Well logging curve high-resolution reconstruction method and system for tight sandstone reservoir

Abstract

The invention discloses a well logging curve high-resolution reconstruction method and system for a tight sandstone reservoir, relates to the field of geological exploration and development, and aims to solve the problems that an existing resolution enhancement technology lacks a dynamic feedback mechanism and is difficult to effectively enhance diagenetic phase recognition work. According to the technical key points, the method comprises the following steps: collecting data of a plurality of logging curves, carrying out wavelet transform on the data, endowing a high-frequency coefficient with a weight, and reconstructing the logging curves; an LSTM model is constructed, a model training loss function is used as a fitness function, a particle swarm algorithm is adopted to optimize the weight endowed by the reconstructed logging curve, and the optimal weight is obtained; in the iteration process of the particle swarm algorithm, according to the training loss response of the model to a reconstructed logging curve under different wavelet decomposition scales, adaptively adjusting the number of wavelet decomposition layers participating in weight optimization or the participation range of a corresponding high-frequency coefficient; and reconstructing the logging curve based on the optimal weight to identify the lithogenous phase.

Description

Technical Field

[0001] This invention relates to the field of geological exploration and development technology, and more specifically, to a high-resolution reconstruction method, system, and storage medium for logging curves in tight sandstone reservoirs. Background Technology

[0002] Diagenesis, through mechanisms such as dissolution, cementation, and compaction, reshapes reservoir structure and properties, and is a key factor in determining reservoir performance. In recent years, with advancements in oil and gas exploration and development technologies, tight sandstone reservoirs have become important development targets for unconventional oil and gas resources. However, these reservoirs generally exhibit complex mineral compositions, intense diagenetic alteration, and multi-scale pore structure development, leading to challenges such as insufficient resolution and weakened response characteristics in conventional logging curves for diagenetic facies analysis, making it difficult to effectively extract sensitive diagenetic facies features. Currently, logging identification methods for tight sandstone reservoir diagenetic facies mainly rely on the correlation analysis between logging curves and core data, and can be divided into two categories: empirical discrimination and statistical analysis. Empirical discrimination methods classify diagenetic facies by constructing the response relationship between logging parameters and diagenetic minerals, but it is highly dependent on core calibration data from cored wells, limiting its application in non-cored well scenarios. Statistical analysis methods utilize machine learning to extract logging curve features, thereby automatically identifying diagenetic facies. Well logging curves exhibit a longitudinally continuous temporal response characteristic as they change with depth. Long short-term memory (LSTM) neural networks have a strong ability to capture long-range dependency features of time-series data, which has led to their widespread application in diagenetic facies identification. However, in low-resolution data scenarios, LSTM has limited ability to extract subtle features of diagenesis and struggles to accurately capture its multi-scale features, resulting in decreased identification accuracy.

[0003] To address these issues, researchers have attempted to improve the quality of basic data through resolution enhancement techniques. Wavelet packet decomposition, by compensating for the high-frequency energy of well logging curves, can effectively enhance the response signals of subtle geological bodies such as thin interbedded layers. However, current research only performs independent high-resolution processing on single sensitive well logging curves, without coordinating the optimization of multiple curves. Furthermore, existing methods often process feature enhancement and classification independently, lacking a dynamic reconstruction mechanism based on classifier feedback, making it difficult to further improve the accuracy of diagenetic facies identification through reconstructed curves. Summary of the Invention

[0004] The technical problem to be solved by this invention is:

[0005] Existing logging curves for tight sandstone reservoirs are easily affected by surrounding rock interference, leading to reduced resolution and severely impacting the accuracy of diagenetic facies identification. Furthermore, existing resolution enhancement technologies lack dynamic feedback mechanisms, making it difficult to effectively enhance diagenetic facies identification.

[0006] To this end, this invention, based on the improvement of logging curve resolution by wavelet transform, integrates LSTM-driven particle swarm optimization (PSO) algorithm to propose a ParticleSwarm Optimization driven wavelet reconstruction logging curve method based on long short-term memory neural network feedback (LSTM-PSO-WR). This method enhances the features of logging curves, and the reconstructed features are input into a classification model to verify the effect of curve reconstruction on improving the accuracy of diagenetic facies identification.

[0007] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0008] This invention provides a high-resolution reconstruction method for logging curves in tight sandstone reservoirs, comprising the following steps:

[0009] S1. Collect multiple logging curve data, preprocess the data, and label the preprocessed logging curve data;

[0010] S2. Perform wavelet transform on the preprocessed logging curve data, assign weights to the high-frequency coefficients obtained after wavelet transform, and reconstruct the logging curve.

[0011] S3. Construct a training dataset based on the reconstructed logging curve data; construct an LSTM model, train the LSTM model based on the training dataset, use the LSTM model training loss function as the fitness function, and use the particle swarm optimization algorithm to optimize the weights assigned to the reconstructed logging curves to obtain the optimal weights; during the iteration of the particle swarm optimization algorithm, according to the model's training loss response to the reconstructed logging curves at different wavelet decomposition scales, adaptively adjust the number of wavelet decomposition layers or the participation range of the corresponding high-frequency coefficients participating in the weight optimization, so as to perform multi-scale adaptive reconstruction for different reservoir characteristics;

[0012] S4. Identify diagenetic facies by reconstructing the logging curves based on the optimal weights.

[0013] Furthermore, the data preprocessing described in S1 includes: noise suppression and outlier correction.

[0014] based on The criterion identifies jump point noise and calculates the mean using a sliding window. with standard deviation ;like Then replace the outlier with the window's midpoint:

[0015]

[0016] Unified depth sequence:

[0017] For low-resolution depth sequences, non-equidistant sampling data is resampled to ensure the depth sequences of each well are properly configured. With corresponding depth value Precise alignment:

[0018]

[0019] in and Points to be interpolated Adjacent reference depth nodes;

[0020] Align the depth sequences of each well with their corresponding depths.

[0021] Furthermore, S2 includes the following process:

[0022] First, the preprocessed well logging curve data is subjected to wavelet transform using db4 wavelet. Layer high frequency coefficient Generated through high-pass filtering and downsampling, satisfying:

[0023]

[0024] in These are the db4 wavelet high-pass filter coefficients; Indicates the first High-frequency detail coefficients after layer wavelet decomposition; It is the first The input sequence of layer decomposition coefficients after downsampling and shifting; For discrete-time indexing; This is the tap index of the filter;

[0025] Then, weights are assigned to the high-frequency coefficients of the decomposition. ,Right now:

[0026]

[0027] Finally, the weighted high-frequency coefficients are upsampled, and after convolution and superposition using a db4 wavelet filter, the curve features are recovered layer by layer, and the reconstructed logging curve is obtained iteratively.

[0028] Furthermore, the wavelet transform has 3 to 5 decomposition layers, particularly 4 layers.

[0029] Furthermore, S3 includes the following steps:

[0030] (1) Initialize the particle swarm algorithm with the weights assigned to the reconstructed logging curves as the optimization objective;

[0031] (2) The reconstructed well logging curves are stitched together to obtain the stitched dataset. , Dimensions ,in Let F be the number of depth point samples and F be the number of features. The stitched data is normalized, and T consecutive depth points are extracted along the depth direction to construct a time series, thus constructing a three-dimensional tensor. The training dataset;

[0032] (3) Construct an LSTM model, train the LSTM model based on the training dataset, use the LSTM model training loss function as the fitness function, and calculate the fitness of each particle;

[0033] (4) Based on the population division results and particle fitness values, determine the individual historical best position of each particle and the global best position of the particle swarm, and update the velocity and position of ordinary particles and locally optimal particles.

[0034] (5) Repeat the above process for iterative calculation, and based on the training loss response of the model to the reconstructed logging curves at different wavelet decomposition scales, adaptively adjust the number of wavelet decomposition layers or the participation range of the corresponding high-frequency coefficients in the weight optimization according to the preset loss change criteria, so as to perform multi-scale adaptive reconstruction for different reservoir characteristics; until the iteration value converges or the maximum number of iterations is reached, the optimal parameter value is obtained.

[0035] Furthermore, the LSTM model described in step (3) includes two-layer LSTM units, with 32 hidden neurons in each layer, and layers are arranged with... The layer, through the Softmax function, outputs the probability distribution of lithogenic facies, and its state update satisfies:

[0036]

[0037] in , , These are the activation values ​​for the forget gate, input gate, and output gate, respectively. , These are the unit state and the hidden state. For the first Layer input; for Activation function.

[0038] Furthermore, the training loss function of the LSTM model in step (3) is:

[0039] .

[0040] Furthermore, the particle velocity in step (4) With position The iterative update process is as follows:

[0041]

[0042]

[0043] in, For the first Individual particles The velocity vector at any given moment; Inertial weight; For the first Individual particles The velocity vector at any given moment; , For learning factors; , A random number within the interval [0,1]. For the first The best position in the history of each particle; For the first Individual particles Position vector at any given time; This is the globally optimal position for the entire particle swarm; For the first Individual particles The position vector at any given time.

[0044] This invention provides a high-resolution reconstruction system for logging curves in tight sandstone reservoirs. The system has program modules corresponding to the steps of any of the above-described technical solutions, and executes the steps in the above-described high-resolution reconstruction method for logging curves in tight sandstone reservoirs during operation.

[0045] The present invention provides a computer-readable storage medium storing a computer program configured to, when invoked by a processor, implement the steps in the high-resolution reconstruction method for logging curves of tight sandstone reservoirs as described in any of the above technical solutions.

[0046] Compared with the prior art, the beneficial effects of the present invention are:

[0047] This invention proposes a long short-term memory feedback particle swarm optimization wavelet method for reconstructing well logging curves. Based on wavelet transform and weighted reconstruction, it optimizes the wavelet reconstruction weights of multiple key well logging curves using a particle swarm optimization algorithm, and dynamically optimizes the fitness function using the training loss of a long short-term memory neural network model. A reconstruction mechanism of "dataset construction - wavelet decomposition - weight optimization - curve reconstruction - iterative evaluation" is established to specifically improve the resolution of well logging curves and the accuracy of diagenetic facies identification. Validation based on actual well logging data shows that the method can effectively enhance high-frequency details of well logging curves, significantly improve the accuracy of diagenetic facies identification, and exhibit good adaptability and generalization ability in different data volumes and cross-regional scenarios, providing a new technical approach and practical reference for the field of well logging curve resolution enhancement. Attached Figure Description

[0048] Figure 1 This is a flowchart of the high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to an embodiment of the present invention.

[0049] Figure 2 This is a flowchart illustrating the dataset construction process in an embodiment of the present invention.

[0050] Figure 3 This is a schematic diagram of the wavelet reconstruction process in an embodiment of the present invention;

[0051] Figure 4 This is a structural diagram of the LSTM unit in an embodiment of the present invention;

[0052] Figure 5 This is a flowchart illustrating the fitness optimization process in an embodiment of the present invention.

[0053] Figure 6 This is a comparison diagram of the reconstructed curve and the original curve in an embodiment of the present invention;

[0054] Figure 7 This is a comparison diagram of the energy of the original curve and the reconstructed curve in an embodiment of the present invention;

[0055] Figure 8 This is a model identification diagenetic facies confusion matrix diagram in an embodiment of the present invention;

[0056] Figure 9 This is a comparison chart of the model recognition performance in the embodiments of the present invention;

[0057] Figure 10 This is a diagram showing the ablation experiment results in an embodiment of the present invention. Detailed Implementation

[0058] To enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described embodiments or examples are merely some, not all, of the embodiments or examples of the present invention. All other embodiments or examples obtained by those skilled in the art based on the embodiments or examples of the present invention without inventive effort should fall within the scope of protection of the present invention.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0060] Combination Figures 1 to 4 As shown, this invention provides a high-resolution reconstruction method for logging curves in tight sandstone reservoirs, comprising the following steps:

[0061] S1. Collect multiple logging curve data, preprocess the data, and label the preprocessed logging curve data;

[0062] S2. Perform wavelet transform on the preprocessed logging curve data, assign weights to the high-frequency coefficients obtained after wavelet transform, and reconstruct the logging curve.

[0063] S3. Construct a training dataset based on the reconstructed logging curve data; construct an LSTM model, train the LSTM model based on the training dataset, use the LSTM model training loss function as the fitness function, and use the particle swarm optimization algorithm to optimize the weights assigned to the reconstructed logging curves to obtain the optimal weights; during the iteration of the particle swarm optimization algorithm, according to the model's training loss response to the reconstructed logging curves at different wavelet decomposition scales, adaptively adjust the number of wavelet decomposition layers or the participation range of the corresponding high-frequency coefficients participating in the weight optimization, so as to perform multi-scale adaptive reconstruction for different reservoir characteristics;

[0064] S4. Identify diagenetic facies by reconstructing the logging curves based on the optimal weights.

[0065] Using the Fuyu oil layer in the Sanzhao Depression of the Songliao Basin as the target area, six key logging curves from 10 wells were selected, specifically including natural gamma ray logging (GR), sonic transit time logging (AC), spontaneous potential logging (SP), caliper logging (CAL), deep lateral logging (LLD), and shallow lateral logging (LLS). To address issues such as missing initial data, anomalies, and inconsistent sampling intervals, the data underwent preprocessing. Taking the depth curve from 1787.3 to 1817.2 m in well x1 (x1 represents hidden information) as an example, the processing procedure is as follows: As shown.

[0066] (1) Noise suppression and outlier correction: based on The criteria identify jump point noise and calculate the mean using a 0.5m sliding window. with standard deviation .like Then replace the outlier with the window's midpoint:

[0067]

[0068] (2) Unified depth sequence: Using 0.05m as the baseline interval, the low-resolution depth sequence is resampled by non-equidistant sampling data to ensure that the depth sequence of each well is consistent. With corresponding depth value Precise alignment:

[0069]

[0070] in and Points to be interpolated The adjacent reference depth nodes.

[0071] After preprocessing, the well logging curve dataset is obtained. The response characteristics of different well logging curves were analyzed, and labels were added to the dataset. The thin section samples of mudstone facies, compacted dense siltstone, and dissolved siltstone in the table are from the literature. .

[0072] Specific implementation plan three: S2 includes the following process:

[0073] Based on the preprocessed dataset, the db4 wavelet is used to perform a four-level decomposition of the logging curves. By assigning weights to high-frequency coefficients and reconstructing the data, the high-frequency details of the curves are enhanced. Taking the SP curve as an example, the process is as follows: As shown.

[0074] First, the curve is decomposed using db4 wavelet decomposition. Layer high frequency coefficient High-pass filtering and downsampling are used to generate curve features that reflect local abrupt changes and high-frequency details. After four-level decomposition, high-frequency trend coefficients are obtained. to ,satisfy:

[0075]

[0076] in, Indicates the first High-frequency detail coefficients after layer wavelet decomposition; It is the first The input sequence of layer decomposition coefficients after downsampling and shifting; For discrete-time indexing; This is the tap index of the filter. These are the db4 wavelet high-pass filter coefficients, with values ​​as follows:

[0077] , , , .

[0078] Then, weights are assigned to the high-frequency coefficients of the decomposition. ,Right now:

[0079]

[0080] The weight values ​​corresponding to the high-frequency coefficients range from 0.5 to 2.0.

[0081] Finally, the weighted high-frequency coefficients are upsampled, and after convolution and superposition using a db4 wavelet filter, the curve features are recovered layer by layer, and the reconstructed logging curve is obtained through iterative reconstruction.

[0082] Specific implementation scheme four: The wavelet transform decomposition layer is 3 to 5 layers, especially 4 layers.

[0083] Specific implementation plan five: S3 includes the following steps:

[0084] A particle swarm optimization algorithm is introduced into the wavelet weighting part to optimize the decomposition of six well logging curves. ~ Corresponding weights The particle swarm optimization algorithm uses the training loss of the LSTM model as the fitness function to iteratively update particle positions and find the optimal weights to improve the accuracy of diagenetic facies identification.

[0085] (1) Initialize the particle swarm optimization algorithm with the weights as the optimization objective: set the particle size. The particle positions follow a uniform distribution. Limiting the search space Initial velocity of the particles In the interval Random initialization between; global optimal position Initialize to empty, optimal fitness Positive infinity ensures the optimal solution is updated after the initial evaluation; momentum coefficient Learning factor .

[0086] (2) The reconstructed well logging curves are stitched together to obtain the stitched dataset. , Dimensions ,in The number of depth point samples. For the number of features, After Max-Min normalization, continuous segments are truncated along the depth direction. A time series is constructed from each depth point, and the structure is organized into a three-dimensional tensor. The training dataset;

[0087] (3) Construct an LSTM model with two layers of LSTM units, each layer having 32 hidden neurons, and between layers To prevent overfitting, input a three-dimensional tensor. The probability distribution of various diagenetic facies is output through the Softmax function. The LSTM element structure is as follows: As shown.

[0088] Its state update satisfies:

[0089]

[0090] in , , These are the activation values ​​for the forget gate, input gate, and output gate, respectively. , These are the unit state and the hidden state. For the first Layer input; for Activation function.

[0091] The training loss function for the LSTM model is:

[0092]

[0093] The loss function value from the 20th training round is used as the fitness to guide particle movement.

[0094] (4) Based on the population partitioning results and particle fitness values, determine the individual historical optimal position of each particle and the global optimal position of the particle swarm, and update the velocity and position of ordinary particles and locally optimal particles; particle velocity With position The iterative update process is as follows:

[0095]

[0096]

[0097] in, These are random numbers, generated during each iteration. For the first Individual particles The velocity vector at any given moment; Inertial weight (used to control the degree to which particles inherit historical velocities); For the first Individual particles The velocity vector at any given moment; , The learning factor (controls the degree to which a particle learns from its own historical best and global best, respectively). For the first The best position in the history of each particle; For the first Individual particles Position vector at any given time; This is the globally optimal position for the entire particle swarm; For the first Individual particles The position vector at any given time.

[0098] And for each dimension Perform cutting to ensure .

[0099] (5) Repeat the above process for iterative calculation, and based on the training loss response of the model to the reconstructed logging curves at different wavelet decomposition scales, adaptively adjust the number of wavelet decomposition layers or the participation range of the corresponding high-frequency coefficients in the weight optimization according to the preset loss change criteria, so as to perform multi-scale adaptive reconstruction for different reservoir characteristics; until the iteration value converges or the maximum number of iterations is reached, the optimal parameter value is obtained.

[0100] During iteration, if the current particle is better, the local / global optimum is updated. When the iteration value converges or the maximum number of iterations is reached, the logging curve is reconstructed with the global optimum weights, and high-resolution features are output.

[0101] The graph shows the trend of particle swarm optimization (PSO) fitness as a function of iterations. The horizontal axis represents the number of iterations, and the vertical axis represents the PSO fitness. The loss value before reconstruction is also included as a benchmark. It can be seen that the fitness decreases rapidly in the early stages of iteration (around 0-20 iterations), indicating that the algorithm quickly searches for a better solution space. After 20 iterations, the fitness curve flattens out, and the algorithm gradually converges and stabilizes at a lower value. The final fitness is lower than the loss value before reconstruction, indicating that the logging curve performance after PSO optimization is significantly better than the initial unoptimized state.

[0102] The high-resolution reconstruction method (algorithm) of logging curves for tight sandstone reservoirs proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0103] Based on the method proposed in this invention, a high-resolution reconstruction system for logging curves in tight sandstone reservoirs is developed using a programming language. This system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned high-resolution reconstruction method for logging curves in tight sandstone reservoirs when running.

[0104] The developed system (software) computer program is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-described high-resolution reconstruction method for logging curves in tight sandstone reservoirs when called by a processor. In other words, the invention is materialized on a carrier, becoming a computer program product.

[0105] Various implementations of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, application-specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include: implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0106] The computational programs (also referred to as programs, software, software applications, or code) of this invention include machine instructions of a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including machine-readable media that receive machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0107] The beneficial effects of the present invention will be described below with reference to specific embodiments.

[0108] Example 1

[0109] To verify the effectiveness of the LSTM-PSO-WR logging curve reconstruction method of this invention in enhancing resolution and improving the accuracy of diagenetic facies identification, this embodiment designed a spectrum analysis experiment, a diagenetic facies identification accuracy comparison experiment, an ablation verification experiment, and a generalization ability evaluation experiment. The diagenetic facies identification accuracy comparison experiment and the ablation verification experiment used the aforementioned LSTM model structure and the aforementioned data to construct a training set to train the model, predicting the diagenetic facies in well x2 (x2 being hidden information), and quantifying the logging curve reconstruction quality by assessing the diagenetic facies identification effect.

[0110] 1. Spectrum Analysis Experiment

[0111] To verify the resolution improvement effect of well logging curve reconstruction, the experiment was designed from two dimensions: visual comparison and quantitative analysis. First, CIFLog software was used to plot a comparison chart of the original and reconstructed curves, visually presenting the detailed differences between the curves before and after reconstruction and qualitatively judging the degree of resolution enhancement. Then, spectral energy analysis was performed on the original and reconstructed signals of the six curves to quantitatively demonstrate the resolution improvement. Spectral energy analysis, based on the Fast Fourier Transform (FFT) theory, converts time-domain well logging data into a frequency-domain representation. By calculating the energy distribution of each frequency component, the frequency characteristics of the well logging curves are revealed. For discrete well logging data… Its spectral energy reflects the contribution of different frequency components to the total energy of the curve, and the calculation method is as follows:

[0112] First, the discrete signal Perform FFT transformation:

[0113]

[0114] in The number of signal sampling points. It is the imaginary unit.

[0115] Considering spectral symmetry, only the first few spectral values ​​are taken. Energy calculation at each point:

[0116]

[0117] After performing the FFT transformation, the correspondence between the frequency axis and the sampling parameters is as follows:

[0118]

[0119] in The sampling frequency of the original signal. Corresponding to the The actual frequency value of each frequency point can be used to index the spectrum using this formula. Mapping to the physical frequency domain for analysis.

[0120] After obtaining the optimal reconstruction weights from the original data using the particle swarm optimization algorithm, a reconstruction curve is generated through wavelet decomposition and weighted reconstruction. Taking the training well x3 data as an example, a comparison of the original and reconstructed curves is shown below. The figure shows that the blue curve represents the original curve and the red curve represents the reconstructed curve. As can be seen from the figure, the reconstructed curve effectively enhances the weaker features and improves the resolution of the thin layer.

[0121] Fourier spectrum analysis results are from The results show that the original and reconstructed logging curves of the six sets of well logging curves exhibit a regular difference in frequency domain characteristics: all curves maintain a consistent basic trend in the low-frequency range of 0-0.4Hz, while in the key frequency range of 0.4-5Hz, the reconstructed curves achieve targeted energy adjustment through wavelet weight optimization, supplementing high-frequency details, mitigating energy decay, and enhancing characteristic responses. Furthermore, since the reconstructed curves only adjust details without altering the overall trend, this adjustment effectively corrects the feature weakening caused by reservoir heterogeneity, instrument bias, and mineral superposition effects in the original curves, ensuring no deviation from the original geological characteristics. It precisely enhances stratigraphic information related to diagenetic facies identification, providing high-resolution data support for subsequent diagenetic facies identification.

[0122] 2. Comparison Experiment on the Accuracy of Diagenetic Facies Identification

[0123] To verify the effect of LSTM-PSO-WR reconstruction curve on improving the accuracy of diagenetic facies identification, an experiment was conducted to compare the specific performance of the LSTM model trained with the original curve and the reconstructed curve in identifying diagenetic facies. The confusion matrix was plotted, and the Matthews correlation coefficient (MCC) and classification accuracy were calculated for quantitative verification. The Matthews correlation coefficient is a commonly used indicator for evaluating the performance of a classification model, while the accuracy represents the proportion of correctly identified numbers in the total number of numbers. The calculation methods for both are shown in equations (13) and (14).

[0124]

[0125] in, This represents the number of samples correctly identified by the model as the target diagenetic facies. This represents the number of samples correctly identified by the model as non-target diagenetic facies. This represents the number of samples that were incorrectly identified as the target diagenetic facies by the model. This indicates the number of samples that were incorrectly identified by the model as non-target diagenetic facies. This represents the total number of samples.

[0126] The confusion matrix calculation results of the original curve training model (Model 1) and the reconstructed curve training model (Model 2) are presented. Figure (a) shows the confusion matrix of diagenetic facies identified by Model 1, and Figure (b) shows the confusion matrix of diagenetic facies identified by Model 2. In the figures, the X-axis represents the actual value and the Y-axis represents the predicted value.

[0127] The two types of model confusion include confusion within the same lithology and confusion across lithologies. In confusion within the same lithology, the main reason is that the GR values ​​are similar and difficult to distinguish. In confusion across lithologies, mudstone and siltstone are weakened by the superposition of argillaceous matrix and diagenesis, forming a characteristic transition zone, which the model finds difficult to capture.

[0128] Model 2 significantly improved the classification effect of diagenetic facies compared with Model 1, reducing the misclassification of categories such as dissolved fine sandstone, compacted dense siltstone facies, and compacted dense fine sandstone by 50% to 60% and improving the accuracy by 12% to 15%. This indicates that the reconstructed curve enhances feature details and can better distinguish the microscopic differences in diagenesis.

[0129] The Matthews correlation coefficient and accuracy evaluation model performance results are as follows: As shown, in the original curve-driven model, the MCC is 0.79 and the accuracy reaches 0.84, indicating that although Model 1 can capture some diagenetic facies features, it is still insufficient in distinguishing complex facies due to feature overlap, which corresponds to the cross-misclassification phenomenon in the confusion matrix. Model 2 shows a significant performance improvement: the MCC increases to 0.89 and the accuracy increases to 0.91, indicating that the reconstructed curve effectively enhances the diagenetic facies differentiation features.

[0130] 3. Ablation Verification Experiment

[0131] To verify the impact of different wavelet basis function choices on the reconstruction curve, the experiment selected the Haar wavelet, Sym4 wavelet, and Coif4 wavelet to construct a control group. Precision, recall, and F1 score were used to quantitatively evaluate the diagenetic facies classification. Precision represents the proportion of correctly labeled diagenetic facies types among the identified types; recall represents the proportion of correctly labeled components in the actual composition; and the F1 score represents the balance between precision and recall. The calculation methods are shown in equations (15), (16), and (17).

[0132]

[0133] The recognition effect of training models with different wavelet reconstruction curves is as follows As shown, (a) is a comparison chart of precision, (b) is a comparison chart of recall, and (c) is a comparison chart of F1 scores.

[0134] Experimental results show that db4 and sym4, due to their balanced characteristics of fourth-order vanishing moments and seven-point support, significantly outperform coif4 and haar. Among them, db4 has the best balance between computational complexity and feature extraction capability; sym4's near-symmetry reduces boundary phase distortion; coif4, due to excessive smoothing of local features, introduces errors in non-smooth logging signals; haar wavelets, with only first-order vanishing moments and extremely short support, can only characterize coarse step changes and cannot provide effective details for the model, thus ranking last in performance.

[0135] Furthermore, the identification results of various diagenetic facies show that Dap, Cip, and Cap suffer from low Recall values ​​due to model overfitting caused by a small sample size. While Fap has a small sample size, its dense facies structure, characterized by significant high resistivity, allows for accurate identification. Therefore, although it performs poorly in precision, it excels in recall.

[0136] 4. Generalization ability evaluation experiment

[0137] To verify the high data volume and cross-regional generalization ability of the well logging curve reconstruction method proposed in this paper, the Fuyu oil layer in the Sanzhao Depression, a benchmark area for diagenetic facies analysis, and the Ordos region, which has a similar geological background but differs in well logging response and diagenetic facies distribution, were selected. Eight subset datasets were constructed with depth nodes of 2000, 4000, 6000, and 8000, with four sets in each region to cover the data volume gradient. The method of this invention was applied to each subset to complete wavelet multi-scale decomposition and particle swarm optimization reconstruction. The original curves and reconstructed curves were used to train the same diagenetic facies identification model. The average F1 score of the diagenetic facies was used as the index to compare and analyze the performance gain of the reconstructed curves under different data volumes in the same region and the stability of the cross-regional model under the same data volume. The generalization ability of the method was systematically verified from the perspectives of data scale adaptability and regional universality.

[0138] This paper presents the F1 scores of models trained using the original and reconstructed curves for different data volumes in the Sanchao and Ordos regions. The relationship between data volume and F1 score shows that the F1 scores of both the original and reconstructed curves increase with increasing data volume at depth nodes. The difference is lowest when the data volume is 2000, and highest when it reaches 6000, indicating a more significant advantage of the reconstructed curve's F1 score over the original curve. With further increases in data volume, the difference between the reconstructed and original curve F1 scores decreases slightly, but still shows a significant improvement. This demonstrates that the LSTM-PSO-WR reconstruction method effectively improves the F1 score for diagenetic facies identification in well logging data scenarios with varying data volumes and regions, validating the method's generalization ability across large data volumes and regions, and providing a reliable technical means for diagenetic facies analysis in complex real-world geological scenarios.

[0139] Table 1 Generalization Experiment Results

[0140] Tab.2 Generalization experiment results

[0141]

[0142] in conclusion:

[0143] (1) Wavelet weighted reconstruction can effectively enhance the high-frequency details of the logging curve in the 0.4-5Hz range, while ensuring that the low-frequency trend in the 0-0.4Hz range remains unchanged, thus improving the logging curve's ability to capture subtle features of diagenesis. Compared with the original curve, the MCC of the reconstructed curve in identifying diagenetic facies increased from 0.79 to 0.89, and the accuracy increased from 0.84 to 0.91, proving that the accuracy of the reconstructed curve has been improved.

[0144] (2) Due to its fourth-order vanishing moment and support characteristics, the db4 wavelet basis function is more suitable for well logging curve reconstruction, and its F1 score is more than 0.04 higher than other types of wavelets.

[0145] (3) The particle swarm optimization algorithm has good adaptability and generalization ability in finding the optimal reconstruction weights. It can improve the F1 score of lithogenic facies identification by more than 0.03 regardless of the amount of data or regional differences. It has application potential in complex geological scenarios.

[0146] The method of this invention can be extended to lithofacies identification of other reservoirs such as shale and volcanic rocks, and its applicability under different lithological conditions can be verified.

[0147] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

[0148] The documents cited in this invention include:

[0149] [1] Yang Shiliang. Study on logging evaluation method of Fuyu oillayer of dense reservoir Sanzhao sag of Songliao Basin[D], the Northeast Petroleum University, 2017.

Claims

1. A high-resolution reconstruction method for logging curves in tight sandstone reservoirs, characterized in that, Includes the following steps: S1. Collect multiple logging curve data, preprocess the data, and label the preprocessed logging curve data; S2. Perform wavelet transform on the preprocessed logging curve data, assign weights to the high-frequency coefficients obtained after wavelet transform, and reconstruct the logging curve. S3. Construct a training dataset based on reconstructed well logging curve data; An LSTM model is constructed and trained based on the training dataset. The LSTM model training loss function is used as the fitness function. The particle swarm optimization algorithm is used to optimize the weights assigned to the reconstructed logging curves to obtain the optimal weights. During the iteration of the particle swarm optimization algorithm, the number of wavelet decomposition layers or the participation range of the corresponding high-frequency coefficients participating in the weight optimization are adaptively adjusted according to the model's training loss response to the reconstructed logging curves at different wavelet decomposition scales, so as to perform multi-scale adaptive reconstruction for different reservoir characteristics. S4. Identify diagenetic facies by reconstructing the logging curves based on the optimal weights.

2. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 1, characterized in that, The data preprocessing described in S1 includes: noise suppression and outlier correction. based on The criterion identifies jump point noise and calculates the mean using a sliding window. with standard deviation ;like Then replace the outlier with the window's midpoint: Unified depth sequence: For low-resolution depth sequences, non-equidistant sampling data is resampled to ensure the depth sequences of each well are properly configured. With corresponding depth value Precise alignment: in and Points to be interpolated Adjacent reference depth nodes; Align the depth sequences of each well with their corresponding depths.

3. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 2, characterized in that, S2 includes the following process: First, the preprocessed well logging curve data is subjected to wavelet transform using db4 wavelet. Layer high frequency coefficient Generated through high-pass filtering and downsampling, satisfying: in These are the db4 wavelet high-pass filter coefficients; Indicates the first High-frequency detail coefficients after layer wavelet decomposition; It is the first The input sequence of layer decomposition coefficients after downsampling and shifting; For discrete-time indexing; This is the tap index of the filter; Then, weights are assigned to the high-frequency coefficients of the decomposition. ,Right now: Finally, the weighted high-frequency coefficients are upsampled, and after convolution and superposition using a db4 wavelet filter, the curve features are recovered layer by layer, and the reconstructed logging curve is obtained iteratively.

4. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 3, characterized in that, The wavelet transform decomposition has 3 to 5 layers, especially 4 layers.

5. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 3, characterized in that, S3 includes the following steps: (1) Initialize the particle swarm algorithm with the weights assigned to the reconstructed logging curves as the optimization objective; (2) The reconstructed well logging curves are stitched together to obtain the stitched dataset. , Dimensions ,in Let F be the number of depth point samples and F be the number of features. The stitched data is normalized, and T consecutive depth points are extracted along the depth direction to construct a time series, thus constructing a three-dimensional tensor. The training dataset; (3) Construct an LSTM model, train the LSTM model based on the training dataset, use the LSTM model training loss function as the fitness function, and calculate the fitness of each particle; (4) Based on the population division results and particle fitness values, determine the individual historical best position of each particle and the global best position of the particle swarm, and update the velocity and position of ordinary particles and locally optimal particles. (5) Repeat the above process for iterative calculation, and based on the training loss response of the model to the reconstructed logging curves at different wavelet decomposition scales, adaptively adjust the number of wavelet decomposition layers or the participation range of the corresponding high-frequency coefficients in the weight optimization according to the preset loss change criteria, so as to perform multi-scale adaptive reconstruction for different reservoir characteristics; until the iteration value converges or the maximum number of iterations is reached, the optimal parameter value is obtained.

6. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 5, characterized in that, The LSTM model described in step (3) includes two-layer LSTM units, with 32 hidden neurons in each layer, and layers are arranged with... The layer, through the Softmax function, outputs the probability distribution of lithogenic facies, and its state update satisfies: in , , These are the activation values ​​for the forget gate, input gate, and output gate, respectively. , These are the unit state and the hidden state. For the first Layer input; for Activation function.

7. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 6, characterized in that, The LSTM model training loss function mentioned in step (3) is: 。 8. The high-resolution reconstruction method for logging curves in tight sandstone reservoirs according to claim 7, characterized in that, Particle velocity in step (4) With position The iterative update process is as follows: in, For the first Individual particles The velocity vector at any given moment; Inertial weights; For the first Individual particles The velocity vector at any given moment; , For learning factors; , A random number within the interval [0,1]. For the first The best position in the history of each particle; For the first Individual particles Position vector at any given time; This is the globally optimal position for the entire particle swarm; For the first Individual particles The position vector at any given time.

9. A high-resolution logging curve reconstruction system for tight sandstone reservoirs, characterized in that, The system has a program module corresponding to the steps of the method described in any one of claims 1 to 8, and executes the steps in the above-described high-resolution reconstruction method of logging curves for tight sandstone reservoirs when it is run.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the high-resolution reconstruction method for logging curves in tight sandstone reservoirs as described in any one of claims 1 to 8.

Images