Salt and salt layer distortion curve reconstruction method and system based on intelligent algorithm

Abstract

The application discloses a salt interlayer and salt underlayer stratum distortion curve reconstruction method and system based on an intelligent algorithm, relates to the technical field of oil and gas well logging data processing, and comprises the following steps: a training data set containing multi-scale physical disturbance is constructed; a mixed reconstruction model is formed by adopting a staged training mode; echo detection signals of a target well section are collected, hierarchical cascade analysis is carried out based on the mixed reconstruction model, and a reconstruction result is determined; and the reconstruction result is displayed on a terminal interface. The application solves the technical problems that in the prior art, in the reconstruction of a salt interlayer and salt underlayer stratum logging curve, it is difficult to comprehensively cover the complex interference scene of salt rock, the accuracy of the reconstruction result is insufficient, and the adaptability to geological rules is poor, achieves accurate reconstruction and visual presentation of a distorted logging curve of a salt interlayer and salt underlayer stratum disturbed by salt rock, and improves the accuracy of curve reconstruction and the geological adaptability.

Description

Technical Field

[0001] This invention relates to the field of oil and gas logging data processing technology, specifically to a method and system for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms. Background Technology

[0002] In the exploration and development of inter- and sub-salt formations, well logging curves are the core basis for obtaining key information such as formation lithology, physical properties, and hydrocarbon potential. However, the complex geometric morphology, special electrical characteristics, and complex contact relationship between salt rock bodies and surrounding rocks can easily lead to distortion of the well logging signal propagation path and distortion of response characteristics, which seriously affects the interpretation accuracy of well logging curves. Existing reconstruction methods mostly rely on simulation of single interference factors or traditional machine learning models, which have problems such as training samples being unable to fully cover the multi-dimensional complex interference scenarios of salt rocks, lack of geological knowledge constraints in the models, and reconstruction results being prone to numerical deviations or violations of geological laws. These methods cannot meet the needs of high-precision formation evaluation and exploration and development decision-making.

[0003] The existing technology has technical problems in reconstructing logging curves in inter- and sub-salt formations, which makes it difficult to fully cover complex interference scenarios in salt rocks, resulting in insufficient accuracy of reconstruction results and poor adaptability to geological laws. Summary of the Invention

[0004] This application provides a method and system for reconstructing distortion curves of inter-salt and sub-salt formations based on intelligent algorithms. It is used to address the technical problems in the existing technology of reconstructing logging curves of inter-salt and sub-salt formations, which is difficult to fully cover complex interference scenarios of salt rocks, and the reconstruction results are not accurate enough and have poor adaptability to geological laws.

[0005] In view of the above problems, this application provides a method and system for reconstructing distortion curves of intersalt and subsalt formations based on intelligent algorithms.

[0006] The first aspect of this application provides a method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms, the method comprising:

[0007] A training dataset containing multi-scale physical perturbations is constructed. This involves generating enhanced training samples covering the real salt rock interference spectrum by establishing a continuous parameterized space for the well logging response distortion field. A hybrid reconstruction model is constructed by configuring a hybrid deep learning network and employing a staged training approach. This hybrid deep learning network includes a one-dimensional convolutional layer and a main reconstruction network. Echo detection signals from the target well section are collected, and hierarchical cascade analysis is performed based on the hybrid reconstruction model to determine the reconstruction results. The reconstruction results are then displayed on a terminal interface.

[0008] A second aspect of this application provides a system for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms, the system comprising:

[0009] The training dataset construction module is used to construct a training dataset containing multi-scale physical perturbations. This involves generating enhanced training samples covering the real salt rock interference spectrum by establishing a continuous parameterized space for the well logging response distortion field. The hybrid reconstruction model construction module is used to construct a hybrid reconstruction model by configuring a hybrid deep learning network and employing a staged training method. The hybrid deep learning network includes a one-dimensional convolutional layer and a main reconstruction network. The reconstruction result determination module is used to determine the reconstruction result by collecting echo detection signals from the target well section and performing hierarchical cascade analysis based on the hybrid reconstruction model. The terminal interface display module is used to display the reconstruction result on a terminal interface.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] A training dataset incorporating multi-scale physical perturbations was constructed. A hybrid reconstruction model was formed by configuring a hybrid deep learning network and employing a phased training approach. Echo detection signals from the target well section were collected, and hierarchical cascade analysis was performed based on the hybrid reconstruction model to determine the reconstruction results. The reconstruction results were then displayed on a terminal interface. This method achieves accurate reconstruction and visualization of distorted logging curves from inter- and sub-salt formations affected by salt rock interference, improving the accuracy and geological adaptability of the curve reconstruction. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A schematic diagram of the process for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms, provided in an embodiment of this application;

[0014] Figure 2 A schematic diagram of the structure of the intelligent algorithm-based distortion curve reconstruction system for intersalt and subsalt formations provided in this application embodiment.

[0015] Figure labeling: Training dataset construction module 10, hybrid reconstruction model construction module 20, reconstruction result determination module 30, terminal interface display module 40. Detailed Implementation

[0016] This application provides a method and system for reconstructing distortion curves of inter-salt and sub-salt formations based on intelligent algorithms. This method addresses the technical problems in existing technologies where well logging curve reconstruction of inter-salt and sub-salt formations is difficult to fully cover complex interference scenarios in salt rocks, resulting in insufficient accuracy of reconstruction results and poor adaptability to geological patterns.

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0018] Example 1, as Figure 1 As shown, this application provides a method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms, the method comprising:

[0019] Step S100: Construct a training dataset containing multi-scale physical perturbations, wherein enhanced training samples covering the real salt rock interference spectrum are generated by establishing a continuous parameterized space of the well logging response distortion field.

[0020] Specifically, when constructing a training dataset containing multi-scale physical disturbances, a set of salt rock disturbance parameters is first defined, which covers at least the geometric morphology parameters, electrical parameters, and contact relationship parameters between the salt rock and the surrounding rock. Then, a continuous mapping relationship of disturbance intensity is established to map the multi-dimensional salt rock disturbance parameters to the one-dimensional disturbance intensity spectrum, forming a continuous parameterization space for the logging response distortion field that can be continuously interpolated. Finally, based on the parameter set and the continuous mapping relationship, enhanced training samples covering the real salt rock disturbance spectrum are generated and incorporated into the training dataset to ensure that the dataset can fully simulate different salt rock disturbance scenarios, providing rich data support that fits the actual geological conditions for the training of the subsequent hybrid reconstruction model.

[0021] Step S200: By configuring a hybrid deep learning network and adopting a staged training method, a hybrid reconstruction model is formed, wherein the hybrid deep learning network includes a one-dimensional convolutional layer and a main reconstruction network.

[0022] Specifically, a high-performance hybrid reconstruction model is constructed by configuring a well-structured, functionally layered hybrid deep learning network and employing a multi-stage refined training approach. The hybrid deep learning network uses a one-dimensional convolutional layer as the front-end feature extraction unit, coupled with a main reconstruction network consisting of a three-layer fully connected encoder, a fused attention layer, and a decoder. The decoder embeds a differentiable simplified well-logging physical response equation as a soft constraint. A dual-discriminator architecture is also introduced. The first discriminator is responsible for judging the distortion level of the reconstruction curve, while the second discriminator generates a simulated distortion curve in reverse and compares it with the original input distortion curve, achieving bidirectional verification. This forms a game-theoretic training mechanism with the hybrid deep learning network until… The model converged; the training process strictly followed a three-stage process of pre-training-main training-fine-tuning. In the pre-training stage, unlabeled pre-training under masked sequence modeling was used to mine the general feature representation of the formation data. In the main training stage, the encoder parameters were fixed, and paired samples of the training dataset were used to focus on training the fusion attention layer, decoder and dual discriminator. A weighted hybrid loss function was introduced, which included point-to-point mean square error loss and morphological loss based on dynamic time warping, to improve the model fitting effect. In the fine-tuning stage, local adaptive fine-tuning of well sections was carried out with the help of enhanced training samples to adapt the model to the detailed features of specific geological scenes, and finally a stable hybrid reconstruction model with both generalization ability and local adaptability was formed.

[0023] Step S300: By acquiring the echo detection signal of the target well section, perform hierarchical cascade analysis based on the hybrid reconstruction model to determine the reconstruction result.

[0024] Specifically, the target well section is first probed to collect echo detection signals, and then the corresponding spectral curves are extracted and input into the trained hybrid reconstruction model. During the model operation, a connection is first established between the geological knowledge graph and the fusion attention layer in the main reconstruction network. By interacting with the geological knowledge graph, structured knowledge such as lithological prior probabilities and sedimentary facies labels at each reconstruction depth point of the target well section is obtained and incorporated as enhancement features into the feature fusion stage of the fusion attention layer. Subsequently, feature attention fusion and sequence reconstruction under stratigraphic multimodal element encoding are performed. Simultaneously, a reconstruction credibility matrix is ​​generated in real time through an uncertainty propagation network. The initial output includes the true response curve of the stratigraphy in the target well section, the salt rock interference component curve, and the reconstruction credibility matrix. Then, relying on the geological rule base containing structured knowledge such as stratigraphic compaction trends, lithological-physical property correlations, sedimentary cycle characteristics, and regional geological laws, the initial reconstruction results are scanned point by point at each depth to detect abnormal sections that violate geological rules. Then, the abnormal sections are corrected by using matched geological rules and contextual information through weighted smoothing and trend fitting, and finally, accurate and reliable reconstruction results are determined.

[0025] Step S400: Display the reconstruction result on the terminal interface.

[0026] Specifically, the final reconstruction results obtained after hierarchical cascade analysis and anomaly correction are visualized on a terminal interface. The displayed content comprehensively covers the actual formation response curve of the target well section, the salt rock interference component curve, and the reconstruction credibility matrix generated in real time through the uncertainty propagation network. At the same time, the interface will clearly present the depth dimension, numerical range, and geological correlation information corresponding to each curve, such as the lithological prior probability and the formation location label corresponding to the sedimentary facies label. This allows staff to intuitively compare the distribution characteristics of the actual formation response and the salt rock interference, and quickly check the reliability of the reconstruction results at different depth points.

[0027] In one possible implementation, step S200 further includes:

[0028] The hybrid reconstruction model uses a hybrid deep learning network as its model architecture. The main reconstruction network includes a three-layer fully connected encoder, a fused attention layer, and a decoder. The decoder is embedded with a differentiable simplified logging physical response equation as a soft constraint.

[0029] During the training of the hybrid deep learning network, a discriminator is introduced to engage in game-theoretic training with the hybrid deep learning network until convergence.

[0030] Specifically, the hybrid reconstruction model uses a modular and hierarchical hybrid deep learning network as its core architecture, forming a complete technical chain of "front-end feature extraction - core feature processing - reconstruction output constraints". Its structure and model details are disclosed as follows: The model architecture consists of a one-dimensional convolutional layer and a main reconstruction network working together. The one-dimensional convolutional layer, as the front-end feature extraction unit, focuses on capturing local features and dimensionality regularization of the input well logging-related data, such as distortion curves in training samples and spectral curves of target well sections. It extracts local correlations and texture features from the data through convolutional kernel sliding operations, providing high-quality basic features for subsequent core processing steps. The main reconstruction network, as the core operation module of the model, adopts a serial structure of encoder-fusion attention layer-decoder to realize the entire process of feature extraction, enhancement and restoration. The encoder employs a three-layer fully connected network structure. The first fully connected layer receives the basic features output from the one-dimensional convolutional layer and performs initial dimensionality compression and feature integration. The second fully connected layer further mines the deep relationships between features. The third fully connected layer completes the final output of high-dimensional abstract features, achieving the transformation from raw features to core abstract features through a progressive layer-by-layer approach. An attention layer is integrated as a key unit for feature enhancement. On one hand, it receives the abstract features output from the encoder; on the other hand, it can establish interactive connections with the geological knowledge graph, introducing structured knowledge such as lithological prior probabilities and sedimentary facies labels as enhanced features. Through an attention weight allocation mechanism, key features, such as those strongly correlated with the true stratigraphic response and features associated with geological knowledge, are further enhanced. Row weight enhancement suppresses redundant features, improving the relevance and effectiveness of feature representation. The decoder also employs a fully connected network structure. Its core function is to inversely restore the enhanced abstract features into a reconstructed curve consistent with the original data dimension. The decoder embeds a differentiable simplified logging physical response equation as a soft constraint. This equation is derived based on the physical principles of logging and is incorporated into the backpropagation process of the decoder through differentiable design. This ensures that the model not only follows the data-level fitting rules during reconstruction but also strictly conforms to the logging physical response mechanism of inter- and sub-salt formations. This effectively avoids physical inconsistencies in the reconstruction results that deviate from geological reality, ensuring that the output reconstructed curve meets both numerical accuracy requirements and conforms to stratigraphic geological laws. The entire hybrid reconstruction model, through the synergistic effect of its modules, forms a reconstruction system that combines data fitting capabilities with physical constraint characteristics, providing solid model support for the accurate reconstruction of distorted curves in inter- and sub-salt formations.

[0031] During the training of the hybrid deep learning network, a dual-discriminator architecture consisting of a first discriminator and a second discriminator is specifically introduced to form a bidirectional game-like training mechanism with the hybrid deep learning network. Through multiple rounds of adversarial iteration, the model is gradually driven to converge. The first discriminator focuses on judging the degree of distortion of the model's output reconstruction curve. By learning the feature distribution of the real stratum response curve, it verifies the authenticity and rationality of the reconstruction curve and feeds back gradient information. The second discriminator undertakes the reverse verification function. It reverse-engineers the simulated distortion curve from the reconstruction curve generated by the model, compares the features of the simulated distortion curve with the original input distortion curve in the training process, and calculates the error, thereby realizing the bidirectional verification of the model's interference removal-reconstruction logic. During game-theoretic training, the hybrid deep learning network aims to generate high-quality reconstruction results that pass the dual discriminator verification. It continuously adjusts the parameters of the one-dimensional convolutional layer, the main reconstruction network (encoder, fusion attention layer, and decoder), while the dual discriminators continuously optimize their discrimination criteria, improving their ability to distinguish between real and reconstructed curves, and between original and simulated distortion curves. This creates a closed-loop adversarial process of generation-discrimination-optimization. Simultaneously, a phased training process of pre-training, main training, and fine-tuning is employed. In the main training phase, a weighted hybrid loss function is introduced, incorporating point-to-point mean squared error loss and morphological loss based on dynamic time warping. This further constrains the direction of game-theoretic training until the reconstructed curves generated by the hybrid deep learning network can stably pass the dual discriminator verification, the model loss value stabilizes, and the game-theoretic training converges. This ensures that the final hybrid reconstruction model possesses both accurate curve reconstruction capabilities and reliable distortion correction effects.

[0032] In one possible implementation, step S220 further includes:

[0033] Step S221: The discriminator includes a first discriminator and a second discriminator.

[0034] The first discriminator is used to determine the distortion of the model reconstruction curve; the second discriminator is used to generate a simulated distortion curve from the model reconstruction curve and compare the simulated distortion curve with the original input distortion curve.

[0035] Specifically, the discriminator adopts a dual-discriminator collaborative architecture, comprising a first discriminator and a second discriminator. The training process utilizes bidirectional verification along both the forward and reverse paths to precisely constrain the hybrid deep learning network, thereby improving training convergence. The forward path follows a logical flow: distortion curve – hybrid reconstruction model – reconstruction curve – first discriminator. The original distortion curve from the training dataset is input into the hybrid reconstruction model. After the model outputs the corresponding reconstruction curve, the first discriminator verifies its authenticity and reasonableness. By learning the characteristic distribution patterns of the real formation response curve, the first discriminator determines whether the reconstruction curve closely matches the actual physical response of the formation and whether there are any distortion deviations. This provides gradient feedback to the hybrid deep learning network, driving the model to optimize reconstruction accuracy. The reverse path follows a sequence: reconstructed curve - second discriminator - simulated distortion curve - comparison with the original input distortion curve. The second discriminator receives the reconstructed curve output by the hybrid reconstruction model and generates a simulated distortion curve through reverse derivation. This simulated distortion curve is then compared with the distortion curve of the original input during training, and errors are calculated to complete the consistency check. By verifying the matching degree between the reverse-generated simulated distortion curve and the original distortion curve, the integrity and reliability of the hybrid reconstruction model's interference removal and reconstruction logic are validated, avoiding insufficient generalization ability caused by one-way fitting. Through the bidirectional linkage of the authenticity check of the forward path and the consistency check of the reverse path, the dual discriminators form a comprehensive constraint on the hybrid reconstruction model, effectively improving the convergence speed and stability of model training, ensuring that the final trained model possesses both accurate reconstruction ability and good generalization performance.

[0036] In one possible implementation, step S220 further includes:

[0037] The training process includes a pre-training phase, a main training phase, and a fine-tuning phase.

[0038] In the pre-training phase, a general representation of the strata is learned through unlabeled pre-training under masked sequence modeling.

[0039] During the main training phase, the encoder parameters are fixed, and the fusion attention layer, decoder, and discriminator are trained using paired samples from the training dataset.

[0040] During the fine-tuning phase, the generated enhanced training samples are used to perform local adaptive fine-tuning and reconstruction training of the well section.

[0041] Specifically, the training process of hybrid deep learning networks strictly follows a three-stage progressive process: pre-training stage, main training stage, and fine-tuning stage. By focusing on different training objectives and optimizing training data and parameter configurations in stages, the model is ensured to gradually form a reconstruction performance that combines generalization ability, accurate fitting ability, and local adaptability.

[0042] In the pre-training stage of the hybrid deep learning network, the core employs a masked sequence modeling-driven unlabeled pre-training method to uncover the inherent patterns and general feature representations of formation data, laying the foundation for subsequent refined training. This stage does not rely on manually labeled distorted curve-true curve pairing samples. Instead, it uses a large amount of unlabeled well logging data from intersalt and subsalt formations, covering raw well logging data and spectral curve data from different salt rock interference scenarios and formation depths, as the training objects. A masked sequence modeling strategy is adopted: according to a preset ratio, 10%–30% of sequence segments are randomly selected to randomly mask the input data sequence. By obscuring some data points or segments, a training task of predicting incomplete input versus complete output is constructed. In this hybrid deep learning network, the one-dimensional convolutional layer and the encoder of the main reconstruction network serve as the core feature extraction units. During training, they focus on learning the local correlations, global trend features, and potential distribution patterns of formation responses in data sequences. By predicting the true data values ​​of masked regions, they gradually master the general feature representations of intersalt and subsalt formation data, such as the morphological variation patterns of formation curves, the feature distribution corresponding to different lithologies, and the common influence patterns of salt rock interference. The entire pre-training process is conducted in an unsupervised learning manner, without the need for manual labeling. This approach not only fully utilizes massive amounts of unlabeled data to expand the scale of training data but also allows the model to escape the constraints of specific scenarios, learning general feature representations applicable to various salt rock interferences and different formation conditions. This effectively improves the model's generalization ability and provides stable and reliable basic feature support for accurate fitting in the subsequent main training stage and local adaptation in the fine-tuning stage.

[0043] In the main training phase of the hybrid deep learning network, to ensure the stability and reliability of core feature extraction, the encoder parameters of the general stratigraphic representation learned in the pre-training phase are first fixed to avoid interference in subsequent training. The focus then shifts to the refined training of the main reconstruction network, which integrates the attention layer, decoder, and dual discriminator. The training process uses paired samples from the training dataset as the core input. These samples consist of a one-to-one correspondence between the original distortion curve and the corresponding real stratigraphic response curve, providing the model with a clear training guide for input-target output. For the fusion attention layer, the weight allocation strategy for key formation features, such as features strongly correlated with the real response and features related to geological knowledge, is continuously optimized through feature mapping learning of paired samples. This strengthens the contribution of effective information and suppresses interference from redundant features. For the decoder, the inverse reconstruction mapping relationship between abstract features and real curves is continuously learned based on paired samples. Combined with its embedded differentiable simplified logging physical response equation, the reconstructed curve is ensured to conform to the formation physical response law while maintaining numerical fitting accuracy. The dual discriminators form a game training with the hybrid deep learning network. The first discriminator calibrates the discrimination criteria through the real curves in the paired samples and verifies the authenticity of the reconstructed curve output by the decoder. The second discriminator generates a simulated distortion curve from the reconstructed curve and compares it with the original distortion curve in the paired samples. Together, they provide gradient feedback for the model. Meanwhile, the main training phase introduces a weighted hybrid loss function that includes point-to-point mean squared error loss to ensure accurate matching of curve numerical points and morphological loss based on dynamic time warping to ensure the consistency of the overall shape of the curve. Through multi-dimensional loss constraints, the training direction is further standardized, and the fusion of attention layer, decoder and dual discriminator is promoted to quickly adapt to task requirements, so as to achieve steady improvement in model performance.

[0044] In the fine-tuning phase of the hybrid deep learning network, the core focus is on the local geological features and logging data characteristics of the target well section. Targeted adaptive fine-tuning and reconstruction training is conducted to further improve the model's reconstruction accuracy in specific application scenarios. This phase fully utilizes the enhanced training samples containing multi-scale physical perturbations built in the previous stage. These samples are generated through a continuous parameterized space of the logging response distortion field, covering diverse scenarios of real salt rock interference spectra and accurately matching the local geological conditions such as salt rock interference intensity and lithological combinations that may exist in the target well section. During fine-tuning, instead of fixing all model parameters, the focus is on adjusting key parameters of the fusion attention layer and decoder based on the stable model architecture established in the main training phase. Simultaneously, the model's adaptability to specific formation response patterns and salt rock interference modes in the target well section is optimized by considering the local data distribution characteristics of the target well section. During training, the model continuously learns and enhances the local features related to the target well section in the training samples, such as the morphology of the formation curves at specific depths and the interference characteristics of the salt rock-surrounding rock contact zone. Through multiple rounds of iteration to correct parameter deviations, the model can accurately capture the local geological details of the well section, effectively solving the reconstruction deviation problem that may occur in general models in specific well sections. Ultimately, it achieves depth adaptation between the model and the geological conditions of the target well section, ensuring that the output reconstruction results have both global consistency and local accuracy.

[0045] In one possible implementation, step S220 further includes:

[0046] A loss function is introduced during the main training phase. The loss function is a weighted hybrid loss, which includes point-to-point mean square error loss and morphological loss based on dynamic time warping.

[0047] Specifically, in the main training phase of the hybrid deep learning network, to achieve dual constraints on the numerical accuracy and morphological consistency of the reconstructed curve, a weighted hybrid loss function consisting of point-to-point mean square error loss and morphological loss based on dynamic time warping is introduced. By reasonably allocating weight coefficients, the optimization objectives of the two types of losses are integrated, providing accurate and comprehensive gradient feedback for model training. The point-to-point mean square error loss focuses on the numerical deviation between the reconstructed curve and the real formation response curve at each data point. By calculating the mean of the squared differences of the corresponding depth point values, the fitting accuracy of a single data point is quantified, driving the model to optimize parameters to reduce numerical errors and ensure that the reconstructed curve closely matches the real curve in specific numerical values. The morphological loss based on dynamic time warping focuses on the consistency of the overall shape and trend of the curve. Considering the potential local offsets and stretching differences in the intersalt and subsalt formation curves, a dynamic programming algorithm is used to find the optimal matching path between the reconstructed curve and the real curve. The cumulative distance along the path is calculated as the morphological loss value, guiding the model to learn the overall changing trend, peak and valley distribution, and morphological characteristics of the real curve, avoiding the problem of numerical fitting being satisfactory but morphological deviation being severe. In the actual training process, by pre-setting weight coefficients, the two types of losses are weighted and summed according to the characteristics of the training data and the reconstruction requirements to form the final weighted mixed loss value. The model aims to minimize this loss value. The parameters of the fusion attention layer, decoder and dual discriminator are iteratively optimized through the backpropagation algorithm, so that the model can achieve the best in both numerical accuracy and morphological restoration, providing strong loss constraint support for the accurate reconstruction of distortion curves of intersalt and subsalt formations.

[0048] In one possible implementation, step S100 further includes:

[0049] Step S110: Define a salt rock interference parameter group, wherein the salt rock interference parameter group includes at least the geometric morphology parameters of the salt rock body, the electrical parameters of the salt rock, and the contact relationship parameters between the salt rock and the surrounding rock.

[0050] Step S120: Establish a continuous mapping relationship for interference intensity, wherein the continuous mapping relationship for interference intensity maps multidimensional salt rock interference parameters to a one-dimensional interference intensity spectrum, forming a distortion space that can be continuously interpolated.

[0051] Step S130: Based on the continuous mapping relationship between the salt rock interference parameter set and the interference intensity, generate enhanced training samples and add them to the training dataset.

[0052] Specifically, a comprehensive and precise set of salt rock interference parameters is defined to lay the foundation for the subsequent construction of a continuous parameterized space for the logging response distortion field. This parameter set takes the interference mechanism of salt rock on the logging response as its core logic and covers parameters in at least three key dimensions: geometric parameters of the salt rock body, which focus on characterizing the spatial distribution characteristics of the salt rock body, including its thickness, lateral extension range, vertical burial depth, and morphological outline, such as regular layered or irregular blocky shapes, which directly affect the propagation path and coverage of the logging signal; electrical parameters of the salt rock body, which focus on the electrical properties of the salt rock itself, including key indicators such as resistivity, dielectric constant, and conductivity. These parameters determine the intensity of the effect of the salt rock on the logging electric and magnetic fields and are the core physical factors causing logging curve distortion; and contact relationship parameters between the salt rock and the surrounding rock, which accurately describe the interface characteristics between the salt rock and the surrounding rock, covering the smoothness of the contact interface, the dip angle, the thickness of the transition zone, and the degree of lithological abrupt change. Differences in physical properties at the interface will cause reflection and refraction of the logging signal, thus causing local curve distortion. By integrating the above three types of parameters, the salt rock interference parameter set can completely and multidimensionally characterize the interference scenarios of salt rock on well logging response, providing comprehensive parameter support for the subsequent generation of enhanced training samples that fit actual geological conditions.

[0053] By establishing a continuous mapping relationship for interference intensity, the normalization and continuous characterization of multidimensional salt rock interference parameters are achieved, thereby constructing a continuous parameterized space for the logging response distortion field that can accurately cover complex interference scenarios. Since the salt rock interference parameter set includes parameters from multiple dimensions such as salt rock geometry, salt rock electrical properties, and salt rock-surrounding rock contact relationships, and these parameters have complex coupling relationships, directly using them for training sample generation can easily lead to fragmented representations of interference scenarios. To address this, a continuous mapping relationship for interference intensity is constructed by introducing a multi-dimensional parameter fusion and normalization algorithm. First, the interference parameters of each dimension of salt rock are standardized to eliminate the influence of dimensional differences and numerical range spans among different parameters. Then, based on the interference mechanism of salt rock on well logging response, the contribution of each parameter to the distortion of well logging curves is quantified through a weight allocation model. For example, the influence weight of salt rock electrical parameters on distortion is higher than that of some geometric parameters. Finally, the processed multi-dimensional salt rock interference parameters are integrated and mapped to a one-dimensional interference intensity spectrum through a nonlinear mapping function, so that salt rock interference of different combinations and intensities can correspond to unique continuous values ​​on the spectrum. This mapping relationship not only realizes the dimensionality reduction and unified representation of multi-dimensional interference parameters, but also forms a distortion space that can be continuously interpolated. By performing gradient interpolation on the one-dimensional interference intensity spectrum, a transitional interference state between known interference scenarios can be generated, filling the scenario gaps under discrete parameter combinations and ensuring that the subsequently generated enhanced training samples can fully cover the continuous distribution range of the real salt rock interference spectrum.

[0054] First, through multi-dimensional parameter combination iterations, parameters such as the geometry of the salt rock mass, its electrical properties, and the contact relationship between the salt rock and the surrounding rock are diversified to form basic parameter combinations covering different types, intensities, and coupling modes of interference. Then, each basic parameter combination is input into a continuous mapping relationship of interference intensity, transforming it into a corresponding one-dimensional interference intensity value. This ensures that each sample accurately corresponds to a specific interference state in the distortion space. Simultaneously, utilizing the continuous interpolation characteristics of the distortion space, a large number of transitional interference scene samples are generated between the basic parameter combinations, filling the scene gaps under discrete parameter combinations and achieving complete coverage of the real salt rock interference spectrum. Furthermore, during the subsequent training of the hybrid deep learning network, this step also supports the implementation of an adversarial sample generation strategy: by monitoring the model training status in real time, the interference parameter regions with large reconstruction errors are accurately located—that is, weak interference scenes that the model has not yet fully learned. More targeted enhanced training samples are proactively generated in these regions to precisely strengthen the model's weak points. All generated enhanced training samples contain distortion curves simulating salt rock interference and corresponding reference information, covering both common interference scenarios and special scenarios such as extreme interference and complex coupled interference. Adding them all to the training dataset can significantly improve the diversity, representativeness, and relevance of the dataset, providing sufficient and realistic training data support for the hybrid reconstruction model. This helps the model to fully learn the curve reconstruction rules under different salt rock interference scenarios, greatly improving the model's generalization ability and reconstruction accuracy.

[0055] In one possible implementation, step S300 further includes:

[0056] Step S310: Probe the target well section and determine the echo detection signal.

[0057] Step S320: Determine the spectral curve based on the echo detection signal, transmit it to the hybrid reconstruction model, perform feature attention fusion and sequence reconstruction under formation multimodal element encoding, and determine the reconstruction result. The reconstruction result includes the formation true response curve of the target well section, the salt rock interference component curve, and the reconstruction confidence matrix. The confidence matrix is ​​generated in real time through an uncertainty propagation network.

[0058] Specifically, specialized logging equipment, such as acoustic logging tools and electromagnetic logging equipment, is deployed to conduct comprehensive and high-precision exploration of the target well section. During the exploration process, the equipment emits a specific frequency detection signal into the formation. When the signal penetrates the salt rock layer, surrounding rock, and various formation interfaces, it is reflected, refracted, and transmitted due to differences in the physical properties of different formation media, such as density, resistivity, and dielectric constant, forming echo signals carrying information about the formation structure, lithological characteristics, and salt rock interference. These echo signals are captured in real time by the equipment's signal receiving unit. Combined with synchronous logging depth recording technology, the depth position of each echo signal segment is accurately associated with the well section depth. Simultaneously, after signal preprocessing, such as noise reduction, filtering, and amplification, invalid information such as environmental interference and equipment noise is eliminated, ultimately determining a high-quality echo detection signal containing the formation response characteristics of the target well section and salt rock interference signals.

[0059] First, the acquired echo detection signals undergo specialized signal processing. Frequency features and amplitude distribution patterns are extracted using techniques such as filtering and denoising, and Fourier transform, generating a spectral curve that accurately reflects the frequency differences between the formation response and salt rock interference, thus transforming the original signal into characteristic spectral data. This spectral curve is then used as the core input to a pre-trained hybrid reconstruction model. After the model starts, a one-dimensional convolutional layer and a three-layer fully connected encoder perform layer-by-layer feature extraction and dimensionality compression on the spectral curve, mining multi-modal elements of the formation, including deep abstract expressions of frequency features, amplitude variations, and depth correlation features. In the feature fusion stage, the fusion attention layer receives the abstract features output by the encoder and, through interaction with the geological knowledge graph, acquires structured knowledge such as lithological prior probabilities and sedimentary facies labels at each reconstruction depth point in the target well section. As an enhancement feature, it strengthens key effective features and suppresses redundant interference features through an attention weight allocation mechanism, thereby improving the relevance and effectiveness of feature expression. Then, under the soft constraint of the differentiable simplified logging physical response equation, the decoder performs inverse restoration and sequence reconstruction of the enhanced multimodal features, and simultaneously drives the uncertainty propagation network to perform real-time computation. By quantifying the parameter uncertainty and feature reliability of the model in the reconstruction process at each depth point, a reconstruction credibility matrix that can intuitively reflect the reliability of the reconstruction results at different depth points is generated. Finally, after a complete model computation process, the reconstruction result containing three types of core information is determined: the true response curve of the target well section that truly reflects the geological characteristics of the formation after removing salt rock interference, the salt rock interference component curve that is separately stripped and clearly presents the salt rock interference characteristics, and the reconstruction credibility matrix that quantifies the reconstruction reliability of each depth point.

[0060] In one possible implementation, step S320 further includes:

[0061] Step S321: Establish the connection between the geological knowledge graph and the fusion attention layer.

[0062] Step S322: By interacting with the geological knowledge map, determine the structured knowledge of the reconstructed depth points in the target well section, wherein the structured knowledge includes at least lithological prior probabilities and sedimentary facies labels.

[0063] Step S323: Use the structured knowledge as an enhanced feature in the feature fusion stage of the attention layer.

[0064] Specifically, by constructing standardized data interaction interfaces and communication protocols, a stable bidirectional connection is established between the geological knowledge graph and the fusion attention layer in the hybrid reconstruction model. This connection, based on data standardization, first transforms the structured knowledge stored in the geological knowledge graph, such as lithology, sedimentary facies, and regional geological patterns, into a feature format recognizable by the model. Simultaneously, a dedicated data receiving and parsing module is configured for the fusion attention layer to ensure that the information output by the knowledge graph accurately matches the model's feature processing logic. On the other hand, the connection supports the model in feeding back basic information about the target well section, such as depth range and geographical location, to the geological knowledge graph, achieving a precise match between model needs and knowledge supply. This connection breaks down the barriers between traditional data-driven models and domain knowledge, enabling professional information from the geological knowledge graph to be injected into the feature fusion process of the fusion attention layer in real time and efficiently.

[0065] Based on the established connection between the geological knowledge graph and the fused attention layer, a two-way data interaction mechanism is used to accurately acquire structured knowledge of the reconstructed depth points of the target well section. The hybrid reconstruction model first inputs the core basic information of the target well section into the geological knowledge graph, including key parameters such as the geographical location, stratigraphic age, depth range, and regional geological background of the well section, providing a precise location basis for knowledge retrieval. Based on this input information, the geological knowledge graph retrieves its own massive amount of stored geological professional data, covering structured knowledge such as regional stratigraphic lithology distribution statistics, sedimentary environment evolution laws, lithology occurrence probability databases at different depth intervals, and sedimentary facies classification standards. Then, for each reconstructed depth point of the target well section, the corresponding core structured knowledge is matched and extracted. This structured knowledge contains at least two key types of information: first, lithological prior probabilities, i.e., the probability distribution data of whether the depth point belongs to different lithological types such as sandstone, mudstone, salt rock, and limestone, derived from the lithological statistical regularities of similar strata in the region and geological exploration experience; second, sedimentary facies labels, which assign clear sedimentary facies classification labels such as fluvial facies, lacustrine facies, deltaic facies, and shallow marine facies based on the sedimentary environment characteristics corresponding to the depth point, intuitively reflecting the paleogeographic environment at the time of strata formation. Through this interactive process, abstract geological knowledge is transformed into quantitative data and classification labels precisely bound to the reconstructed depth point.

[0066] The structured knowledge acquired through interaction with the geological knowledge graph, including at least lithological prior probabilities and sedimentary facies labels, is characterized and transformed into a vector form with the same feature dimensions and format as the encoder output of the hybrid reconstruction model. This vector is then integrated as an enhancement feature into the feature fusion stage of the fusion attention layer. During the fusion process, the fusion attention layer does not simply superimpose this enhancement feature. Instead, based on the logical connections between the geological domain and data features, it deeply fuses the enhancement feature with the multimodal abstract features of the stratigraphy output by the encoder through a dynamic weight allocation mechanism. For depth points with high lithological prior probabilities, the feature weights related to the physical properties of the corresponding stratigraphy are strengthened. Combined with the sedimentary environment patterns reflected by the sedimentary facies labels, the expression of features to specific stratigraphic response patterns is optimized, while suppressing redundant features or noise interference that contradict the structured knowledge. Through this fusion approach, the model's feature expression possesses both data-driven accuracy and incorporates the constraints of prior geological knowledge, effectively improving the feature's ability to represent the true response of the stratigraphy and providing more geologically plausible feature support for subsequent sequence reconstruction by the decoder.

[0067] In one possible implementation, step S300 further includes:

[0068] Step S330: Construct a geological rule base, wherein the geological rule base contains structured knowledge based on stratigraphic compaction trends, lithological-physical property relationships, sedimentary cycle characteristics, and regional geological regularities.

[0069] Step S340: Perform a depth-by-depth point scan on the reconstruction results to detect abnormal sections that violate the geological rule base.

[0070] Step S350: Correct the abnormal section, wherein weighted smoothing and trend fitting using matching geological rules and contextual information are used as correction methods.

[0071] Specifically, focusing on the geological evolution patterns and exploration practices of inter- and sub-salt strata, a highly binding and practical geological rule base is constructed to provide a standardized basis for verifying the geological rationality of the reconstruction results. This rule base integrates four key types of structured knowledge: first, the rule of stratigraphic compaction trends, based on the gradual compaction of particles and the regular changes in porosity and permeability caused by increasing burial depth, which clarifies the reasonable range of variation in the physical properties of strata at different depths; second, the lithology-property correlation rule, covering the inherent correspondence between major lithologies such as sandstone, mudstone, and salt rock and their corresponding porosity, resistivity, and density parameters, defining the normal value boundaries of physical property parameters under specific lithologies; third, the sedimentary cycle rule, reflecting the periodic alternation of grain size and lithology formed during stratigraphic deposition, clarifying the periodic variation patterns and reasonable transition ranges of curve morphology; and fourth, the regional geological regularity rule formed by combining the tectonic background, stratigraphic distribution, and lithological assemblage characteristics of the target well section, ensuring that the reconstruction results are consistent with the regional geological background. By standardizing and formalizing this structured knowledge, the geological rule base forms a comprehensive and quantifiable set of geological judgment standards.

[0072] Using the constructed geological rule base as the core judgment standard, a comprehensive depth-by-depth scan was conducted on the reconstruction results, including the true response curve of the formation and the salt rock interference component curve, output by the hybrid reconstruction model. During the scan, for the reconstruction data at each depth point, the structured rules in the geological rule base, such as formation compaction trend, lithology-physical property correlation, sedimentary cycle characteristics, and regional geological regularities, were matched one by one. The rationality was verified by quantifying the fit between the data and the rules: for example, whether the changes in well logging values ​​with depth conformed to the gradual change law caused by formation compaction, whether the physical property parameters corresponding to a specific lithology were within the reasonable range defined by the rules, and whether the curve shape conformed to the periodic characteristics of the sedimentary cycle. At the same time, the numerical changes of adjacent depth points were monitored in detail to accurately identify abnormal sections that violated geological rules, such as drastic jumps in well logging values ​​that did not conform to the sedimentary regularity, mismatch between lithology and physical property parameters, and data trends that contradicted the regional geological background. This provided a clear target range and anomaly type basis for subsequent targeted correction.

[0073] For anomaly sections detected that violate the geological rule base, precise correction is implemented to ensure the geological rationality and data consistency of the reconstruction results. During the correction process, firstly, based on the anomaly type, such as drastic numerical fluctuations, lithological-physical property mismatches, or trends that violate regional geological patterns, corresponding matching rules are retrieved from the geological rule base. For example, sedimentary cycle rules are matched for drastic fluctuation sections that do not conform to sedimentary cycle characteristics, and lithological-physical property correlation rules are matched for sections with abnormal physical property parameters. Simultaneously, normal reconstruction data from adjacent depth points before and after the anomaly section are fully extracted as contextual information to clarify the natural variation trend and data distribution characteristics of the strata in the area. Based on this, a weighted smoothing algorithm is used to perform gradient correction on anomalous data points. Normal data points near the edge of anomalous sections are given higher weights, while core anomalous data points are gradually adjusted according to geological rules to avoid the data after correction becoming disconnected from the surrounding normal data. At the same time, trend fitting technology is combined with matching geological rules and contextual trends to construct curve models that conform to stratigraphic evolution laws, such as compaction gradients and cyclical patterns. Overall trend optimization is performed on anomalous sections, so that the corrected curves not only eliminate anomalous fluctuations without geological origins, but also strictly conform to geological rules and maintain morphological consistency and logical coherence with the surrounding normal reconstructed data, significantly improving the reliability of the reconstruction results and the geological interpretation value.

[0074] Example 2, based on the same inventive concept as the intelligent algorithm-based method for reconstructing distortion curves of inter- and sub-salt formations in the aforementioned examples, such as... Figure 2 As shown, this application provides a system for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms. The system and method embodiments in this application are based on the same inventive concept. The system includes:

[0075] Training dataset construction module 10 is used to construct a training dataset containing multi-scale physical perturbations, wherein enhanced training samples covering the real salt rock interference spectrum are generated by establishing a continuous parameterized space of the well logging response distortion field.

[0076] The hybrid reconstruction model building module 20 is used to construct a hybrid reconstruction model by configuring a hybrid deep learning network and adopting a staged training method. The hybrid deep learning network includes a one-dimensional convolutional layer and a main reconstruction network.

[0077] The reconstruction result determination module 30 is used to determine the reconstruction result by collecting echo detection signals from the target well section and performing hierarchical cascade analysis based on the hybrid reconstruction model.

[0078] The terminal interface display module 40 is used to display the reconstruction result on the terminal interface.

[0079] Furthermore, the system is also used to implement the following functions:

[0080] The hybrid reconstruction model uses a hybrid deep learning network as its model architecture. The main reconstruction network includes a three-layer fully connected encoder, a fused attention layer, and a decoder. The decoder is embedded with a differentiable simplified well logging physical response equation as a soft constraint. During the training of the hybrid deep learning network, a discriminator is introduced to play a game with the hybrid deep learning network until convergence.

[0081] Furthermore, the system is also used to implement the following functions:

[0082] The discriminator includes a first discriminator and a second discriminator; wherein, the first discriminator is used to judge the distortion of the model reconstruction curve; the second discriminator is used to generate a simulated distortion curve from the model reconstruction curve in reverse, and compare the simulated distortion curve with the distortion curve of the original input.

[0083] Furthermore, the system is also used to implement the following functions:

[0084] The training process includes a pre-training phase, a main training phase, and a fine-tuning phase. In the pre-training phase, a general representation of the formation is learned through unlabeled pre-training under masked sequence modeling. In the main training phase, the encoder parameters are fixed, and paired samples from the training dataset are used to train the fusion attention layer, decoder, and discriminator. In the fine-tuning phase, the generated enhanced training samples are used to perform local adaptive fine-tuning and reconstruction training of well sections.

[0085] Furthermore, the system is also used to implement the following functions:

[0086] A loss function is introduced during the main training phase. The loss function is a weighted hybrid loss, which includes point-to-point mean square error loss and morphological loss based on dynamic time warping.

[0087] Furthermore, the system is also used to implement the following functions:

[0088] Define a set of salt rock interference parameters, which includes at least the geometric morphology parameters of the salt rock mass, the electrical properties of the salt rock, and the contact relationship parameters between the salt rock and the surrounding rock. Establish a continuous mapping relationship for interference intensity, which maps the multidimensional salt rock interference parameters to a one-dimensional interference intensity spectrum, forming a distortion space that can be continuously interpolated. Based on the set of salt rock interference parameters and the continuous mapping relationship for interference intensity, generate enhanced training samples and add them to the training dataset.

[0089] Furthermore, the system is also used to implement the following functions:

[0090] The target well section is probed to determine the echo detection signal; the spectral curve based on the echo detection signal is determined and transmitted to the hybrid reconstruction model, where feature attention fusion and sequence reconstruction under formation multimodal element encoding are performed to determine the reconstruction result. The reconstruction result includes the formation true response curve of the target well section, the salt rock interference component curve, and the reconstruction confidence matrix. The confidence matrix is ​​generated in real time through an uncertainty propagation network.

[0091] Furthermore, the system is also used to implement the following functions:

[0092] Establish a connection between the geological knowledge graph and the fusion attention layer; determine the structured knowledge of the reconstructed depth points in the target well section by interacting with the geological knowledge graph, wherein the structured knowledge includes at least lithological prior probabilities and sedimentary facies labels; use the structured knowledge as an enhancement feature of the feature fusion stage of the fusion attention layer.

[0093] Furthermore, the system is also used to implement the following functions:

[0094] A geological rule base is constructed, which contains structured knowledge based on stratigraphic compaction trends, lithological-physical property correlations, sedimentary cycle characteristics, and regional geological regularities. The reconstruction results are scanned at depth points to detect abnormal segments that violate the geological rule base. The abnormal segments are corrected by weighted smoothing and trend fitting using matched geological rules and contextual information.

[0095] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0096] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0097] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms, characterized in that, The method includes: A training dataset containing multi-scale physical perturbations is constructed, wherein enhanced training samples covering the real salt rock disturbance spectrum are generated by establishing a continuous parameterized space of the well logging response distortion field. By configuring a hybrid deep learning network and adopting a staged training method, a hybrid reconstruction model is constructed, wherein the hybrid deep learning network includes a one-dimensional convolutional layer and a main reconstruction network; By acquiring echo detection signals from the target well section, a hierarchical cascade analysis is performed based on the hybrid reconstruction model to determine the reconstruction results; The reconstruction results are displayed on the terminal interface; The continuous parameterization space of the logging response distortion field includes: Define a salt rock interference parameter set, wherein the salt rock interference parameter set includes at least the geometric morphology parameters of the salt rock body, the electrical parameters of the salt rock, and the contact relationship parameters between the salt rock and the surrounding rock; A continuous mapping relationship for interference intensity is established, wherein the continuous mapping relationship for interference intensity maps multidimensional salt rock interference parameters to a one-dimensional interference intensity spectrum, forming a distortion space that can be continuously interpolated; Based on the continuous mapping relationship between the salt rock interference parameter set and the interference intensity, enhanced training samples are generated and added to the training dataset; Based on the aforementioned hybrid reconstruction model, a hierarchical cascade analysis is performed to determine the reconstruction results, including: The target well section is probed to determine the echo detection signal; The spectral curve based on the echo detection signal is determined and transmitted to the hybrid reconstruction model. Feature attention fusion and sequence reconstruction under formation multimodal element encoding are performed to determine the reconstruction result. The reconstruction result includes the formation true response curve of the target well section, the salt rock interference component curve, and the reconstruction confidence matrix. The confidence matrix is ​​generated in real time through an uncertainty propagation network.

2. The method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms as described in claim 1, characterized in that, The hybrid reconstruction model uses a hybrid deep learning network as its model architecture. The main reconstruction network includes a three-layer fully connected encoder, a fused attention layer, and a decoder. The decoder is embedded with a differentiable simplified logging physical response equation as a soft constraint. During the training of the hybrid deep learning network, a discriminator is introduced to engage in game-theoretic training with the hybrid deep learning network until convergence.

3. The method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms as described in claim 2, characterized in that, The discriminator includes a first discriminator and a second discriminator; The first discriminator is used to determine the distortion of the model reconstruction curve; the second discriminator is used to generate a simulated distortion curve from the model reconstruction curve and compare the simulated distortion curve with the original input distortion curve.

4. The method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms as described in claim 3, characterized in that, The training process includes a pre-training phase, a main training phase, and a fine-tuning phase. In the pre-training phase, a general representation of the strata is learned through unlabeled pre-training under masked sequence modeling. During the main training phase, the encoder parameters are fixed, and paired samples from the training dataset are used to train the fusion attention layer, decoder, and discriminator. During the fine-tuning phase, the generated enhanced training samples are used to perform local adaptive fine-tuning and reconstruction training of the well section.

5. The method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms as described in claim 4, characterized in that, A loss function is introduced during the main training phase. The loss function is a weighted hybrid loss, which includes point-to-point mean square error loss and morphological loss based on dynamic time warping.

6. The method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms as described in claim 1, characterized in that, Establish a connection between the geological knowledge graph and the fusion attention layer; By interacting with the geological knowledge graph, the structured knowledge of the reconstructed depth points in the target well section is determined, wherein the structured knowledge includes at least lithological prior probabilities and sedimentary facies labels; The structured knowledge is used as an enhanced feature in the feature fusion stage of the attention layer.

7. The method for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms as described in claim 1, characterized in that, After determining the reconstruction result, the following is included: Construct a geological rule base, which contains structured knowledge based on stratigraphic compaction trends, lithological-physical property relationships, sedimentary cycle characteristics, and regional geological regularities; The reconstruction results are scanned point by point at depth to detect abnormal sections that violate the geological rule base. The abnormal sections are corrected by using weighted smoothing and trend fitting based on matching geological rules and contextual information as correction methods.

8. A system for reconstructing distortion curves of inter- and sub-salt formations based on intelligent algorithms, characterized in that, The system is used to implement the intelligent algorithm-based method for reconstructing distortion curves of inter- and sub-salt formations as described in any one of claims 1-7. The system comprises: The training dataset construction module is used to construct a training dataset containing multi-scale physical perturbations. In this module, enhanced training samples covering the real salt rock interference spectrum are generated by establishing a continuous parameterized space of the well logging response distortion field. The hybrid reconstruction model building module is used to construct a hybrid reconstruction model by configuring a hybrid deep learning network and adopting a staged training method. The hybrid deep learning network includes a one-dimensional convolutional layer and a main reconstruction network. The reconstruction result determination module is used to determine the reconstruction result by collecting echo detection signals from the target well section and performing hierarchical cascade analysis based on the hybrid reconstruction model. The terminal interface display module is used to display the reconstruction results on the terminal interface.

Images