Ultra-deep while-drilling electromagnetic wave logging method based on deep learning

Abstract

The invention relates to the technical field of oil and gas exploration logging, in particular to an ultra-deep electromagnetic wave logging-while-drilling method based on deep learning. Comprising the following steps: (1) multi-source data acquisition: acquiring an electromagnetic wave original signal of an ultra-deep stratum through a while-drilling electromagnetic wave detection device, and synchronously acquiring resistivity, sound wave and density logging data and drilling real-time monitoring data at the same time to form a multi-modal logging data set; (2) data preprocessing: performing noise suppression, abnormal value elimination and normalization processing on the acquired multi-modal data, performing targeted noise reduction on an electromagnetic wave signal by adopting a bidirectional LSTM recurrent neural network, and converting logging data with azimuth information into normalized data under geodetic coordinates through coordinate conversion; the method has the beneficial effects that the inversion precision is remarkably improved, geological physical constraints and a quantum state dynamic coding mechanism are fused through the GeoResNet-1D model, the transverse positioning error is reduced to 0.4 m, the vertical signal resolution reaches 0.5 m, and the complex structure recognition rate is improved to 89%.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration logging technology, specifically to an ultra-deep electromagnetic logging method based on deep learning. Background Technology

[0002] Exploration of ultra-deep oil and gas resources is a crucial direction for ensuring energy security. However, ultra-deep formations are characterized by complex geological structures, high pressure and temperature, and severe signal attenuation, posing numerous challenges to traditional electromagnetic logging-while-drilling (EML) methods. Traditional logging inversion methods rely on numerical simulations and human experience, making it difficult to handle strongly coupled multi-parameter and nonlinear problems, resulting in low inversion accuracy. Lateral positioning errors are typically above 2.1 meters, and vertical resolution is only around 5 meters. Furthermore, traditional methods lack the ability to fuse multi-modal logging data, failing to fully utilize the complementary information from other logging data, and their slow signal processing speed cannot meet the millisecond-level decision-making requirements during drilling. In addition, the complex geological structures of ultra-deep formations (such as salt domes and faults) easily lead to signal distortion, further reducing the analytical accuracy of logging data and hindering the efficient exploration and development of ultra-deep oil and gas resources.

[0003] To address the aforementioned issues, a deep learning-based method for ultra-deep drilling electromagnetic wave logging is proposed. Summary of the Invention

[0004] To address the problems of low inversion accuracy, poor real-time performance, and weak adaptability to complex geological conditions in existing ultra-deep drilling electromagnetic wave logging methods, this invention provides an ultra-deep drilling electromagnetic wave logging method based on deep learning. By using multi-modal data fusion and deep learning technology, it achieves high-precision and rapid inversion of ultra-deep formation parameters, providing reliable support for drilling guidance.

[0005] The present invention provides a deep learning-based method for ultra-deep drilling electromagnetic wave logging, which specifically includes the following steps:

[0006] A deep learning-based method for ultra-deep drilling electromagnetic wave logging, characterized by the following steps:

[0007] (1) Multi-source data acquisition: The original electromagnetic wave signals of ultra-deep formations (depth ≥ 3000 meters) are acquired by the drilling electromagnetic wave detection equipment, and resistivity, acoustic wave, density logging data and drilling real-time monitoring data are acquired simultaneously to form a multi-modal logging dataset.

[0008] (2) Data preprocessing: Noise suppression, outlier removal and normalization are performed on the collected multimodal data. A bidirectional LSTM recurrent neural network is used to perform targeted noise reduction on the electromagnetic wave signal. The well logging data with azimuth information is transformed into normalized data in geodetic coordinates through coordinate transformation.

[0009] (3) Deep learning model construction and training: Construct a GeoResNet-1D deep learning model that integrates geophysical constraints. The model includes a one-dimensional residual block of geophysical constraints and a quantum state dynamic encoding module. The preprocessed multimodal data is used as training samples, and the actual stratigraphic parameter measurement values ​​are used as labels. The model is trained until convergence.

[0010] (4) Real-time logging inversion: Input the real-time collected and preprocessed multimodal logging data into the trained GeoResNet-1D model, and obtain key parameters such as formation resistivity, porosity, permeability and hydrocarbon content through model inference inversion. The inversion response time is ≤8ms.

[0011] (5) Results output and visualization: The visualization interface built on the VC++MFC class library displays the results of formation parameter inversion and three-dimensional dynamic images of formation structure in real time, and supports data query, curve drawing and result saving.

[0012] As a further aspect of the present invention, the original electromagnetic wave signal in step (1) is acquired by sensors distributed in a 360° circle, containing four electromagnetic wave data with azimuth information. The sampling frequency is adapted to the signal attenuation characteristics of ultra-deep strata to ensure signal integrity.

[0013] As a further aspect of the present invention, the bidirectional LSTM recurrent neural network in step (2) includes an input gate, a forget gate, and an output gate. By learning the time series characteristics of electromagnetic square wave noise, it achieves precise noise suppression and improves the signal-to-noise ratio by ≥31%.

[0014] As a further embodiment of the present invention, the one-dimensional residual block of the geophysical constraint in step (3) adopts a dual-path residual structure, including a signal feature path and a physical constraint path, wherein the physical constraint path is embedded in the discretization layer of the differentiable Maxwell equation, and the model output is forced to satisfy the electromagnetic wave propagation law.

[0015] As a further aspect of the present invention, the model training process in step (3) adopts a dynamic weight fusion gate, which adjusts the dual-path output fusion ratio according to the formation density and conductivity parameters. When the formation non-uniformity is significant, the weight of the physical constraint path is automatically increased to more than 0.7.

[0016] As a further aspect of the present invention, the GeoResNet-1D model described in step (4) is designed to be lightweight, with the number of parameters controlled to 3.2M, which meets the computing power limitations of downhole edge equipment, with a lateral positioning error of ≤0.4 meters and a vertical signal resolution of 0.5 meters.

[0017] As a further embodiment of the present invention, the visualization interface described in step (5) supports real-time rendering of multi-source data, achieves stratum imaging through two-dimensional interpolation algorithm and static color calibration, and supports scaling, translation and curve overlay comparison operations.

[0018] The beneficial effects of this invention are:

[0019] Significantly improved inversion accuracy: By integrating geophysical constraints and quantum state dynamic coding mechanism with GeoResNet-1D model, the lateral positioning error is reduced to 0.4 meters, the vertical signal resolution reaches 0.5 meters, and the recognition rate of complex structures is improved to 89%, which is a significant performance improvement compared with traditional methods;

[0020] Real-time performance meets engineering requirements: The model is lightweight with only 3.2M parameters and an inference response time of ≤8ms, supporting millisecond-level decision-making in downhole drilling, thus solving the problem of slow processing speed of traditional methods;

[0021] Strong adaptability to complex geology: Through dynamic weighted fusion gate and deformable convolution technology, it can adapt to the non-homogeneity of ultra-deep strata and signal distortion, and the accuracy of anomaly detection reaches 92%;

[0022] High utilization of multi-source data: It effectively integrates electromagnetic wave signals with multi-modal data such as resistivity and acoustic waves, and improves the reliability of inversion through data complementarity, with a multi-level prediction consistency of 94%;

[0023] High engineering application value: The visual interface is easy to operate, supports real-time monitoring and decision support, can reduce the number of invalid wells, reduce exploration costs and risks, and provide technical support for the development of ultra-deep oil and gas resources.

[0024] Obviously, based on the above description of the present invention, and according to common technical knowledge and conventional methods in the field, various other modifications, substitutions or alterations can be made without departing from the basic technical concept of the present invention.

[0025] The following detailed embodiments further illustrate the above-described content of the present invention. However, this should not be construed as limiting the scope of the present invention to the following examples. All technologies implemented based on the above-described content of the present invention fall within the scope of the present invention. Detailed Implementation

[0026] The present invention is illustrated below with specific embodiments, which are not intended to limit the scope of the invention.

[0027] A deep learning-based method for ultra-deep drilling electromagnetic wave logging specifically includes the following steps:

[0028] Multi-source data acquisition: Raw electromagnetic wave signals from ultra-deep formations (depth ≥ 3000 meters) are acquired using an electromagnetic wave detection device while drilling. The sensor adopts a 360° circular layout to simultaneously acquire four electromagnetic wave data with azimuth information. At the same time, auxiliary logging data such as resistivity, acoustic waves, and density are acquired, as well as real-time monitoring data such as depth and tool face angle during the drilling process, forming a multi-modal logging dataset to ensure that the data covers formation physical characteristics and drilling conditions.

[0029] Data preprocessing: A bidirectional LSTM recurrent neural network is used to suppress noise in the raw electromagnetic wave signal. The gating mechanism is used to fully learn the information before and after the noise sequence, which improves the signal-to-noise ratio by ≥31%. The well logging data containing azimuth information is transformed into normalized data in geodetic coordinates. Through outlier removal and normalization, the data format and scale are unified to provide a high-quality data foundation for model input. Two-dimensional interpolation algorithm is used to enrich the electromagnetic wave data in the well perimeter and well axis directions, forming a large two-dimensional matrix for subsequent imaging.

[0030] Deep Learning Model Construction and Training: Constructing the GeoResNet-1D deep learning model, the core of which includes geophysically constrained one-dimensional residual blocks and quantum state dynamic encoding modules:

[0031] Geophysical constraint one-dimensional residual block: adopts a dual-path structure. The signal feature path captures local temporal features through 1D dilated convolution, while the physical constraint path is embedded in a discretization layer of differentiable Maxwell's equations to ensure that the model output conforms to the laws of electromagnetic wave propagation.

[0032] Model training: Using preprocessed multimodal data as input and actual formation parameter measurements as labels, a dynamic weight fusion gate is used to adjust the dual-path fusion ratio. During training, regularization techniques and data augmentation strategies are used to avoid overfitting and improve the model's generalization ability.

[0033] Real-time logging inversion: The multimodal data acquired and preprocessed in real time is input into the trained GeoResNet-1D model. The model is adapted to downhole edge equipment through lightweight design, with an inference response time of ≤8ms. Key parameters such as formation resistivity, porosity, permeability and hydrocarbon content are obtained through inversion. The lateral positioning error is ≤0.4 meters, the vertical signal resolution reaches 0.5 meters, and the recognition rate of complex structures is improved to 89%.

[0034] Results Output and Visualization: The inversion results are displayed in real time through a visualization interface built on the VC++ MFC class library. Static color calibration is used to assign RGB colors to each imaging data to achieve three-dimensional dynamic imaging of the formation structure. The interface supports data query, real-time curve drawing, historical curve analysis and result saving functions. Users can accurately view the details of data at different depths through zoom and pan operations, providing intuitive support for drilling guidance decisions.

[0035] Example 1

[0036] Specific implementation steps of a deep learning-based ultra-deep drilling electromagnetic wave logging method:

[0037] Multi-source data acquisition: In ultra-deep drilling operations, electromagnetic wave signals at depths of 3000-5000 meters are acquired using a drilling electromagnetic wave detection device with a sampling frequency of 2500Hz. Simultaneously, resistivity logging data (measurement range 1-200Ω・m), acoustic transit time data, and parameters such as drilling depth and tool face angle are acquired to form a multi-modal dataset.

[0038] Data preprocessing: A bidirectional LSTM recurrent neural network was used to denoise the electromagnetic wave signal. The network layer was set to 3 layers and the number of hidden units was 256. After denoising, the signal-to-noise ratio was improved by 35%. The well logging data with azimuth information was subjected to coordinate transformation. After outlier removal and normalization, a standardized data matrix was formed.

[0039] Model Training: A GeoResNet-1D model was constructed. The one-dimensional residual blocks with geophysical constraints were subjected to 1D dilated convolutions with a kernel length of 7 and an expansion rate of 3. The physical constraint paths were embedded in the Maxwell equation discretization layer. The quantum state dynamic encoding module encoded the four signal levels into four-dimensional orthogonal basis vectors. The model converged after 68 epochs of training using measured data from 100 ultra-deep wells as the training set. The dual-path fusion ratio was dynamically adjusted during the training process.

[0040] Real-time inversion: The multimodal data collected in real time is input into the model, which runs on an NVIDIA Jetson edge device to invert parameters such as formation resistivity and porosity. The lateral positioning error is 0.38 meters and the vertical resolution is 0.5 meters.

[0041] Results visualization: The visualization interface displays the formation imaging results in real time, supports the superposition and comparison of resistivity curves and electromagnetic wave signal curves, and allows operators to adjust the drilling trajectory based on the formation structure information displayed on the interface to ensure accurate drilling into oil reservoirs.

[0042] This invention effectively solves the problems of low accuracy and poor real-time performance in ultra-deep electromagnetic wave logging inversion by using multimodal data fusion and deep learning technology, providing a reliable technical means for ultra-deep oil and gas resource exploration and having broad engineering application prospects.

[0043] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0044] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made using the contents of the present invention under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based method for ultra-deep drilling electromagnetic wave logging, characterized in that, Includes the following steps: (1) Multi-source data acquisition: The original electromagnetic wave signals of ultra-deep formations (depth ≥ 3000 meters) are acquired by the drilling electromagnetic wave detection equipment, and resistivity, acoustic wave, density logging data and drilling real-time monitoring data are acquired simultaneously to form a multi-modal logging dataset. (2) Data preprocessing: Noise suppression, outlier removal and normalization are performed on the collected multimodal data. A bidirectional LSTM recurrent neural network is used to perform targeted noise reduction on the electromagnetic wave signal. The well logging data with azimuth information is transformed into normalized data in geodetic coordinates through coordinate transformation. (3) Deep learning model construction and training: Construct a GeoResNet-1D deep learning model that integrates geophysical constraints. The model includes a one-dimensional residual block of geophysical constraints and a quantum state dynamic encoding module. The preprocessed multimodal data is used as training samples, and the actual stratigraphic parameter measurement values ​​are used as labels. The model is trained until convergence. (4) Real-time logging inversion: Input the real-time collected and preprocessed multimodal logging data into the trained GeoResNet-1D model, and obtain key parameters such as formation resistivity, porosity, permeability and hydrocarbon content through model inference inversion. The inversion response time is ≤8ms. (5) Results output and visualization: The visualization interface built on the VC++MFC class library displays the results of formation parameter inversion and three-dimensional dynamic images of formation structure in real time, and supports data query, curve drawing and result saving.

2. The method according to claim 1, characterized in that, The original electromagnetic wave signal mentioned in step (1) is collected by sensors distributed in a 360° circle. It contains four electromagnetic wave data with azimuth information. The sampling frequency is adapted to the signal attenuation characteristics of ultra-deep strata to ensure signal integrity.

3. The method according to claim 1, characterized in that, The bidirectional LSTM recurrent neural network described in step (2) includes an input gate, a forget gate, and an output gate. By learning the time series characteristics of electromagnetic square wave noise, it achieves precise noise suppression and improves the signal-to-noise ratio by ≥31%.

4. The method according to claim 1, characterized in that, The one-dimensional residual block with geological and physical constraints described in step (3) adopts a dual-path residual structure, including a signal feature path and a physical constraint path. The physical constraint path is embedded in the discretization layer of the differentiable Maxwell equations, which forces the model output to satisfy the electromagnetic wave propagation law.

5. The method according to claim 4, characterized in that, In step (3), the model training process uses a dynamic weight fusion gate to adjust the dual-path output fusion ratio according to the formation density and conductivity parameters. When the formation non-uniformity is significant, the weight of the physical constraint path is automatically increased to more than 0.

7.

6. The method according to claim 1, characterized in that, The GeoResNet-1D model described in step (4) is designed to be lightweight, with the number of parameters controlled to 3.2M, which meets the computing power limit of downhole edge equipment, with a lateral positioning error of ≤0.4 meters and a vertical signal resolution of 0.5 meters.

7. The method according to claim 1, characterized in that, The visualization interface described in step (5) supports real-time rendering of multi-source data, achieves stratum imaging through two-dimensional interpolation algorithm and static color calibration, and supports scaling, translation and curve overlay comparison operations.