Continental facies shale oil reservoir lithofacies identification method and system based on time-frequency conjoint analysis

Abstract

The invention relates to a continental facies shale oil reservoir lithofacies identification method and a continental facies shale oil reservoir lithofacies identification system based on time-frequency conjoint analysis. Constructing a logging data set with a label based on a data set preprocessing model constructed by a hybrid method; comprehensively capturing multi-scale characteristics of lithofacies change based on a logging curve data adaptive fusion model of TFAA-TCN; a logging curve time-frequency information extraction model based on Mama-MHSA is combined with Mama long-term dependency relationship capture and multi-head attention mechanism MHSA feature screening, the long-term dependency relationship of spatial information of a logging curve and attention weights among different features are mined, and continental facies shale oil reservoir lithofacies identification is carried out. According to the method, the accuracy of continental facies shale oil reservoir lithofacies identification is effectively improved, the extraction and fusion mode of logging curve characteristics is optimized, and the adaptive capacity to a complex reservoir environment is enhanced.

Description

Technical field:

[0001] The present invention relates to the field of geological exploration technology, and in particular to a method and system for identifying lithofacies of continental shale oil reservoirs based on time-frequency joint analysis. Background technology:

[0002] With the continuous advancement of unconventional oil and gas exploration and development technologies, continental shale oil resources have demonstrated tremendous development potential and commercial value, becoming a crucial component of China's future energy strategy. Lithofacies, the fundamental unit for describing and characterizing unconventional oil and gas reservoirs, directly reflects the reservoir's storage properties and oil-bearing characteristics. Accurate identification of lithofacies is crucial for the detailed evaluation and efficient development of shale oil reservoirs.

[0003] Currently, lithofacies identification of continental shale oil reservoirs relies primarily on core sampling observations, laboratory laboratory analysis, and well log analysis. Because core and laboratory data are difficult and costly to obtain, well log data is relatively easy to obtain and contains rich lithofacies information. Therefore, lithofacies identification based on well log data has become an important method. Geologists typically rely on experience and statistical analysis to interpret well logs to identify reservoir lithofacies types.

[0004] In recent years, the development of computer artificial intelligence technology has promoted the advancement of well log facies identification methods. Compared with traditional core description and experimental analysis methods, neural network methods based on well logs can improve identification and reduce the influence of human subjective factors. However, they have the following shortcomings in the process of identifying continental shale oil reservoir facies:

[0005] 1. Insufficient time-frequency feature extraction capabilities: Traditional well log lithofacies identification methods rely primarily on raw time series data and fail to fully utilize the frequency domain information contained in the well logs. The lack of joint time-frequency analysis prevents the model from fully capturing the multi-scale characteristics of lithofacies variations, affecting identification accuracy.

[0006] 2. Current methods primarily rely on template matching during data anomaly detection, which can easily misidentify samples that don't meet predefined characteristics as outliers. This leads to high recognition error rates and computational complexity. While the isolation forest algorithm effectively addresses these issues, it requires a pre-defined threshold for the number of outlier outputs, reducing the algorithm's accuracy. Furthermore, detected outliers are often directly deleted, compromising the continuity and integrity of the well logging data.

[0007] 3. Modal fusion methods have limitations: Existing methods usually use simple splicing or weighted averaging when fusing multiple logging attributes. They fail to fully consider the different contributions of different logging curves to lithofacies identification, resulting in low information fusion efficiency and affecting the classification ability of the model.

[0008] 4. Insufficient ability to capture long-term dependencies: When processing well logging curves, traditional neural network methods often find it difficult to simultaneously take into account local feature extraction and long-term dependency modeling, resulting in the loss of key information in the lithofacies evolution process.

[0009] 5. Lack of adaptive weight adjustment mechanism: In the process of lithofacies identification, the contribution of logging curve characteristics under different geological conditions is different. Existing methods are difficult to adaptively adjust the weights, resulting in weak adaptability and generalization ability of the model in different reservoir environments. Summary of the invention:

[0010] The purpose of the present invention is to provide a method for identifying continental shale oil reservoir lithofacies based on time-frequency joint analysis. This method for identifying continental shale oil reservoir lithofacies based on time-frequency joint analysis is used to solve the problems existing in existing continental shale oil reservoir lithofacies identification methods, such as insufficient time-frequency feature extraction, limited modal fusion methods, weak ability to capture long-term dependencies, and lack of an adaptive weight adjustment mechanism. Another purpose of the present invention is to provide a method for identifying continental shale oil reservoir lithofacies based on time-frequency joint analysis.

[0011] The technical solution adopted by the present invention to solve the technical problem is: the method for identifying the lithofacies of continental shale oil reservoirs based on time-frequency joint analysis includes the following steps:

[0012] Step 1: Obtain core experimental data from continental shale oil reservoirs. Based on a dataset preprocessing model constructed using a hybrid approach, perform lithofacies calibration on the well logging data to create a dataset, standardize the well logging curves, process outliers, and filter features to obtain a labeled well logging dataset. The dataset preprocessing model uses an improved isolation forest method based on the quartile principle to calculate thresholds to detect outliers, and uses a dynamic window-modified Lagrangian interpolation method to replace abnormal data. The well logging curve features are filtered using an improved ReliefF algorithm.

[0013] Step 2: Based on the TFAA-TCN well logging curve data adaptive fusion model, a time-frequency feature joint analysis module is used to perform wavelet transform on the well logging curve data in the labeled well logging dataset to extract frequency domain information, and the frequency domain information and time-frequency features are adaptively fused;

[0014] Step 3: Based on the Mamba-MHSA well logging curve time-frequency information extraction model, by combining Mamba long-term dependency capture with the multi-head attention mechanism MHSA feature screening, the long-term dependency of the spatial information of the well logging curve and the attention weights between different features are mined to obtain the continental shale oil reservoir lithofacies identification model;

[0015] Step 4: Train the continental shale oil reservoir lithofacies identification model to obtain the optimal continental shale oil reservoir lithofacies identification model for continental shale oil reservoir lithofacies identification.

[0016] Step 1 in the above scheme is specifically as follows:

[0017] S1.1 Dataset Creation: First, nine conventional logging curves, including depth and natural gamma, acoustic transit time, deep lateral resistivity, shallow lateral resistivity, microsphere focused resistivity, lithologic density, natural potential, compensated neutron, and wellbore, were selected as input data. The corresponding lithofacies obtained from core experiments were used as label data to create a dataset, which was divided into a validation set and a test set with a ratio of 7:3.

[0018] S1.2 Dataset preprocessing, including data standardization and outlier processing; well logging curve standardization, using the well closest to the regional coordinate center as the reference well, and performing offset correction using formula (1);

[0019] WellA norm =WellA+(Well base -WellA) (1)

[0020] Among them, WellA norm Represents the corrected logging curve data, WellA represents the logging curve data of the corrected well, Well base Well logging data representing the benchmark well;

[0021] Outlier processing: use the improved isolation forest method based on the quartile principle to calculate the threshold to detect outliers, and use the Lagrange interpolation method improved by dynamic window to replace abnormal data;

[0022] Quartile principle: First, for the distribution of each well logging curve data in the data set, find the upper quartile Q1 and the lower quartile Q3. Then, based on this, define Q1-1.5(Q3-Q1) and Q3+1.5(Q3-Q1) as the lower limit and upper limit of the data. The data outside the upper and lower limits are defined as abnormal data. The abnormal number threshold of the isolation forest method is determined by formula (2):

[0023]

[0024] where count un represents the number of abnormal values ​​in the logging curve, m represents the number of logging curves, x i Represents the i-th logging curve data;

[0025] For outliers, the Lagrangian interpolation method improved by dynamic window is used to fill in the outliers. First, the kernel density estimation (KDE) is used to calculate the data distribution density in the depth direction. Then, a small window is used in the data-intensive area and a large window is used in the data-sparse area to improve the local adaptability of the interpolation. Finally, Lagrangian interpolation is used within the dynamic window to calculate the value of the missing point to approximate the outlier point data to be inserted. The calculation formulas are shown in (3)-(5):

[0026]

[0027]

[0028] Where N is the known data point, h is the smoothing bandwidth, K is the kernel function, w(x) is the window size obtained by density calculation, and w max is the maximum value of the window, w min is the window minimum, ρ(x) is the density at the data point x, ρ max is the maximum value of the density in the data set, ρ min is the minimum value of density in the data set;

[0029] S1.3 Feature screening: By improving the ReliefF algorithm, the feature weights of the data input are calculated and proportionalized, the feature importance is highlighted according to the feature weight ratio, and important logging curve features are retained; based on the original Relief algorithm, the improved ReliefF algorithm uses a stratified equidistant sampling method to construct a lithofacies feature screening sample set, and the cosine similarity distance CS method is used to replace the Euclidean distance to enhance sample representativeness.

[0030] Step 2 in the above solution is specifically as follows:

[0031] S2.1 uses the Daubechies wavelet basis to decompose the logging sequence data line by line to obtain the corresponding frequency domain information. The frequency domain information consists of two parts: low-frequency components and high-frequency components. Since the length of the converted frequency domain information is half of the length of the original logging sequence information, the binary interpolation reconstruction method is used to process the frequency domain information to make it consistent with the length of the original logging sequence information, thus obtaining the complete frequency domain information.

[0032] S2.2 Adaptive fusion of time-frequency features: First, the TCN network is used to extract sequence features from the time domain information and frequency domain information respectively to obtain dataT and dataF, while maintaining the sequence feature extraction sequence following the causal relationship rule while obtaining the same output length as the input length; secondly, the weight W is calculated through global average pooling, one-dimensional convolution and sigmoid activation function according to formula (6); then, dataT and dataF are added according to formula (7) to obtain the final time-frequency fusion feature dataTF;

[0033] W(dataT,dataF)=σ(Conv1D(GAP(dataT+dataF))) (10)

[0034]

[0035] Step 3 in the above solution is specifically as follows:

[0036] S3.1Mamba captures long-term dependencies. Using dataTF as input, Mamba overcomes the limitations of traditional SSM by introducing time-varying parameters, enabling efficient processing of long sequence data. It also introduces a dynamic scanning direction mechanism, using horizontal scanning in data-dense areas and vertical scanning in data-sparse areas, to improve the model's adaptability to different data distributions.

[0037] S3.2 Multi-head attention mechanism MHSA feature screening. The features processed by the Mamba module are further screened by MHSA for feature data related to lithofacies. MHSA calculates the attention weights between different features to highlight important features and suppress irrelevant features. Finally, the extracted features are sent to the output layer for lithofacies identification of continental shale oil reservoirs.

[0038] A terrestrial shale oil reservoir lithofacies identification system based on time-frequency joint analysis includes a data set preprocessing model constructed based on the hybrid method, a well logging curve data adaptive fusion model based on TFAA-TCN, a terrestrial shale oil reservoir lithofacies identification model and a user management unit.

[0039] The dataset preprocessing model built based on the hybrid method is used to receive and store well logging curve data for lithofacies calibration, standardization, outlier and feature screening, and generate labeled well logging datasets;

[0040] The adaptive fusion model of well logging curve data based on TFAA-TCN is used to obtain reservoir characteristics at different scales based on wavelet transform, and the time domain and frequency domain features are fused to form the final time-frequency features.

[0041] The continental shale oil reservoir lithofacies identification model uses the Mamba-MHSA architecture for feature extraction and classification. Mamba processes the temporal patterns of well logging curves to enhance the modeling capabilities of long-term dependencies. MHSA calculates attention weights between different features to highlight important features and achieve continental shale oil reservoir lithofacies identification.

[0042] User management unit, used for system user information management and user authority settings.

[0043] Beneficial effects:

[0044] 1. The present invention improves the isolation forest algorithm based on the quartile principle, improves the objectivity of the isolation forest threshold, realizes adaptive detection of abnormal data in well logging curves, avoids the deficiency of pre-setting the number of abnormal outputs in the traditional isolation forest algorithm, and improves the accuracy of anomaly detection.

[0045] 2. This method uses a dynamic window-based Lagrangian interpolation method to complete detected outliers, effectively eliminating the information gaps caused by directly deleting outliers and improving the flexibility and adaptability of interpolation. This fundamentally ensures the smooth continuity of the data sequence, providing high-quality input for subsequent feature extraction and lithofacies identification.

[0046] 3. In the feature screening stage, the present invention uses a stratified equidistant sampling method to construct a lithofacies feature screening sample set, and improves the ReliefF algorithm by replacing the Euclidean distance with the cosine similarity distance (CS) method, thereby enhancing sample representativeness. It can efficiently and robustly screen out important well logging curve features and provide an interpretable feature importance ranking, making it suitable for feature screening tasks of large-scale, high-dimensional data.

[0047] 4. This paper utilizes the TFAA-TCN module to perform frequency-domain decomposition of logging curves using a Daubechies wavelet basis and employs binary interpolation reconstruction techniques to obtain complete time-frequency data. TCN extracts local features from each of these time-frequency data, and calculates the weight W through global average pooling, one-dimensional convolution, and a sigmoid activation function. This adaptively fuses time-domain and frequency-domain information, comprehensively capturing the multi-scale characteristics of lithofacies variations and addressing the insufficient feature extraction capabilities of traditional single-time series data processing methods.

[0048] The present invention constructs a Mamba-MHSA neural network. A dynamic scanning direction mechanism is introduced into the Mamba module, using horizontal scanning in data-dense areas and vertical scanning in data-sparse areas. This improves the model's adaptability to diverse data distributions and enables effective feature extraction within the global context. Furthermore, a multi-head self-attention module (MHSA) is used to further screen key features related to lithofacies, significantly enhancing the model's generalization and adaptability to complex reservoir environments.

[0049] 6. The trained lithofacies identification model for continental shale oil reservoirs can accurately identify lithofacies using data from uncored wells, breaking through the limitations of traditional reliance on core data and providing strong technical support for the detailed evaluation of continental shale oil reservoirs and the efficient development of unconventional oil and gas resources. Description of the drawings:

[0050] Figure 1The process of lithofacies identification of continental shale oil reservoirs based on time-frequency joint analysis.

[0051] Figure 2 It is a time-frequency adaptive analysis module for continental shale oil reservoir data.

[0052] Figure 3 Mamba-MHSA network structure for lithofacies identification of continental shale oil reservoirs.

[0053] Figure 4 Confusion matrix for lithofacies identification percentage of continental shale oil reservoirs. Specific implementation method:

[0054] The present invention will be further described below in conjunction with the appended drawings:

[0055] This method and system for identifying lithofacies in continental shale oil reservoirs based on joint time-frequency analysis integrates TFAA-TCN, an improved isolation forest, a Lagrangian interpolation method with a dynamic window, a ReliefF algorithm with an improved cosine similarity distance, and Mamba-MHSA technology. First, the well log sequence data is preprocessed using the improved isolation forest and Lagrangian interpolation methods with a dynamic window. This data is then fed into the TFAA-TCN joint time-frequency analysis module, where it undergoes frequency-domain analysis and adaptively fuses the time-frequency information. This addresses the issue of traditional fusion methods that fail to fully account for the varying importance of information from different modalities. Secondly, the Mamba-MHSA neural network extracts features from the fused time-frequency information of the well log curves, enabling the model to better handle long-term dependencies while also taking into account memory capacity. Finally, the data is fed into the output layer to obtain lithofacies classification results.

[0056] This lithofacies identification method for continental shale oil reservoirs based on time-frequency joint analysis includes (1) a data set preprocessing model constructed based on a hybrid method, (2) a well logging curve data adaptive fusion model based on TFAA-TCN, (3) a well logging curve time-frequency information extraction model based on Mamba-MHSA for lithofacies feature analysis, and (4) training and application of the continental shale oil reservoir lithofacies identification model.

[0057] Step 1: Dataset preprocessing model based on hybrid algorithm, such as Figure 1 As shown in S101, it includes dataset creation, data preprocessing (standardization and outlier processing) and feature screening.

[0058] S1.1 Dataset Preparation: Nine conventional well logs, including depth and natural gamma ray, acoustic transit time, deep lateral resistivity, shallow lateral resistivity, microsphere focused resistivity, lithologic density, spontaneous potential, compensated neutron, and wellbore, were selected as input data. The corresponding lithofacies obtained from core experiments were used as label data to create a dataset, with a validation and test set split at a ratio of 7:3. Based on factors such as organic matter enrichment, mineral composition, microstratification, and reservoir properties, and combined with bedding descriptions, the following types of lithofacies were selected: siliceous laminated organic shale (SL-T), calcareous laminated organic shale (CM-T), clayey massive organic shale (CL-B), mixed bioturbated medium organic shale (MX-B), silica-calcium mixed laminated organic shale (SF-T), microfracture-developed mixed shale (FR-M), high-organic laminated shale (OP-T), volcanic ash interbedded massive shale (AS-B), and other transitional lithofacies (OT).

[0059] S1.2 Dataset preprocessing, including data normalization and outlier processing. Well logging curve normalization, the well closest to the regional coordinate center is used as the reference well (Well base ), and the logging data of other wells are offset corrected using formula (1).

[0060] WellA norm =WellA+(Well base -WellA) (1)

[0061] Among them, WellA norm Represents the corrected logging curve data, WellA represents the logging curve data of the corrected well, Well base Well logging data representing the benchmark well.

[0062] Outlier processing uses the improved isolation forest method based on the quartile principle to calculate the threshold to detect outliers, and uses the Lagrange interpolation method improved by the dynamic window to replace abnormal data. First, for the distribution of each well logging curve data in the data set, find the upper quartile (Q1) and lower quartile (Q3), and then define Q1-1.5 (Q3-Q1) and Q3+1.5 (Q3-Q1) as the lower and upper limits of the data based on this. Data outside the upper and lower limits are defined as abnormal data. Through formula (2), we can get the abnormal number threshold:

[0063]

[0064] where count un represents the number of abnormal values ​​in the logging curve, m represents the number of logging curves, x i Represents the i-th logging curve.

[0065] For outliers, Lagrangian interpolation based on a dynamic window is used to fill in the outliers. First, kernel density estimation (KDE) is used to calculate the data distribution density in the depth direction. Then, a small window is used in data-dense areas and a large window is used in data-sparse areas to improve the local adaptability of the interpolation. Finally, Lagrangian interpolation is used within the dynamic window to calculate the values ​​of the missing points and approximate the outlier data to be inserted. The calculation formulas are shown in (3)-(5).

[0066]

[0067] Where N is the known data point, h is the smoothing bandwidth, K is the kernel function, w(x) is the window size obtained by density calculation, and w max is the maximum value of the window, w min is the window minimum, ρ(x) is the density at the data point x, ρ max is the maximum value of the density in the data set, ρ min is the minimum density in the dataset.

[0068] S1.3 Feature screening: By improving the ReliefF algorithm, the feature weights of the data input are calculated and proportionalized. The feature importance is highlighted according to the feature weight ratio, and important well logging curve features are retained. Based on the original Relief algorithm, the improved ReliefF algorithm uses a stratified equidistant sampling method to construct a lithofacies feature screening sample set, and replaces the Euclidean distance with the cosine similarity distance (CS) method to enhance sample representativeness. First, a weight vector is initialized for each feature of the data set. Then, a stratified equidistant sampling method is used to construct a lithofacies feature screening sample set. Then, based on formula (6), the k nearest guessed neighbors (Near-Hit, NH) are found, and the well logging feature weight ratio is calculated according to formulas (7)-(9).

[0069]

[0070]

[0071] Where CS(A,B) is the cosine similarity distance between sample A and sample B, x 1i is the i-th eigenvalue of sample A, x 2i is the i-th eigenvalue of sample B, diff(A, R i , NH j ) is the measure of feature A in sample R i and its neighbor NH j The degree of difference between them, k is the number of neighbors, which is used to control the range and number of neighbors considered, n is the total number of samples, p(class(Ri )) represents sample R i Prior probability of the category or the proportion in the sample set, NM j (C) represents the sample R i The neighboring samples from different categories C are used to measure the difference between feature A and samples of different categories, W(A i )The weight of the i-th eigenvalue of sample A.

[0072] S1.1 was used to construct the dataset for this case study, totaling approximately 15,000 records, with a training set and test set split at a ratio of 7:3. S1.2 selected the Gu 1* well as the base well and standardized the log curves of the other 29 wells, completing the construction of the dataset for this case study. S1.2 processed 276 data points for outliers. Feature screening in S1.3 revealed significant differences between the shallow lateral resistivity, compensated neutron, and microsphere focused resistivity curves compared to the other six curves. Therefore, these six other log curves and depth data were selected as input for subsequent model training.

[0073] Step 2: Adaptive fusion model of well logging curve data based on TFAA-TCN, such as Figure 1 As shown in S102 in FIG, it includes wavelet frequency domain information conversion and time-frequency feature adaptive fusion. The time-frequency adaptive analysis module of continental shale oil reservoir data is as follows: Figure 2 The time-frequency adaptive analysis module for continental shale oil reservoir data is an adaptive fusion model of well logging curve data based on TFAA-TCN.

[0074] S2.1 For the six logging curves in the data set, the Daubechies (db1) wavelet basis is used to decompose the logging sequence data one by one to obtain the corresponding frequency domain information (including low-frequency components and high-frequency components). The frequency domain information is then processed using the binary interpolation reconstruction method to make it consistent with the length of the original logging sequence information.

[0075] S2.2 Adaptive fusion of time-frequency features. First, the TCN network is used to fusion the time-frequency information dataT and dataF of the logging curve. Secondly, the weight W is calculated by global average pooling, one-dimensional convolution and sigmoid activation function according to formula (10). Then, dataT and dataF are added according to formula (11) to obtain the final time-frequency fusion feature dataTF.

[0076] W(dataT,dataF)=σ(Conv1D(GAP(dataT+dataF))) (10)

[0077]

[0078] Step 3: The lithofacies feature extraction model based on the Mamba-MHSA well logging curve characteristics, that is, the continental shale oil reservoir lithofacies identification model, also known as the Mamba-MHSA network structure for continental shale oil reservoir lithofacies identification, as shown in Figure 1 S103 in the figure includes Mamba long-term dependency capture and multi-head attention mechanism (MHSA) feature screening, and the network result is shown in the figure below. Figure 3 shown.

[0079] S3.1 Mamba captures long-term dependencies. The Mamba model is constructed, and the dataTF obtained in step 2 is used as Mamba's input. The Mamba model captures the long-term contextual information of the input dataset (depth data plus six well logs). Specifically, the Mamba module introduces time-varying parameters to overcome the limitations of traditional SSMs, enabling efficient processing of long-sequence data. It also incorporates a dynamic scanning direction mechanism, using horizontal scanning in data-dense areas and vertical scanning in data-sparse areas, to improve the model's adaptability to diverse data distributions and effectively address the vanishing gradient problem of traditional time series algorithms when handling long-term dependencies. dataTF serves as Mamba's input.

[0080] S3.2 Multi-Head Attention (MHSA) Feature Screening: Features processed by the Mamba network are further fed into the MHSA, where attention weights between different features are calculated to further extract feature information. This feature data is then passed to the fully connected layer. Specifically, features processed by the Mamba network are further filtered by the MHSA for lithofacies-related features. By calculating attention weights between different features, the MHSA highlights important features and suppresses irrelevant ones, thereby improving the model's generalization capabilities. Finally, the extracted features are fed into the output layer for lithofacies identification of continental shale oil reservoirs.

[0081] Step 4: Training and application of continental shale oil reservoir lithofacies identification model, such as Figure 1 S104 includes model parameter training and application to lithofacies identification of uncored wells.

[0082] S4.1 Model parameter training: Based on the preprocessed continental shale oil reservoir lithofacies identification data, the data set is split into sub-datasets of every 1,000 data items using a sliding window method, with a step size of 500 data items. Each sub-dataset is used as a whole to train the continental shale oil reservoir identification model to obtain the corresponding model weight data.

[0083] S4.2 is used for the lithofacies identification of uncored wells. The depth and logging curve data of the continental shale oil reservoir of the uncored well are input to realize data preprocessing, time-frequency data fusion and calculation based on the trained model to obtain the corresponding reservoir lithofacies.

[0084] Through the training of the continental shale oil reservoir lithofacies recognition model and its application to the test data set, the confusion matrix obtained is as follows: Figure 4 shown.

[0085] A terrestrial shale oil reservoir lithofacies identification system based on time-frequency joint analysis, comprising:

[0086] The data preprocessing unit, a dataset preprocessing model built using a hybrid approach, receives and stores well logging data, including nine conventional well logging curves: natural gamma ray, acoustic transit time, deep lateral resistivity, shallow lateral resistivity, microsphere focused resistivity, lithologic density, spontaneous potential, compensated neutron, and caliper. This unit also performs lithofacies calibration, standardization, and outlier and feature screening to ensure data quality. This lithofacies calibration is also performed with a small amount of core data to generate a labeled well logging dataset, providing foundational data for subsequent model training.

[0087] The time-frequency feature joint analysis unit is an adaptive fusion model of well logging curve data based on TFAA-TCN. It obtains reservoir characteristics at different scales based on wavelet transform and fuses time domain and frequency domain features to form the final time-frequency features.

[0088] The continental shale oil reservoir lithofacies identification model uses the Mamba-MHSA architecture for feature extraction and classification. Mamba processes the temporal patterns of well log curves to enhance the modeling of long-term dependencies, while MHSA calculates attention weights between different features to highlight important features. This model parameter training enables continental shale oil reservoir lithofacies identification.

[0089] User management unit, used for system user information management and user authority settings.

[0090] This application also provides a memory, a processor, and a program for identifying terrestrial shale oil reservoir lithofacies stored in the memory and executable on the processor. When the processor executes the program, the steps of one of the aforementioned methods are implemented. The aforementioned computer-readable storage medium includes a mobile storage device, ROM, RAM, flash memory, or other medium capable of storing program code.

[0091] The present invention improves the isolation forest through the quartile principle and uses the Lagrangian interpolation method based on dynamic windows to replace abnormal data, avoiding the manual setting of parameters, improving the efficiency and effectiveness of the algorithm in processing logging curves, and ensuring the continuity and integrity of the data. By designing a method and system for identifying terrestrial shale oil reservoir lithofacies based on time-frequency joint analysis, the present invention first uses the TFAA-TCN time-frequency feature joint analysis module to perform wavelet transform frequency domain information extraction and time-frequency feature adaptive fusion on the logging curve, fully exploring the time-frequency characteristics of the logging curve, and solving the problem of insufficient single feature information recognition in traditional fusion methods. Then, Mamba is used to extract features from the time-frequency fusion information of the logging curve, enhancing the model's ability to capture local patterns and long-term dependencies. In addition, by introducing the multi-head self-attention module (MHSA), feature data related to lithofacies are screened out, improving the generalization ability of the model and making it more applicable in different reservoir environments. The present invention effectively improves the accuracy of lithofacies identification of continental shale oil reservoirs, optimizes the extraction and fusion methods of logging curve features, and enhances the model's adaptability to complex reservoir environments, providing strong technical support for the efficient development of unconventional oil and gas resources.

Claims

1. A method for identifying terrestrial shale oil reservoir lithofacies based on time-frequency joint analysis, characterized in that The steps include: Step 1: Obtain core experimental data from continental shale oil reservoirs. Based on the dataset preprocessing model constructed using the hybrid method, perform lithofacies calibration on the well logging curve data to create a dataset, standardize the well logging curves, process outliers, and filter features to obtain a labeled well logging dataset. The dataset preprocessing model uses an improved isolation forest method based on the quartile principle to calculate the threshold to detect outliers, and uses a dynamic window improved Lagrange interpolation method to replace abnormal data; the improved ReliefF algorithm is used to screen the characteristics of the well logging curve; Step 2: Based on the TFAA-TCN well logging curve data adaptive fusion model, a time-frequency feature joint analysis module is used to perform wavelet transform, frequency domain information extraction and time-frequency feature adaptive fusion on the well logging curves in the labeled well logging data set; Step 3: Based on the Mamba-MHSA well logging curve time-frequency information extraction model, by combining Mamba long-term dependency capture with the multi-head attention mechanism MHSA feature screening, the long-term dependency of the spatial information of the well logging curve and the attention weights between different features are mined to obtain the continental shale oil reservoir lithofacies identification model; Step 4: Train the continental shale oil reservoir lithofacies identification model to obtain the optimal continental shale oil reservoir lithofacies identification model for continental shale oil reservoir lithofacies identification.

2. The method for identifying terrestrial shale oil reservoir lithofacies based on time-frequency joint analysis according to claim 1, characterized in that: The step 1 is specifically as follows: S1.1 Dataset Creation: First, nine conventional logging curves, including depth and natural gamma, acoustic transit time, deep lateral resistivity, shallow lateral resistivity, microsphere focused resistivity, lithologic density, natural potential, compensated neutron, and wellbore, were selected as input data. The corresponding lithofacies obtained from core experiments were used as label data to create a dataset, which was divided into a validation set and a test set with a ratio of 7:

3. S1.2 Dataset preprocessing, including data standardization and outlier processing; well logging curve standardization, using the well closest to the regional coordinate center as the reference well, and performing offset correction using formula (1); WellA norm =WellA+(Well base -WellA) (1) Among them, WellA norm Represents the corrected logging curve data, WellA represents the logging curve data of the corrected well, Well base Well logging data representing the benchmark well; Outlier processing: use the improved isolation forest method based on the quartile principle to calculate the threshold to detect outliers, and use the Lagrange interpolation method improved by dynamic window to replace abnormal data; Quartile principle: First, for the distribution of each well logging curve data in the data set, find the upper quartile Q1 and the lower quartile Q3. Then, based on this, define Q1-1.5 (Q3-Q1) and Q3+1.5 (Q3-Q1) as the data lower limit and data upper limit. The data outside the upper and lower limits are defined as abnormal data. The abnormal number threshold of the isolation forest method is determined by formula (2): where count un represents the number of abnormal values ​​in the logging curve, m represents the number of logging curves, x i represents the i-th logging curve; For outliers, the Lagrangian interpolation method improved by dynamic window is used to fill in the outliers. First, the kernel density estimation (KDE) is used to calculate the data distribution density in the depth direction. Then, a small window is used in the data-intensive area and a large window is used in the data-sparse area to improve the local adaptability of the interpolation. Finally, Lagrangian interpolation is used within the dynamic window to calculate the value of the missing point to approximate the outlier point data to be inserted. The calculation formulas are shown in (3)-(5): Where N is the known data point, h is the smoothing bandwidth, K is the kernel function, w(x) is the window size obtained by density calculation, and w max is the maximum value of the window, w min is the window minimum, ρ(x) is the density at the data point x, ρ max is the maximum value of the density in the data set, ρ min is the minimum value of density in the data set; S1.3 Feature screening: By improving the ReliefF algorithm, the feature weights of the data input are calculated and proportionalized. The importance of features is highlighted according to the feature weight ratio, and important logging curve features are retained; The improved ReliefF algorithm is based on the original Relief algorithm. The stratified equal-interval sampling method is used to construct the lithofacies feature screening sample set, and the cosine similarity distance CS method is used to replace the Euclidean distance to enhance the sample representativeness.

3. The method for identifying terrestrial shale oil reservoir lithofacies based on time-frequency joint analysis according to claim 2, characterized in that: The step 2 is specifically as follows: S2.1 uses the Daubechies wavelet basis to decompose the logging sequence data line by line to obtain the corresponding frequency domain information. The frequency domain information consists of two parts: low-frequency components and high-frequency components. Since the length of the converted frequency domain information is half of the length of the original logging sequence information, the binary interpolation reconstruction method is used to process the frequency domain information to make it consistent with the length of the original logging sequence information, thus obtaining the complete frequency domain information. S2.2 Adaptive fusion of time-frequency features; First, the TCN network is used to extract sequence features from time domain information and frequency domain information respectively to obtain dataT and dataF, while maintaining the sequence feature extraction sequence following the causal relationship rule while obtaining the same output length as the input length; secondly, the weight W is calculated according to formula (6) through global average pooling, one-dimensional convolution and sigmoid activation function; Then, dataT and dataF are added together according to formula (7) to obtain the final time-frequency fusion feature dataTF; W(dataT,dataF)=σ(Conv1D(GAP(dataT+dataF))) (10) 4. The method for identifying terrestrial shale oil reservoir lithofacies based on time-frequency joint analysis according to claim 3, characterized in that: The step three is specifically as follows: S3.1 Mamba captures long-term dependencies. Using dataTF as input, Mamba introduces time-varying parameters to efficiently process long sequence data. It also introduces a dynamic scanning direction mechanism, using horizontal scanning in data-dense areas and vertical scanning in data-sparse areas, to improve the model's adaptability to different data distributions. S3.2 Multi-head attention mechanism MHSA feature screening. The features processed by the Mamba module are further screened by MHSA for feature data related to lithofacies. MHSA calculates the attention weights between different features to highlight important features and suppress irrelevant features. Finally, the extracted features are sent to the output layer for lithofacies identification of continental shale oil reservoirs.

5. A terrestrial shale oil reservoir lithofacies identification system based on time-frequency joint analysis according to claim 4, characterized in that: The identification system includes the data set preprocessing model constructed based on the hybrid method, the well logging curve data adaptive fusion model based on TFAA-TCN, the continental shale oil reservoir lithofacies identification model and a user management unit; The dataset preprocessing model built based on the hybrid method is used to receive and store well logging curve data for lithofacies calibration, standardization, outlier and feature screening, and generate labeled well logging datasets; The adaptive fusion model of well logging curve data based on TFAA-TCN is used to obtain reservoir characteristics at different scales based on wavelet transform, and the time domain and frequency domain features are fused to form the final time-frequency features. The continental shale oil reservoir lithofacies identification model uses the Mamba-MHSA architecture for feature extraction and classification. Mamba processes the temporal patterns of well logging curves to enhance the modeling capabilities of long-term dependencies. MHSA calculates attention weights between different features to highlight important features and achieve continental shale oil reservoir lithofacies identification. User management unit, used for system user information management and user authority settings.

Images