Controllable source electromagnetic apparent resistivity curve correction method and system based on deep learning

Abstract

The invention discloses a controllable source electromagnetic apparent resistivity curve correction method and system based on deep learning, and the method comprises the steps: generating N apparent resistivity curves through a wave impedance recursion theory, and recording the N apparent resistivity curves as noise-free apparent resistivity curves; random noise is added to the apparent resistivity curve without noise according to an electromagnetic noise influence rule, and an apparent resistivity curve with noise is obtained; a ResNet-Transformer network architecture is constructed, the apparent resistivity curve containing noise and the apparent resistivity curve not containing noise corresponding to the apparent resistivity curve serve as training input and output, and an apparent resistivity curve correction model is obtained through training; and inputting a controllable source electromagnetic apparent resistivity curve which is acquired in real time and is influenced by noise into the apparent resistivity curve correction model to correct the input apparent resistivity curve so as to obtain a smooth apparent resistivity curve. According to the method, the controllable source electromagnetic apparent resistivity curve influenced by noise can be effectively corrected, and reliable data support is provided for subsequent inversion interpretation.

Description

Technical Field

[0001] The present invention belongs to the technical field of controlled-source electromagnetic data processing, and particularly relates to a method and system for correcting apparent resistivity curves of controlled-source electromagnetic based on deep learning. Background Art

[0002] As one of the important means of geophysical exploration, the controlled-source electromagnetic method plays an irreplaceable role in deep resource exploration. However, the surface electromagnetic noise interference is becoming increasingly serious, resulting in the disorder and distortion of its apparent resistivity curves, making it difficult to accurately reflect the underground deep electrical structure.

[0003] Traditional methods for correcting apparent resistivity curves, such as polynomial fitting, rely on preset model assumptions and are difficult to capture the sudden changes in apparent resistivity curves caused by non-stationary noise. In addition, methods such as filtering require manual selection of thresholds or frequency bands and have limited ability to distinguish between overlapping noise and signals in the frequency domain. Summary of the Invention

[0004] Aiming at the technical problems that traditional methods for correcting apparent resistivity curves only pursue the smoothness of curve shapes and may violate the physical constraints of underground electrical structures, the present invention provides a method and system for correcting apparent resistivity curves of controlled-source electromagnetic based on deep learning to achieve high-precision correction of apparent resistivity curves under noise interference.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solutions: A method for correcting apparent resistivity curves of controlled-source electromagnetic based on deep learning, comprising: S1, generating N apparent resistivity curves through wave impedance recursion theory, denoted as apparent resistivity curves without noise; S2, adding random noise to the apparent resistivity curves without noise according to the influence law of electromagnetic noise to obtain apparent resistivity curves containing noise; S3, constructing a ResNet-Transformer network architecture, using the apparent resistivity curves containing noise as training inputs and the corresponding apparent resistivity curves without noise as training outputs, and training to obtain an apparent resistivity curve correction model; S4, inputting the real-time collected controlled-source electromagnetic apparent resistivity curves affected by noise into the apparent resistivity curve correction model, and correcting the input apparent resistivity curves through the model to output smooth apparent resistivity curves.

[0006] Further, generating each apparent resistivity curve through wave impedance recursion theory includes: S1.1, establishing a multi-layer geoelectric structure model and defining the value ranges of resistivity and layer thickness of each layer: ; where \(k\) represents the layer index of the multi-layer geoelectric structure, and respectively represent the resistivity and layer thickness of the \(k\)-th layer geoelectric structure; and respectively represent the minimum and maximum resistivities of the \(k\)-th layer, and respectively represent the minimum and maximum thicknesses of the \(k\)-th layer; S1.2. For each layer of the geoelectric structure, within the value range of the defined resistivity and layer thickness, randomly sample the resistivity and thickness, and at the same time introduce a resistivity flipping mechanism with different probabilities to simulate various geological anomalies; S1.3. Based on the electromagnetic field diffusion theory, recursively calculate the equivalent wave impedance on the surface of each layer of the geoelectric structure from the bottom layer upwards: ; where represents the equivalent wave impedance on the surface of the \(k\)-th layer geoelectric structure, represents the equivalent wave impedance on the surface of the \((k + 1)\)-th layer geoelectric structure, represents the intrinsic wave impedance of the \(k\)-th layer geoelectric structure, represents the propagation constant of the \(k\)-th layer geoelectric structure, represents the thickness of the \(k\)-th layer geoelectric structure; among them, the intrinsic wave impedance and propagation constant of each layer of the geoelectric structure are both related to the resistivity and emission frequency of the geoelectric structure of this layer; S1.4. According to the equivalent wave impedance on the surface of the first layer geoelectric structure, generate the apparent resistivity : ; where is the angular frequency, is the magnetic permeability, is the emission frequency of the controlled-source electromagnetic method, and ; S1.5. Based on the apparent resistivity obtained at different emission frequencies of the controlled-source electromagnetic method, construct an apparent resistivity curve with the emission frequency as the independent variable.

[0007] Furthermore, the calculation formulas for the intrinsic wave impedance and propagation constant of each layer of the geoelectric structure are: ; where \(i\) is the imaginary unit.

[0008] Furthermore, each apparent resistivity curve generated by S1 through the wave impedance recursion theory is filtered and then used as the data source for S2 and / or S3.

[0009] Further, adding random noise to the apparent resistivity curve without noise according to the law of electromagnetic noise influence includes: S2.1. Let the i-th apparent resistivity curve without noise be expressed as: ; In the formula, represents the apparent resistivity value in the apparent resistivity curve , respectively represent the emission frequencies of the controlled-source electromagnetic method; S2.2. Add the following generated random noise : ; In the formula, A represents the intensity of the random noise, represents a sequence composed of random values generated between [-1, 1]; S2.3. Add to to obtain the i-th apparent resistivity curve with noise in the logarithmic domain : ; S2.4. Finally, calculate , which is the i-th apparent resistivity curve with noise in the final conventional number domain.

[0010] Further, the constructed ResNet-Transformer network architecture includes, in sequence from input to output: a ResNet network layer, a Transformer network layer, and a regression layer;

[0011] For the ResNet network layer, first perform: one-dimensional convolution, one-dimensional pooling, and ReLU activation on the input apparent resistivity sequence with noise; then continue to perform: one-dimensional convolution, one-dimensional pooling, and ReLU activation on the output of the first ReLU activation; after performing one-dimensional convolution on the output of the first ReLU activation, then add it to the output of the second ReLU activation as the output of the ResNet network layer; For the Transformer network layer, first perform positional encoding on the output of the ResNet network layer, then process it using the self-attention mechanism, and then output after passing through the dropout layer and the fully connected layer; The regression layer performs regression output on the output of the Transformer network layer to obtain the apparent resistivity sequence without noise.

[0012] Further, when training the apparent resistivity curve correction model in S3, the apparent resistivity curves to be input and output from the model are first transformed into the logarithmic domain and then used as the training input and training output of the model. When performing correction using the trained apparent resistivity curve correction model in S4, for the real-time collected controlled-source electromagnetic apparent resistivity curve affected by noise, after converting its apparent resistivity value to the value in the logarithmic domain, it is input into the apparent resistivity curve correction model; and the output data of the apparent resistivity curve correction model is converted from the logarithmic domain to the conventional number domain, which is the finally smoothed apparent resistivity curve.

[0013] A system based on the above-mentioned controlled-source electromagnetic apparent resistivity curve correction method, which includes: A smoothed apparent resistivity curve generation module, configured to: generate N apparent resistivity curves through the wave impedance recursion theory, denoted as the apparent resistivity curves without noise; A noisy apparent resistivity curve generation module, configured to: add random noise to the apparent resistivity curve without noise according to the law of electromagnetic noise influence to obtain the apparent resistivity curve containing noise; An apparent resistivity curve correction model training module, configured to: on the basis of constructing a ResNet-Transformer network architecture, use the apparent resistivity curve containing noise as the training input and the corresponding apparent resistivity curve without noise as the training output to train and obtain the apparent resistivity curve correction model; An apparent resistivity curve correction model, configured to: correct the real-time input controlled-source electromagnetic apparent resistivity curve affected by noise and output the smoothed apparent resistivity curve.

[0014] Compared with the prior art, the present invention has the following technical effects: The controlled-source electromagnetic apparent resistivity curve correction method based on deep learning provided by the present invention, on the one hand, ResNet extracts local features of the input data (such as noise distortion, curve mutation, etc.) through the convolutional layers in the residual blocks, and retains the input details through skip connections to avoid the problem of gradient disappearance, so as to effectively capture the local distortion features under noise interference. On the other hand, the self-attention mechanism of Transformer can model the global correlation between any two points in the sequence to ensure the overall smoothness and physical consistency of the corrected curve. Moreover, ResNet-Transformer combines the deep learning ability of ResNet and the advantages of the self-attention mechanism of Transformer, and can realize the collaborative optimization of local noise suppression and the overall curve trend. Again, the wave impedance recursion theory generates training data that conforms to geological laws, increasing the reliability of the apparent resistivity curve correction training.

[0015] The present invention utilizes the non - linear mapping relationship between the noisy apparent resistivity curve and the smoothed apparent resistivity curve in ResNet - Transformer learning of deep learning, and then applies it to the correction of measured apparent resistivity curves. Compared with traditional correction methods, the correction reliability and efficiency of the method of the present invention are higher. Compared with traditional deep learning networks such as CNN or LSTM networks, the ResNet - Transformer framework constructed in the present invention is more suitable for the controllable - source electromagnetic apparent resistivity correction task, and the improvement effect is better. Brief Description of the Drawings

[0016] Figure 1 It is a flowchart of the method for correcting the controllable - source electromagnetic apparent resistivity curve according to the embodiment of the present invention.

[0017] Figure 2 It is a comparison chart of apparent resistivity curves before and after adding random noise in the embodiment of the present invention.

[0018] Figure 3 It is a graph of apparent resistivity curves generated by the wave impedance recurrence theory in the embodiment of the present invention, where (a) to (i) respectively represent different apparent resistivity curves.

[0019] Figure 4 It is a structural diagram of the ResNet - Transformer network in the embodiment of the present invention.

[0020] Figure 5 It is a training curve graph of the ResNet - Transformer in the embodiment of the present invention, where (a) is the root - mean - square error RMSE and (b) is the loss.

[0021] Figure 6 It is a comparison chart of measured controllable - source electromagnetic apparent resistivity curves before and after correction in the embodiment of the present invention, where (a) to (i) respectively represent different apparent resistivity curves. Detailed Embodiment

[0022] The following makes a detailed description of the embodiments of the present invention. This embodiment is carried out based on the technical solution of the present invention, and gives a detailed implementation manner and specific operation process, further explaining the technical solution of the present invention.

[0023] Embodiment 1

[0024] This embodiment provides a method for correcting the controllable - source electromagnetic apparent resistivity curve based on deep learning. Referring to Figure 1 as shown, it includes the following steps.

[0025] S1, Generate N apparent resistivity curves through the wave impedance recurrence theory, denoted as apparent resistivity curves without noise.

[0026] S1.1. Establish a multi-layer geoelectric structure model and define the value ranges of the resistivity and layer thickness of each layer: ; where k represents the layer index of the multi-layer geoelectric structure, and represent the resistivity and layer thickness of the k-th layer geoelectric structure respectively; and represent the minimum and maximum resistivity of the k-th layer respectively, and represent the minimum and maximum thickness of the k-th layer respectively.

[0027] S1.2. For each layer of the geoelectric structure, randomly sample the resistivity and thickness within the defined value ranges of the resistivity and layer thickness, and introduce a resistivity flipping mechanism with different probabilities to simulate various geological anomalies.

[0028] S1.3. Based on the electromagnetic field diffusion theory, recursively calculate the equivalent wave impedance of the surface of each layer of the geoelectric structure from the bottom layer upwards: ; where represents the equivalent wave impedance of the surface of the k-th layer of the geoelectric structure, represents the equivalent wave impedance of the surface of the (k + 1)-th layer of the geoelectric structure, represents the intrinsic wave impedance of the k-th layer of the geoelectric structure, represents the propagation constant of the k-th layer of the geoelectric structure, represents the thickness of the k-th layer of the geoelectric structure. Among them, the intrinsic wave impedance and propagation constant of each layer of the geoelectric structure are both related to the resistivity and emission frequency of this layer of the geoelectric structure, and the calculation formulas are: ; where i is the imaginary unit, is the angular frequency, is the magnetic permeability, is the emission frequency of the controlled-source electromagnetic method, and .

[0029] S1.4. Generate the apparent resistivity of the multi-layer geoelectric structure according to the equivalent wave impedance of the surface of the first layer of the geoelectric structure: ; S1.5. Based on the apparent resistivity obtained at different emission frequencies of the controlled-source electromagnetic method, construct an apparent resistivity curve with the emission frequency as the independent variable.

[0030] In a more optimal embodiment, after converting each apparent resistivity curve generated by the wave impedance recursion theory into the logarithmic domain for filtering processing, it is used as the data source for the subsequent step S2 and step S3. As Figure 2 shown is the apparent resistivity curve diagram generated by the wave impedance recursion theory in this embodiment, which includes a total of 50 apparent resistivity curves.

[0031] S2. Add random noise to the apparent resistivity curve without noise according to the law of electromagnetic noise influence to obtain an apparent resistivity curve containing noise.

[0032] S2.1. Let any i-th apparent resistivity curve without noise be expressed as: ; In the formula, represents the apparent resistivity value in the apparent resistivity curve , respectively represent the emission frequencies of the controlled-source electromagnetic method; in this embodiment, .

[0033] S2.2. Add the following generated random noise : ; In the formula, A represents the intensity of the random noise, represents a sequence composed of random values generated between [-1, 1]; S2.3. Add to to obtain the i-th apparent resistivity curve containing noise in the logarithmic domain : ; S2.4. Finally, calculate , which is the i-th apparent resistivity curve containing noise in the final conventional number domain.

[0034] As Figure 3 shown is the comparison diagram of the apparent resistivity curves before and after adding random noise in this embodiment.

[0035] S3. Construct a ResNet-Transformer network architecture, use the apparent resistivity curve containing noise as the training input, and the corresponding apparent resistivity curve without noise as the training output, and train to obtain an apparent resistivity curve correction model.

[0036] The ResNet-Transformer network architecture constructed in the embodiment of the present invention, as Figure 4As shown, from input to output, it successively includes: a ResNet network layer, a Transformer network layer, and a regression layer; For the ResNet network layer, first, perform one-dimensional convolution, one-dimensional pooling, and ReLU activation on the input apparent resistivity sequence containing noise; then continue to perform one-dimensional convolution, one-dimensional pooling, and ReLU activation on the output of the first ReLU activation; after performing one-dimensional convolution on the output of the first ReLU activation, then add it to the output of the second ReLU activation as the output of the ResNet network layer; For the Transformer network layer, first perform positional encoding on the output of the ResNet network layer, then process it using the self-attention mechanism, and then output after passing through a dropout layer and a fully connected layer; The regression layer performs regression output on the output of the Transformer network layer to obtain an apparent resistivity sequence without noise.

[0037] In this embodiment, 4 / 5 of the training data is used as the training set to train the apparent resistivity curve correction model, and the other 1 / 5 of the training data is used as the validation set to verify the trained apparent resistivity curve correction model.

[0038] In a more optimal embodiment, when training the apparent resistivity curve correction model in step S3, first convert the apparent resistivity curve that needs to be input and output from the model to the logarithmic domain, and then use it as the training input and training output of the model.

[0039] Figure 5 The following shows the training curve of ResNet-Transformer in the embodiment of the present invention.

[0040] S4. Input the real-time collected controlled-source electromagnetic apparent resistivity curve affected by noise into the apparent resistivity curve correction model, and correct the input apparent resistivity curve through the model to output a smooth apparent resistivity curve.

[0041] In a more optimal embodiment, when using the trained apparent resistivity curve correction model for correction in step S4, for the real-time collected controlled-source electromagnetic apparent resistivity curve affected by noise, convert its apparent resistivity value to the value in the logarithmic domain and then input it into the apparent resistivity curve correction model; and convert the output data of the apparent resistivity curve correction model from the logarithmic domain to the conventional number domain, which is the final smooth apparent resistivity curve.

[0042] S4.1: Assume that the multiple collected controlled-source electromagnetic apparent resistivity curves containing noise are: ; In the formula, represents the nth apparent resistivity curve, denotes the apparent resistivity value in the nth apparent resistivity curve, denotes the main frequency value of the controlled-source electromagnetic method.

[0043] S4.2: Convert the controlled-source electromagnetic apparent resistivity curve containing noise in the following manner: ; S4.3: Sequentially input the converted noisy apparent resistivity curves into the trained apparent resistivity curve correction model, and the model outputs the corrected apparent resistivity curve .

[0044] Step 4.4: Restore the apparent resistivity curve corrected by the model to , and obtain the final corrected controlled-source electromagnetic apparent resistivity curve.

[0045] As Figure 6 shown is the comparison chart of the measured controlled-source electromagnetic apparent resistivity curve before and after correction. After being corrected by the method of the embodiment of the present invention, the apparent resistivity curves tend to be smooth, indicating that the method of the present invention can effectively improve the disorder phenomenon of the apparent resistivity curve caused by noise interference, and the data quality is significantly improved.

[0046] Embodiment 2 This embodiment provides a system based on the controlled-source electromagnetic apparent resistivity curve correction method described in Embodiment 1, including: Smoothing apparent resistivity curve generation module, for: generating N apparent resistivity curves through the wave impedance recursion theory, denoted as apparent resistivity curves without noise; Noisy apparent resistivity curve generation module, for: adding random noise to the apparent resistivity curve without noise according to the electromagnetic noise influence law to obtain the apparent resistivity curve containing noise; Apparent resistivity curve correction model training module, for: on the basis of constructing the ResNet-Transformer network architecture, using the apparent resistivity curve containing noise as the training input and the corresponding apparent resistivity curve without noise as the training output to train and obtain the apparent resistivity curve correction model; Apparent resistivity curve correction model, for: correcting the real-time input controlled-source electromagnetic apparent resistivity curve affected by noise and outputting a smooth apparent resistivity curve.

[0047] It should be understood that: the functional unit modules in each embodiment of the present invention can be concentrated in one processing unit, or each unit module can exist physically alone, or two or more unit modules can be integrated in one unit module, and can be implemented in the form of hardware or software.

[0048] The above embodiments are the preferred embodiments of the present application. Those of ordinary skill in the art can also make various transformations or improvements based on this. Without departing from the overall concept of the present application, these transformations or improvements should all fall within the scope of protection required by the present application.

Claims

1. A controllable-source electromagnetic apparent resistivity curve correction method based on deep learning, characterized in that, Including: S1. Generate N apparent resistivity curves through wave impedance recursion theory, denoted as apparent resistivity curves without noise. S2. Add random noise to the apparent resistivity curves without noise according to the influence law of electromagnetic noise to obtain apparent resistivity curves with noise. S3. Construct a ResNet-Transformer network architecture, use the apparent resistivity curves with noise as the training input, and the corresponding apparent resistivity curves without noise as the training output to train and obtain an apparent resistivity curve correction model. S4. Input the controllable-source electromagnetic apparent resistivity curves affected by noise collected in real time into the apparent resistivity curve correction model, and the model corrects the input apparent resistivity curves to output smooth apparent resistivity curves.

2. The controllable-source electromagnetic apparent resistivity curve correction method according to claim 1, characterized in that Generate each apparent resistivity curve through wave impedance recursion theory, including: S1.

1. Establish a multi-layer geoelectric structure model and define the value ranges of resistivity and layer thickness for each layer. ； where k represents the layer index of the multi-layer geoelectric structure, and respectively represent the resistivity and layer thickness of the k-th layer geoelectric structure; and respectively represent the minimum and maximum resistivities of the k-th layer, and respectively represent the minimum and maximum thicknesses of the k-th layer; S1.

2. For each layer of the geoelectric structure, randomly sample resistivity and thickness within the defined value ranges of resistivity and layer thickness, and introduce a resistivity flipping mechanism with different probabilities to simulate various geological anomalies. S1.

3. Based on the electromagnetic field diffusion theory, recursively calculate the equivalent wave impedance on the surface of each layer of the geoelectric structure from the bottom layer upwards. ； In the formula, represents the equivalent wave impedance of the surface of the k-th layer of geoelectric structure, represents the equivalent wave impedance of the surface of the (k + 1)-th layer of geoelectric structure, represents the intrinsic wave impedance of the k-th layer of geoelectric structure, represents the propagation constant of the k-th layer of geoelectric structure, represents the thickness of the k-th layer of geoelectric structure; among them, the intrinsic wave impedance and propagation constant of each layer of geoelectric structure are both related to the resistivity and emission frequency of the geoelectric structure of this layer; S1.4, generate the apparent resistivity of the multi-layer geoelectric structure based on the equivalent wave impedance on the surface of the first-layer geoelectric structure : ； In the formula, is the angular frequency, is the magnetic permeability, is the emission frequency of the controlled-source electromagnetic method, and ; S1.5, based on the apparent resistivity obtained at different transmitting frequencies of the controlled-source electromagnetic method, a curve of apparent resistivity with the transmitting frequency as the independent variable is formed.

3. The controllable-source electromagnetic apparent resistivity curve correction method according to claim 2, characterized in that, The intrinsic wave impedance and propagation constant of each layer of the geoelectric structure, the calculation formula is: ； In the formula, i is the imaginary unit.

4. The controllable-source electromagnetic apparent resistivity curve correction method according to claim 1, characterized in that The apparent resistivity curves generated by S1 through wave impedance recursion theory are filtered and then used as the data source for S2 and / or S3.

5. The controllable-source electromagnetic apparent resistivity curve correction method according to claim 1, wherein Adding random noise to the apparent resistivity curves without noise according to the influence law of electromagnetic noise includes: S2.1, Let any i-th apparent resistivity curve without noise be expressed as: ； In the formula, represents the apparent resistivity curve in the apparent resistivity value, respectively represent the emission frequencies of the controlled-source electromagnetic method; S2.2, add the randomly generated noise as follows : ； where A represents the intensity of random noise, represents a sequence composed of random values generated between [-1, 1]; S2.3, add to to obtain the i-th apparent resistivity curve containing noise in the logarithmic domain : ； S2.4, Final calculation , which is the i-th noisy apparent resistivity curve containing noise in the final conventional number field.

6. The controllable-source electromagnetic apparent resistivity curve correction method according to claim 1, wherein The constructed ResNet-Transformer network architecture includes, from input to output: a ResNet network layer, a Transformer network layer, and a regression layer. For the ResNet network layer, first perform one-dimensional convolution, one-dimensional pooling, and ReLU activation on the input apparent resistivity sequence with noise; then continue to perform one-dimensional convolution, one-dimensional pooling, and ReLU activation on the output of the first ReLU activation; after performing one-dimensional convolution on the output of the first ReLU activation, then add it to the output of the second ReLU activation as the output of the ResNet network layer. For the Transformer network layer, first perform position encoding on the output of the ResNet network layer, then process it using the self-attention mechanism, and then output after passing through the dropout layer and the fully connected layer. The regression layer performs regression output on the output of the Transformer network layer to obtain an apparent resistivity sequence without noise.

7. The controllable-source electromagnetic apparent resistivity curve correction method according to claim 1, characterized in that When training the apparent resistivity curve correction model in S3, first convert the apparent resistivity curves to be input and output from the model to the logarithmic domain and then use them as the training input and training output of the model. When performing calibration using the calibrated apparent resistivity curve calibration model in S4, for the real-time collected controllable-source electromagnetic apparent resistivity curve affected by noise, after converting its apparent resistivity value to the numerical value in the logarithmic domain, it is then input into the apparent resistivity curve calibration model; and the output data of the apparent resistivity curve calibration model is converted from the logarithmic domain to the conventional number domain, which is the finally smoothed apparent resistivity curve.

8. A system based on the controllable-source electromagnetic apparent resistivity curve correction method according to any one of claims 1-7, characterized in that, It includes: A smoothed apparent resistivity curve generation module, which is used to: generate N apparent resistivity curves through the wave impedance recursion theory, denoted as the apparent resistivity curves without noise; A noisy apparent resistivity curve generation module, which is used to: add random noise to the apparent resistivity curve without noise according to the electromagnetic noise influence law to obtain the apparent resistivity curve containing noise; An apparent resistivity curve calibration model training module, which is used to: on the basis of constructing the ResNet-Transformer network architecture, use the apparent resistivity curve containing noise as the training input and the corresponding apparent resistivity curve without noise as the training output to train and obtain the apparent resistivity curve calibration model; An apparent resistivity curve calibration model, which is used to: calibrate the real-time input controllable-source electromagnetic apparent resistivity curve affected by noise and output the smoothed apparent resistivity curve.

Images