A method for multi-layer formation inversion of a while-drilling electromagnetic boundary-probe instrument

Abstract

The present application relates to well logging technology, and discloses a multilayer formation inversion method of a while-drilling electromagnetic boundary detection instrument, which comprises two stages of offline training and online inversion. In the offline stage, first, a plurality of formation structure types are constructed and training samples are generated; then, the original response vector of the forward modeling of the samples is enhanced and reconstructed to improve the data dimension and sensitivity; then, for each formation type, a dedicated gradient optimization set is trained by a supervised descent method. In the online stage, the original response measured in real time is subjected to the same feature reconstruction, and its belonging formation type is intelligently judged, and the matched gradient optimization set is automatically called to perform rapid iterative inversion. In the inversion iteration, dynamic attenuation regularization and geological prior constraint are integrated to effectively suppress noise, guarantee stability and geological rationality. The present application realizes rapid, accurate and stable inversion of complex multilayer formation, and significantly improves the reliability of while-drilling geosteering.

Description

TECHNICAL FIELD

[0001] The present application relates to well logging technology, in particular to a while-drilling data processing technology. BACKGROUND

[0002] The while-drilling electromagnetic boundary detection instrument (or the while-drilling electromagnetic boundary detection instrument) is a key means of horizontal well geosteering, and the relative position of the drill bit and the formation boundary can be judged in real time through inversion to guide the well trajectory adjustment, and the precision and stability of the inversion result directly affect the reliability of the well trajectory optimization and reservoir development decision. At present, the inversion technology in this field mainly faces the following bottlenecks:

[0003] I. In terms of formation model:

[0004] (1) The precision is limited due to the simplification of the inversion model: the existing mainstream inversion method for the while-drilling electromagnetic boundary detection instrument is mostly based on a three-layer (upper surrounding rock-reservoir-lower surrounding rock) formation model. However, with the increasing demand for exploitation, the actual geological conditions are more complex, and five-layer or more layer formation structures (such as thin interbedding, complex interlayer) are often encountered. The three-layer model is too simplified to accurately describe such complex formations, resulting in distorted inversion results and misleading geosteering decisions.

[0005] (2) Special challenges of multi-layer formation inversion: when the inversion model is expanded to five layers or more, the parameters to be inverted (resistivity of each layer, boundary position) increase sharply, and in the case of maintaining the amount of measurement response data, it will lead to the problem of "underdetermination" of inversion, and it is difficult to converge, and other problems such as strong multi-solution, unstable inversion, etc. At the same time, the measurement signal has low sensitivity to the detection of the outermost layer parameters, further exacerbating the uncertainty of the inversion.

[0006] II. In terms of inversion algorithm:

[0007] The existing traditional inversion algorithm cannot meet the needs of multi-layer formation inversion of the while-drilling electromagnetic boundary detection instrument in actual well logging operations:

[0008] (1) Deterministic inversion algorithm: the inversion accuracy of the gradient-based deterministic inversion method (such as gradient descent method, Gauss-Newton method, etc.) depends heavily on the quality of the initial model; the inversion efficiency and stability are insufficient, and it is easy to fall into local extremum; the Jacobian matrix and Hessian matrix need to be calculated for each iteration, and a large number of forward simulations are required, which has high computational cost.

[0009] (2) Heuristic global search algorithm (such as genetic algorithm, etc.): although it can globally optimize, it usually requires a large number of forward calculations, and the inversion speed is slow, which cannot meet the real-time requirements at all.

[0010] (3) Neural network method of pure data-driven method: Neural network realizes rapid inversion through "black box" fitting, but its training needs massive data and is divorced from the physical forward process, the inversion result lacks interpretability, is poor in stability in complex and unseen formation structure, has poor model generalization ability, and has high engineering application risk.

[0011] In summary, the existing traditional technology cannot achieve a good balance among calculation efficiency, inversion accuracy, method stability and physical interpretability, which restricts the application effect of the while-drilling electromagnetic boundary exploration technology in complex oil and gas reservoirs.

[0012] In the face of the inversion limitations of the above methods, the inversion method based on supervised descent combines the traditional algorithm and the supervised descent method of deep learning, and fuses the physical constraint of traditional deterministic inversion and the training idea of machine learning, which can speed up the convergence speed and improve the global optimization ability of inversion, and can improve the limitations of the existing inversion algorithm to a certain extent.

[0013] However, in the face of the multi-formation and multi-parameter inversion problem of the while-drilling electromagnetic boundary exploration model, the inversion method based on supervised descent still has the following technical difficulties:

[0014] (1) The average gradient optimization set obtained by training based on fixed sample data has poor adaptability to any actual formation structure model and poor accurate optimization ability;

[0015] (2) The matrix dimension of the gradient optimization set obtained by training is low, and the generalization ability is weak and the detection sensitivity is poor.

[0016] (3) The inversion algorithm has weak noise resistance and insufficient stability, and is difficult to adapt to the inversion needs of real well data.

[0017] With the increasing complexity of logging formation structure and the improvement of exploitation demand, the traditional inversion method is difficult to meet the actual engineering application demand in the aspects of inversion model and inversion algorithm of while-drilling electromagnetic boundary exploration:

[0018] (1) Insufficient model precision and complexity: the simplified three-layer model cannot depict complex formations, and after being expanded to a multi-layer model, the parameter increases dramatically, leading to "underdetermined" inversion problem, multiple solutions, convergence difficulty and poor inversion ability of weak sensitive parameters (such as outer resistivity).

[0019] (2) Existing inversion algorithm is difficult to adapt: deterministic algorithm (such as Gauss-Newton method) has high calculation cost, depends on initial value and is easy to fall into local extremum; heuristic algorithm has too large calculation amount and cannot meet real-time requirements; pure data-driven neural network method has weak generalization ability, lacks physical interpretation and is insufficient in stability.

[0020] (3) Supervised descent-based inversion methods face technical challenges when dealing with multi-layer, multi-parameter inversion problems of electromagnetic boundary exploration instruments while drilling:

[0021] Limited data information utilization and sensitivity: The limited dimension of the original measurement data restricts the expression ability of the gradient optimization set, resulting in insufficient sensitivity to changes in formation parameters (especially outer boundary and resistivity), affecting the inversion resolution.

[0022] Insufficient model generalization and adaptability: The "average" gradient optimization set trained based on a fixed data set is difficult to accurately adapt to any variable and unknown actual formation structure underground, resulting in a decline in the precise optimization ability of a specific model.

[0023] Weak robustness and anti-interference ability of formation inversion: The inversion process lacks effective constraints on the physical prior of the formation, is sensitive to noise in the measured data, has poor stability and anti-interference ability, and is difficult to guarantee the reliability and geological reasonableness of the inversion results in real well environments. SUMMARY

[0024] The technical problem to be solved by the present application is to provide a more rapid, accurate and stable formation inversion method that can be applied to electromagnetic boundary exploration instruments while drilling for five-layer and more complex formations.

[0025] The technical solution adopted by the present application to solve the above technical problem is a multi-layer formation inversion method for electromagnetic boundary exploration instruments while drilling, comprising the steps of:

[0026] Offline training phase:

[0027] Constructing several types of formation structure, and generating a training sample set containing a formation parameter vector and a forward original response vector for each formation structure type;

[0028] Wherein, the formation parameter vector includes a formation resistivity and a formation boundary, is an integer greater than or equal to 5; the formation structure type is constructed based on the apparent resistivity layer type and the position area type based on the geological signal; the apparent resistivity and the geological signal under a specific frequency and source distance are included in the forward original response vector;

[0029] Performing feature enhancement and reconstruction on the forward original response vector of each training sample to generate an enhanced feature response vector containing original response and enhanced features;

[0030] For each formation structure type, use its training sample set to train a set of gradient optimization sets corresponding to each formation structure type by supervised descent method, and each set of gradient optimization sets contains a gradient descent matrix and a bias term for each iteration.

[0031] Online inversion stage:

[0032] Obtaining the forward original response vector of real-time measurement, and performing feature enhancement and reconstruction to obtain an enhanced feature response vector; meanwhile, judging the stratum structure type to which the current measurement point belongs according to the forward original response vector;

[0033] Calling a set of gradient optimization corresponding to the stratum structure type matched with the current measurement point, and using the set of gradient optimization and the enhanced feature response vector to perform inversion iteration on the stratum parameter vector, and outputting the final stratum model parameter after the inversion iteration is completed.

[0034] Further, the enhanced feature response vector is generated from the forward original response vector, including but not limited to:

[0035] Extracting the peak value information of the forward original response vector as a first enhanced feature:

[0036] Extracting the start and end net change of the forward original response vector as a second enhanced feature:

[0037] Extracting the average change trend of the forward original response vector as a third enhanced feature:

[0038] Extracting the convexity of the forward original response vector as a fourth enhanced feature:

[0039] The forward original response vector and the first to fourth enhanced features extracted above are spliced into a vector to obtain an enhanced feature response vector.

[0040] Further, in the online inversion stage, a dynamic attenuation regularization term and a geological prior constraint term are introduced during the inversion iteration; the geological prior constraint term uses the stratum parameter vector of the known stratum type as prior information to constrain the inversion result.

[0041] The present application proposes an inversion execution method based on stratum type intelligent discrimination and matching, which combines offline classification training and online selection calling, pre-constructs a special inversion model library for different typical stratum structures, and automatically selects the optimal model according to the measurement data characteristics in real-time inversion, solving the problem of insufficient adaptability of traditional methods to complex and variable strata, and improving the generalization ability of the inversion framework.

[0042] Meanwhile, a feature enhancement and reconstruction means for the while-drilling electromagnetic boundary detection response is designed: by extracting and fusing the peak value, change trend and other key physical features in the while-drilling electromagnetic boundary detection measurement response, an enhanced input data with higher information dimension and more sensitive to stratum parameters is constructed, directly improving the resolution ability of the inversion system to weak sensitive parameters such as outer boundary.

[0043] Further, a multiple regularization constraint mode fusing dynamic attenuation and geological prior is proposed, which combines the regularization term with attenuation characteristics and the prior knowledge constraint such as stratum continuity to act on the inversion iteration process together, effectively suppresses the noise interference and multi-solution, and ensures the inversion stability and result interpretability of the algorithm in the complex environment of real well.

[0044] The beneficial effects of the present application compared with the prior art are: the present application forms a multilayer stratum inversion scheme for the while-drilling electromagnetic boundary detection logging instrument through the systematic optimization of "feature enhancement-model optimization-constraint optimization", improves the calculation efficiency of inversion through the rapid iteration of supervised descent method, improves the inversion accuracy through feature enhancement and model classification training, improves the inversion stability through multiple regularization, and improves the physical interpretability in the inversion process through the training driven by physical forward and the constraint of geological prior, thereby providing reliable technical support for real-time high-precision geosteering of complex oil and gas reservoirs. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 It is an example inversion flowchart;

[0046] Figure 2 It is a measured apparent resistivity curve of a real well;

[0047] Figure 3 It is a measured geological signal curve of a real well;

[0048] Figure 4 It is an example stratum inversion result diagram of a real well. DETAILED DESCRIPTION

[0049] The overall inversion flowchart of the example includes two stages of offline preprocessing and training and online real-time inversion, as shown in Figure 1

[0050] Stage one: offline preprocessing and training

[0051] This stage is completed through offline operation before the real-time inversion stage, and the purpose is to construct a training sample set suitable for different stratum types, and to perform feature reconstruction and supervised descent training on the training samples, learn and save the gradient descent set suitable for different stratum structures, and prepare for online real-time inversion.

[0052] Step 1: constructing a stratum structure type according to prior information The constructed stratum structure type contains different stratum structure types to be inverted, and labels are set for different stratum structure types . The total number of stratum structure types .

[0053] The total number of apparent resistivity layers is​ to resistivity layering, For example, according to the prior threshold of apparent resistivity, the apparent resistivity is layered into high-resistivity layer, medium-high-resistivity layer, medium-resistivity layer, medium-low-resistivity layer and low-resistivity layer, and then prior threshold of apparent resistivity , apparent resistivity is divided into , ,…, 5 apparent resistivity layering types according to the prior threshold. Or, through 4 prior thresholds of apparent resistivity, the apparent resistivity is layered into high-resistivity layer, medium-resistivity layer and low-resistivity layer, 3 apparent resistivity layering types.

[0054] to the total number of position area types, position area type sequence number, For example, according to the prior threshold of geological signal , the position area type is divided into 3 position area types according to the relationship between the current position and the layer boundary, i.e. upper boundary area, middle area and lower boundary area, geological signal is divided into , ,…, , position area types according to the prior threshold.

[0055] Then, according to the apparent resistivity area type and the geological signal area type, the layer structure type to which the current layer belongs is set , so that different layer structure types can be constructed according to the prior information.

[0056] Step 2: For different layer types, use forward calculation to generate respective training sample data sets.

[0057] Each sample in the training data set contains:

[0058] 1) layer parameter vector , for example, in the layer layer inversion model, then includes layer resistivity and layer boundary, and in the present application is an integer greater than or equal to 5;

[0059] 2) the corresponding forward original response vector of the electromagnetic boundary detection instrument while drilling , is the transpose; represents the i-th measurement position of the actual instrument a raw response, each raw response including two components: apparent resistivity and geology signals at a specific frequency and source distance; the total number of raw responses measured by the actual instrument at a measurement position, The value is related to the performance of the electromagnetic boundary exploration instrument while drilling.

[0060] Step 3: Response reconstruction of the forward raw response vector in the training data set to increase the response dimension and extract the response features. For each sample of the forward raw response vector Generate enhanced feature response vectors, including but not limited to:

[0061] Extract the peak information of the forward raw response vector as the first enhanced feature:

[0062] ;

[0063] Extract the start and end net change of the forward raw response vector as the second enhanced feature:

[0064] ;

[0065] Extract the average change trend of the forward raw response vector as the third enhanced feature:

[0066] ;

[0067] Extract the convexity of the forward raw response vector as the fourth enhanced feature:

[0068] ;

[0069] Concatenate the forward raw response vector and the first to fourth enhanced features extracted above to obtain an enhanced feature response vector .

[0070] The embodiment maps the low-dimensional abstract measurement data to a high-dimensional feature space that is more sensitive to changes in formation parameters through feature-enhanced response reconstruction, improves the data dimension and detection sensitivity of the measurement response, and improves the defect of insufficient gradient optimization set fitting ability in the original supervised descent inversion method. Through the integration of multiple regularization schemes, the stability of the inversion process and the rationality of the inversion results are guaranteed, and finally a solution to the electromagnetic boundary exploration inversion problem while drilling in multiple layered formations is formed.

[0071] Step 4: Group training of gradient optimization sets dedicated to each formation structure type.

[0072] For different formation structure types of training sample sets, for each type of formation structure, perform supervised descent training respectively:

[0073] Supervised descent inversion method replaces the complex matrix computation in traditional Newton method by training a collection of gradient optimization where, is the Jacobian matrix; is the Hessian matrix, denotes the inverse, is the gradient descent matrix for the th iteration, is the bias term for the th iteration. Assuming that training samples are generated by forward modeling for a certain formation type, given any initial formation parameter vector , the electromagnetic (EM) while-drilling (WD) profile inversion problem for iterations can be converted to the following optimization problem:

[0074]

[0075] where, , is the true measured formation parameter vector for the th training sample, is the formation update parameter for the th training sample at the th iteration, is the formation parameter update direction for the th training sample at the th iteration; the enhanced feature response vector update direction for the th training sample at the th iteration , is the forward modeling algorithm for the EM while-drilling profile instrument, is the enhanced feature response vector for the th training sample.

[0076] Solving this linear least square problem, the gradient descent matrix and the bias term at the th iteration are obtained:

[0077]

[0078] where, denotes the operation of extending to an augmented vector.

[0079] ​​​​Then, the gradient descent matrix corresponding to the above formation structure type and the bias term are saved, and in the actual logging process, the enhanced feature response vector measured in real time is used to update the model.

[0080] Each formation structure type corresponds to a gradient optimization set , and the class formation structure type corresponds to gradient optimization sets, and after collecting the gradient optimization sets corresponding to all formation structure types, phase one ends.

[0081] In phase two, the gradient optimization set corresponding to the formation type saved in the training phase is brought into the inversion, and the formation parameter vector in each iteration of the inversion process can be directly updated:

[0082] ;

[0083] wherein, is the total number of iterations in the supervised descent algorithm, is the gradient descent matrix of the iteration, is the bias term of the iteration, is the iteration number, , is the total number of iterations of the supervised descent algorithm; is the enhanced feature response vector, is the formation parameter vector at the iteration.

[0084] The embodiment realizes high adaptability to any layered formation structure through formation type classification training and intelligent selection of the inversion model, and enhances the generalization ability of the inversion framework to any formation structure.

[0085] Phase two: online real-time inversion

[0086] This phase uses the results of the offline training in phase one to quickly and stably invert the real-time measurement data. The core innovation lies in the inversion strategy of model self-adaptation selection and the multiple regularization scheme.

[0087] Step 1: input the forward original response vector to be inverted as the forward original response vector, the same response reconstruction process as in the offline training phase is adopted to reconstruct the enhanced feature response .

[0088] Step 2: Automatically determine and match the stratum type set in the training stage according to the forward original response vector to be inverted, and determine in real time which stratum structure type the current measurement point belongs to.

[0089] The specific manner is as follows:

[0090] According to the apparent resistivity in the forward original response and the geological signal, the apparent resistivity curve and the geological signal curve are respectively generated;

[0091] According to the apparent resistivity curve, the apparent resistivity layer number matched with the resistivity of the stratum where the electromagnetic boundary exploration instrument while drilling is located is determined ; according to the geological signal curve, the regional type number of the position where the electromagnetic boundary exploration instrument while drilling is located is determined ; so as to determine that the stratum structure type of the current measurement point is .

[0092] Step 3: According to the stratum structure type determination result of the current measurement point, the specific gradient optimization set matched with the stratum type and saved in the training stage is automatically called for inversion iteration, and the stratum parameter in each iteration process is updated and inverted according to the following formula:

[0093] ;

[0094] Step 4: In the iteration and updating process of the stratum parameter in the inversion stage, multiple regularization constraints are introduced:

[0095] ;

[0096] Wherein, represents the L2 norm;

[0097] The first regularization term is a dynamic attenuation regularization term, and the first regularization term coefficient is used to limit the update step, improve the stability and anti-noise ability of the inversion process. The second regularization term is a geological prior constraint term, and the second regularization term coefficient is used to constrain the inversion result by taking the stratum parameter vector of the known stratum type as prior information , so as to ensure the continuity and interpretability of the stratum structure. The embodiment integrates multiple regularization schemes to ensure the stability of the inversion process and the rationality of the inversion result, and finally forms a solution to the electromagnetic boundary exploration inversion problem while drilling in a multi-layer stratum.

[0098] Step 5: Output the stratum parameter vector obtained by inversion as the final stratum model parameter.

[0099] The core means of the embodiments are as follows:

[0100] (1) To solve the problem of low matrix dimension of the gradient optimization set obtained by training, poor generalization ability and poor detection sensitivity, deep feature mining and reconstruction are performed on limited forward original response data to expand the dimension and physical meaning of input information, enhance the detection sensitivity of data to formation parameters (especially low sensitivity parameters), and improve the representation ability and inversion accuracy of the gradient optimization set.

[0101] (2) To solve the problem of low model adaptability and poor accurate optimization ability of the average gradient optimization set trained based on fixed sample data to any actual formation structure, the rigid mode of using a single and general gradient optimization set is abandoned, specific inversion gradient optimization sets are trained for multiple typical formation structures in advance, and in the online real-time stage, the most matched model is intelligently identified and called according to the measured response characteristics. By using the mechanism of "type training and on-demand calling", the supervised descent method can dynamically adapt to different well sections and different types of complex formation structures, instead of relying on a single and general optimization set, thereby improving the accurate inversion ability to different actual formation models.

[0102] (3) To solve the problem of weak noise resistance and insufficient stability of the inversion algorithm, which is difficult to adapt to the inversion requirements of real well data, multiple regularization schemes are introduced in the inversion iteration, and the dynamically decaying regularization ensures fast and stable convergence in the early stage of iteration and fine adjustment in the later stage. The geological prior constraint forces the inversion result to comply with the physical laws of formation continuity and smoothness. The combination of the two effectively suppresses the interference of measurement noise, avoids the oscillation of the inversion process and the generation of unreasonable "pseudo-solutions" in geology, and improves the robustness, stability and reliability of the algorithm in the real well complex environment. The output result is more reliable and interpretable.

[0103] The application effect of the multi-layer formation inversion method of the electromagnetic boundary detection while drilling instrument is as follows:

[0104] Based on the measured response data of the real well, the test analysis is carried out, as shown in Figure 2 The four azimuthal resistivity logging curves while drilling in a real well are shown. The apparent resistivity curve can reflect the overall trend of the resistivity of the surrounding formation when the instrument moves along the well trajectory. As shown in Figure 3 The four geological signal curves while drilling in a real well are shown. The geological signal curve represents the distance change information between the instrument and the surrounding formation interface.

[0105] The formation model obtained based on the same inversion scheme is shown in Figure 4 The comparison of Figure 2 , 3The inversion results are consistent with the resistivity variation details reflected by the apparent resistivity curve and the stratum boundary position revealed by the geological signal curve, which shows that the method is suitable for complex well environments and strongly heterogeneous strata and has good inversion accuracy.

Claims

1. A method for multilayer formation inversion of a LWD EM boundary probe instrument, the method comprising: The method comprises the following steps: Offline training phase: Construct several types of stratigraphic structure, and generate a training sample set containing a stratigraphic parameter vector and an original forward response vector for each stratigraphic structure type; wherein the formation parameter vector includes a formation resistivity and a formation boundary, is an integer greater than or equal to 5; the formation structure type is constructed based on a combination of a resistivity layering type and a location area type based on geological signals; the forward original response vector includes a resistivity and a geological signal at a specific frequency and offset. Feature enhancement and reconstruction are performed on the original forward response vector of each training sample to generate an enhanced feature response vector containing original response and enhanced features; the feature enhancement and reconstruction are to extract four specific features of the peak value information, the start-stop net change, the average change trend and the convexity degree of the original forward response vector, and splice them with the original response vector; For each stratigraphic structure type, a set of gradient optimization corresponding to each stratigraphic structure type is trained by using the training sample set through a supervised descent method, and each set of gradient optimization contains a gradient descent matrix and a bias term for each iteration; Online inversion phase: Obtain the original forward response vector measured in real time, and perform feature enhancement and reconstruction to obtain an enhanced feature response vector; at the same time, generate a apparent resistivity curve and a geological signal curve according to the apparent resistivity and the geological signal of the source distance in the original forward response vector, and then analyze the apparent resistivity curve and the geological signal curve to determine the corresponding type number, and then determine the stratigraphic structure type to which the current measurement point belongs; Call a set of gradient optimization corresponding to the stratigraphic structure type matched with the current measurement point, and use the gradient optimization set and the enhanced feature response vector to perform inversion iteration on the stratigraphic parameter vector; after the inversion iteration is completed, the final stratigraphic model parameter is output.

2. The multi-layer formation inversion method of the electromagnetic edge-finding instrument as described in claim 1, characterized in that, forward original response vector generating an enhanced feature response vector, comprising: The peak value information of the original forward response vector is extracted as the first enhanced feature: ； The start-stop net change of the original forward response vector is extracted as the second enhanced feature: ； The average change trend of the original forward response vector is extracted as the third enhanced feature: ； The convexity degree of the original forward response vector is extracted as the fourth enhanced feature: ； wherein, represents the number of total original responses measured at a measurement location, represents the number of total original responses measured at a measurement location, is the number of total original responses measured at a measurement location by the electromagnetic shoulder probe while drilling instrument. The forward original response vector is spliced with the first to fourth enhanced features extracted above to obtain an enhanced feature response vector .

3. The multi-layer formation inversion method of the electromagnetic edge-finding instrument as described in claim 1, characterized in that, The iteration update formula for using the gradient optimization set and the enhanced feature response vector to perform inversion iteration on the stratigraphic parameter vector is: ； wherein is the gradient descent matrix for the th iteration, is the bias term for the th iteration, is the iteration number, , is the total number of iterations for the supervised descent algorithm; is the enhanced feature response vector, is the formation parameter vector at the th iteration.

4. The method of claim 3, wherein the multi-layer formation inversion of the LWD EM boundary probe tool is performed by solving the following equation: ###0001### where: ###0002### and ###0003### are the unknowns, and ###0004### is the known value of the apparent conductivity. 5 In the offline training phase, a set of gradient optimization corresponding to each formation structure type is obtained by solving the following least square problem : ； in, , For the actual measurement of the first The formation parameter vector of each training sample For the first During the nth iteration The formation parameter vector of each training sample For the first During the nth iteration The formation parameter update direction of the training sample; During the nth iteration The direction of updating the enhanced feature response vector of each training sample. , For the first The enhanced feature response vector of each training sample. Forward modeling calculations for the electromagnetic edge detection instrument while drilling, This represents the L2 norm.

5. The multi-layer formation inversion method of the electromagnetic edge-finding instrument as described in claim 4, characterized in that, In the online inversion phase, multiple regularization constraints are introduced during inversion iteration: ； wherein, is the first updated direction of the formation parameters at the nth iteration; a first regularization term is a dynamic attenuation regularization term, is a first regularization term coefficient, used to limit the update step size, improve the stability and anti-noise ability of the inversion process; a second regularization term is a geological prior constraint term, is a second regularization term coefficient, using the known stratum parameter vector of the stratum type as prior information to constrain the inversion result, ensure the continuity and interpretability of the stratum structure.

Images