027] When the error reaches the preset accuracy or the number of learning times is greater than the designed maximum number, the calculation ends, otherwise it returns to step S2.1 for the next round of learning.

[0028] Furthermore, the step of predicting the geoelectric model parameters is:

[0029] Step S3.1: The selected apparent resistivity data and the corresponding geoelectric model parameters are sorted into the input and output formats of the BP neural network, wherein the first 80% of the data is used as a training set and the last 20% of the data is used as an output set to verify the prediction accuracy of the network;

[0030] Step S3.2: Setting the neural network parameters according to the number of inputs and outputs, and training the BP neural network;

[0031] Assuming the number of AB pole distances is m and the number of layered model layers is n, the number of model parameters is 2n-1, and the input of the BP neural network is the apparent resistivity at different AB pole distances:

[0032] p s =(p s1， p s2， ,…,p sm， ) T

[0033] The output is the model parameters:

[0034] λ=(λ 1， ,λ 2， …, λ 2n-1， )=(h 1， , p1, h 2， , p 2， …h n-1， , p n-1，， , p n )

[0035] Among them, h is the layer thickness of the model, and p is the layer resistivity of the model.

[0036] Furthermore, the method of optimizing the BP neural network model by an optimization algorithm comprises:

[0037] Set the optimization population size and calculate the updated position and velocity components of the optimized individuals:

[0038]

[0039] The inertia weight of the optimized individual is ω, the learning factors are ρ1 and ρ2, the random numbers are e1 and e2, and the optimal value of the individual searched in the dimension c where the optimized individual i is located is The global optimal value in all solution spaces is The position component of the optimized individual i in dimension c for the a+1th time is The optimized position component of individual i in dimension c for the ath time is The velocity component of the optimized individual i in dimension c is The change function is used to improve the weight, and the expression is:

[0040]

[0041] The preset weight upper limit is ω max , the preset weight lower limit is ω min , the improved weight is The number of iterations is t, and the maximum number of iterations is t max , introduce tent mapping to generate chaotic sequence, the expression is:

[0042]

[0043] The chaotic parameter is γ, the chaotic sequence of the a+1th time is Q(a+1), and the chaotic sequence of the ath time is Q(a). The chaotic sequence is used to update the speed of the optimization individual, and the expression is:

[0044]

[0045] The optimized individual speed after updating is Iterate until the optimal weight is output.

[0046] The beneficial effects of the present invention are:

[0047] The present invention is a resistivity method data inversion method based on BP neural network. Compared with the prior art, the present invention has the following technical effects:

[0048] The invention overcomes the serious dependence of the traditional linear inversion method on the initial model and greatly reduces the calculation difficulty compared with the linear inversion method. The inversion method can accurately approximate the real model, is effective and feasible, and reduces the inversion difficulty. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a flow chart of the resistivity method data inversion steps based on BP neural network described in the patent of the present invention;

[0050] Figure 2 It is the flow chart of the BP neural network algorithm described in the patent of this invention;

[0051] Figure 3 The measuring point position described in the present invention is the measured data - the 40-meter measuring point measured and forward-modeled apparent resistivity fitting curve;

[0052] Figure 4 The measuring point position described in the present invention is the measured data 280 meters measuring point actual measurement and forward modeling apparent resistivity fitting curve diagram;

[0053] Figure 5 The measuring point position described in the present invention is the measured data 760 meters measuring point actual measurement and forward modeling apparent resistivity fitting curve diagram;

[0054] Figure 6 The measuring point position described in the present invention is the measured data 1320 meters measuring point actual measurement and forward modeling apparent resistivity fitting curve diagram;

[0055] Figure 7 It is the resistivity cross-section diagram inverted by BP neural network of measured data described in the patent of the present invention;

[0056] Figure 8 It is a resistivity cross-section diagram inverted by the least square method of the measured data described in the patent of this invention. DETAILED DESCRIPTION

[0057] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0058] The present invention provides a resistivity data inversion method based on BP neural network, comprising the following steps:

[0059] like Figure 1 and Figure 2 As shown, in this embodiment, the following steps are included:

[0060] Step S1 constructs a horizontal layered model based on the preprocessed geoelectric model parameters, and obtains an apparent resistivity data set by forward modeling the geoelectric model parameters;

[0061] Step S2 builds and trains a BP neural network model according to the constructed data set, and optimizes the BP neural network model through an optimization algorithm;

[0062] Step S3: input the test sample, predict the geoelectric model parameters through the BP neural network, and output the prediction result;

[0063] Step S4 analyzes and evaluates the prediction results.

[0064] In this implementation example, the preprocessing method includes removing duplicate data, deleting missing values, processing outliers, data conversion, data normalization and data reduction.

[0065] In this implementation example, the field measured data are divided into different types according to different measuring points (the measuring point locations are -200m to 1320m, with one measuring point every 80 meters), the data are preliminarily processed, the model parameters are set, and a certain range of values ​​are selected for forward calculation to obtain the apparent resistivity data of the geoelectric model for training the BP neural network.

[0066] In this implementation example, the steps of constructing the BP neural network are:

[0067] Step S2.1: Input apparent resistivity data, divide it into training set and test set, and normalize the data;

[0068] Step S2.2: Specify the number of hidden layers, use the training set data to train the BP neural network, build a BP neural network model, and use the Levenberg-Marquardt algorithm to train the samples; the input and output characteristics of the BP neural network are expressed by the following equations:

[0069] The input to this neuron is: The output of this neuron is:

[0070] Where w i is the connection weight between the ith neuron and this neuron; x i is the input of the i-th neuron; f is the activation function; θ is the threshold of the hidden layer neural node;

[0071] The number of hidden layer nodes is related to the number of input and output parameters. The value range of the hidden layer can be determined according to the empirical formula:

[0072]

[0073] Among them, l is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and the value range of a is 1-10.

[0074] Step S2.3: Set the error function; select random numbers in [-1,1] to assign connection weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer; set the number of training times, error threshold, and learning rate training parameters.

[0075] The apparent resistivity data is used as the input set of the BP neural network, and the model parameters corresponding to the apparent resistivity are used as the output set. 80% of the apparent resistivity data of the classic model is used for training the BP neural network, and multiple inversion predictions are performed on 20% of the model data and the apparent resistivity data obtained by actual measurement.

[0076] Assuming the number of AB pole distances is m, the number of layered model layers is n, then the number of model parameters is 2n-1, and the input of the BP neural network is the apparent resistivity at different AB pole distances;

[0077] Step S2.4: Correct the connection weights and thresholds between the hidden layer and the output layer, and between the input layer and the hidden layer according to the network output;

[0078] Step S2.5: perform global error calculation to determine whether the requirements are met;

[0079] When the error reaches the preset accuracy or the number of learning times is greater than the designed maximum number, the calculation ends, otherwise it returns to step S2.1 for the next round of learning.

[0080] In this implementation example, the steps of predicting the geoelectric model parameters are:

[0081] Step S3.1: The selected apparent resistivity data and the corresponding geoelectric model parameters are sorted into the input and output formats of the BP neural network, wherein the first 80% of the data is used as a training set and the last 20% of the data is used as a test set to verify the prediction accuracy of the network;

[0082] Step S3.2: Setting the neural network parameters according to the number of inputs and outputs, and training the BP neural network;

[0083] Assuming the number of AB pole distances is m and the number of layered model layers is n, the number of model parameters is 2n-1, and the input of the BP neural network is the apparent resistivity at different AB pole distances:

[0084] p s =(p s1， p s2， ,…,p sm， ) T

[0085] The output is the model parameters:

[0086] λ=(λ 1， ,λ 2， …, λ 2n-1， )=(h 1， , p1, h 2， , p 2， …h n-1， , p n-1，， , p n )

[0087] Among them, h , is the layer thickness of the model, p , is the layer resistivity of the model.

[0088] In this implementation example, the method of optimizing the BP neural network model by the optimization algorithm includes:

[0089] Set the optimization population size and calculate the updated position and velocity components of the optimized individuals:

[0090]

[0091] The inertia weight of the optimized individual is ω, the learning factors are ρ1 and ρ2, the random numbers are e1 and e2, and the optimal value of the individual searched in the dimension c where the optimized individual i is located is The global optimal value in all solution spaces is The position component of the optimized individual i in dimension c for the a+1th time is The optimized position component of individual i in dimension c for the ath time is The velocity component of the optimized individual i in dimension c is The change function is used to improve the weight, and the expression is:

[0092]

[0093] The preset weight upper limit is ω max , the preset weight lower limit is ω min , the improved weight is The number of iterations is t, and the maximum number of iterations is t max , introduce tent mapping to generate chaotic sequence, the expression is:

[0094]

[0095] The chaotic parameter is γ, the chaotic sequence of the a+1th time is Q(a+1), and the chaotic sequence of the ath time is Q(a). The chaotic sequence is used to update the speed of the optimization individual, and the expression is:

[0096]

[0097] The optimized individual speed after updating is Iterate until the optimal weight is output.

[0098] Four groups of models were selected to plot the measured and forward-fitted apparent resistivity ρ s Curve (such as Figures 3 to 6 ). After calculation, the maximum fitting error is 7.68% and the minimum is 0.63%. At the same time, by integrating the data, the BP neural network inversion resistivity cross-section diagram is drawn ( Figure 7 ), and the resistivity cross section is inverted by the least squares method ( Figure 8 ) comparison, confirming the reliability and feasibility of the method. At the same time, when the embodiment is applied, the data contained in the training set is all model data from the second layer to the sixth layer. After training with different types of model data, the global inversion ability and inversion accuracy of the BP neural network are improved. The introduction of the BP artificial neural network can not only avoid the huge amount of calculation, but also the BP neural network has a high degree of self-learning and self-adaptation ability and strong nonlinear mapping ability, which can significantly improve the inversion accuracy.

[0099] 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, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A resistivity data inversion method based on BP neural network, characterized in that: The following steps are involved: Step S1 constructs a horizontal layered model based on the preprocessed geoelectric model parameters, and obtains an apparent resistivity data set by forward modeling the geoelectric model parameters; Step S2 builds and trains a BP neural network model according to the constructed data set, and optimizes the BP neural network model through an optimization algorithm; Step S3: input the test sample, predict the geoelectric model parameters through the BP neural network, and output the prediction result; Step S4 analyzes and evaluates the prediction results.

2. The resistivity data inversion method based on BP neural network according to claim 1 is characterized in that: The preprocessing method described in step S1 includes removing duplicate data, deleting missing values, processing outliers, data conversion, data normalization and data reduction.

3. The resistivity data inversion method based on BP neural network according to claim 1 is characterized in that: A two-layer and three-layer horizontal layered model was constructed. The layer thickness ranged from 10 to 100 m with an interval of 10 m, and the resistivity ranged from 100 to 1000 Ω·m with an interval of 100 Ω·m. 21 groups of different pole distances were set. Through forward modeling, the apparent resistivity data corresponding to each geoelectric model parameter were obtained. The two-layer horizontal layered model contains 10 3 The combination includes all the models of type G and type D of the two-layer horizontal layer model, and the three-layer horizontal layer model includes 10 5 A combination of types, including all models of A, Q, H and K types of the three-layer horizontal layered model.

4. The resistivity data inversion method based on BP neural network according to claim 1 is characterized in that: The apparent resistivity is obtained by forward modeling different types of geoelectric models to form a data set.

5. The resistivity data inversion method based on BP neural network according to claim 1 is characterized in that: The construction steps of the BP neural network are: Step S2.1: Input apparent resistivity data, divide it into training set and test set, and normalize the data; Step S2.2: Specify the number of hidden layers, use the training set data to train the BP neural network, build a BP neural network model, and use the Levenberg-Marquardt algorithm to train the samples; the input and output characteristics of the BP neural network are expressed by the following equations: The input to this neuron is: The output of this neuron is: Where w i is the connection weight between the ith neuron and this neuron; x i is the input of the i-th neuron; f is the activation function; θ is the threshold of the hidden layer neural node; The number of hidden layer nodes is related to the number of input and output parameters. The value range of the hidden layer can be determined according to the empirical formula: Among them, l is the number of hidden layer nodes, m is the number of input layer nodes, n is the number of output layer nodes, and the value range of a is 1-10; Step S2.3: setting the error function; Select random numbers within [-1,1] to assign connection weights and thresholds between the input layer and the hidden layer, and between the hidden layer and the output layer; set the training parameters of the number of trainings, error threshold, and learning rate, use the apparent resistivity data as the input set of the BP neural network, and the model parameters corresponding to the apparent resistivity as the output set. Use 80% of the apparent resistivity data of the classic model for the training of the BP neural network, and perform multiple inversion predictions on 20% of the model data and the apparent resistivity data obtained by actual measurement; Assuming the number of AB pole distances is m, the number of layered model layers is n, then the number of model parameters is 2n-1, and the input of the BP neural network is the apparent resistivity at different AB pole distances; Step S2.4: Correct the connection weights and thresholds between the hidden layer and the output layer, and between the input layer and the hidden layer according to the network output; Step S2.5: perform global error calculation to determine whether the requirements are met; When the error reaches the preset accuracy or the number of learning times is greater than the designed maximum number, the calculation ends, otherwise it returns to step S2.1 for the next round of learning.

6. The resistivity data inversion method based on BP neural network according to claim 1 is characterized in that: The steps of predicting the geoelectric model parameters are as follows: Step S3.1: The selected apparent resistivity data and the corresponding geoelectric model parameters are sorted into the input and output formats of the BP neural network, wherein the first 80% of the data is used as a training set and the last 20% of the data is used as a test set to verify the prediction accuracy of the network; Step S3.2: Setting the neural network parameters according to the number of inputs and outputs, and training the BP neural network; Assuming the number of AB pole distances is m and the number of layered model layers is n, the number of model parameters is 2n-1, and the input of the BP neural network is the apparent resistivity at different AB pole distances: p s =(p s1, p s2, ,...,p sm, ) T The output is the model parameters: λ=(λ 1, ,l 2, ...,l 2n-1, )=(h 1, ,p 1, h 2, ,p 2, ...h n-1 ,,p n-1 ,,,p n ) Among them, h , is the layer thickness of the model, p , is the layer resistivity of the model.

7. The resistivity data inversion method based on BP neural network according to claim 1 is characterized in that: The method for optimizing the BP neural network model by an optimization algorithm comprises: Set the optimization population size and calculate the updated position and velocity components of the optimized individuals: The inertia weight of the optimized individual is ω, the learning factors are ρ1 and ρ2, the random numbers are e1 and e2, and the optimal value of the individual searched in the dimension c where the optimized individual i is located is The global optimal value in all solution spaces is The position component of the optimized individual i in dimension c for the a+1th time is The optimized position component of individual i in dimension c for the ath time is The velocity component of the optimized individual i in dimension c is The change function is used to improve the weight, and the expression is: The preset weight upper limit is ω max , the preset weight lower limit is ω min , the improved weight is The number of iterations is t, and the maximum number of iterations is t max , introduce tent mapping to generate chaotic sequence, the expression is: The chaotic parameter is γ, the chaotic sequence of the a+1th time is Q(a+1), and the chaotic sequence of the ath time is Q(a). The chaotic sequence is used to update the speed of the optimization individual, and the expression is: The optimized individual speed after updating is Iterate until the optimal weight is output.

Images