A method and system for automatic interpretation of post-casing saturation logging based on multi-modal deep learning fusion

Abstract

This invention discloses an automatic spectral interpretation method and system for casing well saturation logging based on multimodal deep learning fusion, belonging to the field of oil and gas geophysical logging technology. The method includes: acquiring gamma-ray energy spectrum data and time spectrum data of the formation to be tested; preprocessing the energy spectrum data and time spectrum data to obtain standardized input features; inputting the standardized input features into a pre-trained multimodal deep learning model; the multimodal deep learning model outputting spectral interpretation results, including relative production of formation elements and macroscopic capture cross section Σ; and calculating the formation oil saturation based on the spectral interpretation results. This invention also introduces a physical constraint loss function to ensure that the solution conforms to formation physical laws and adapts to small sample conditions through a transfer learning strategy. Compared with existing technologies, this invention significantly improves spectral interpretation accuracy and processing efficiency, and can be widely applied to the evaluation of remaining oil saturation in casing wells.

Description

Technical Field

[0001] This invention relates to the field of oil and gas geophysical logging technology, specifically to gamma energy spectrum and time spectrum data obtained by pulsed neutron logging in the evaluation of formation saturation in casing wells, and particularly to an automatic spectral interpretation method and system for post-casing saturation logging based on multimodal deep learning fusion. Background Technology

[0002] In the later stages of oil and gas field development, assessing the remaining oil saturation of the casing well formation is crucial for formulating potential tapping strategies. Pulsed neutron logging technology transmits high-energy neutrons into the formation, recording the inelastic scattered gamma rays and captured gamma rays generated by the interaction between neutrons and formation atomic nuclei. This yields two key types of data: gamma spectral data (reflecting the count distribution of gamma rays at different energies, including characteristic peak information of elements such as C, O, Si, Ca, Fe, Cl, and H) and temporal spectral data (recording the count sequence of thermal neutrons or captured gamma rays decaying over time, allowing for the extraction of thermal neutron lifetime or macroscopic capture cross-section Σ).

[0003] Traditional spectral analysis methods mainly include least squares stripping, weighted multichannel analysis, and nonlinear fitting. These methods have the following limitations: ① Mismatch between the standard spectrum and the measured spectrum, with actual well conditions causing spectral distortion that is difficult to adaptively correct; ② Severe overlap of characteristic peaks, such as Si and Ca, S and Cl, limiting the accuracy of traditional linear decomposition; ③ Significant impact of statistical fluctuations, especially the low tunnel count rate in the later stages of the time spectrum, making traditional fitting susceptible to noise interference; ④ Insufficient information utilization, with the complementarity of the energy spectrum and time spectrum not effectively constrained, leading to high interpretability issues; ⑤ Low processing efficiency, with high-precision spectral analysis requiring complex iterative fitting, making it difficult to meet the needs of continuous processing of massive depth points.

[0004] In recent years, deep learning technology has shown potential in well logging data processing, but there have been no reports of applying multimodal deep learning technology systems to the joint automatic interpretation of energy spectrum and time spectrum of post-casing saturation logging. Summary of the Invention

[0005] To address the problems of overlapping characteristic peaks, poor noise resistance, and insufficient information utilization in traditional post-casing saturation logging spectral interpretation methods when processing energy and time spectra, this invention discloses an automatic post-casing saturation logging spectral interpretation method and system based on multimodal deep learning fusion. By constructing an end-to-end deep learning model, it simultaneously learns the local characteristic peak information of the energy spectrum, the temporal decay characteristics of the time spectrum, and the higher-order interaction information between the two, thereby achieving joint automatic spectral interpretation of the energy and time spectra, significantly improving spectral accuracy and processing efficiency.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] An automatic spectral interpretation method for post-jacket saturation logging based on multimodal deep learning fusion includes the following steps:

[0008] Acquire gamma-ray spectral data and time-spectral data of the formation to be tested;

[0009] The energy spectrum data and time spectrum data are preprocessed to obtain standardized input features;

[0010] Standardized input features are fed into a pre-trained multimodal deep learning model;

[0011] The multimodal deep learning model outputs spectral results, including relative yields of formation elements and macroscopic capture sections Σ;

[0012] The formation oil saturation was calculated based on the spectral results.

[0013] Optionally, the preprocessing steps for energy spectrum data and time spectrum data may include one or more combinations of energy scale correction, time scale correction, background subtraction, normalization, and smoothing and noise reduction.

[0014] Optionally, the step of inputting standardized input features into a pre-trained multimodal deep learning model, which includes:

[0015] The energy spectrum processing branch is used to extract characteristic peak information and spectral shape features from gamma energy spectrum data;

[0016] The time spectrum processing branch is used to extract the decay features and temporal dependencies of time spectrum data;

[0017] The multimodal feature fusion module is used for high-order interactive fusion of energy spectrum features and time spectrum features;

[0018] The decoding output module is used to map the fused features to the target spectral parameters.

[0019] Optionally, the energy spectrum processing branch employs a one-dimensional convolutional neural network combined with a residual connection structure to extract spectral features at different scales in parallel through multi-scale convolutional kernels.

[0020] Optionally, the time spectrum processing branch includes a variational mode decomposition preprocessing module and a bidirectional long short-term memory network. The variational mode decomposition decomposes the original time series into intrinsic mode functions of different frequencies, and the bidirectional long short-term memory network extracts the bidirectional time series features of each intrinsic mode function.

[0021] Optionally, the multimodal feature fusion module includes an explicit tensor interaction layer, which constructs a binary interaction plane and a ternary interaction core for energy spectrum features and time spectrum features, thereby realizing high-order feature interaction between modes.

[0022] Optionally, the training of the multimodal deep learning model adopts a joint loss function, which includes a combination of element yield prediction loss, capture cross section prediction loss, physical constraint loss and temporal spectrum reconstruction loss. Among them, the physical constraint loss includes element yield normalization constraint and capture cross section physical range constraint, and the temporal spectrum reconstruction loss calculates the similarity between the reconstructed spectrum and the original spectrum based on the dynamic time warping distance.

[0023] Optionally, it also includes the step of constructing a training dataset:

[0024] Using standard spectral libraries or Monte Carlo simulation methods, we construct mixed spectral data and their corresponding real component labels under different formation and wellbore parameter conditions to form a training dataset;

[0025] Data augmentation of the training dataset includes one or more combinations of adding random noise, simulating spectral drift, and time axis scaling.

[0026] Optionally, it also includes a step of optimizing the input features:

[0027] Collect multi-source logging data, including conventional logging curves and geological stratification information;

[0028] Construct an XGBoost model and calculate the SHAP value of each feature;

[0029] Important features are selected based on the cumulative contribution rate of SHAP values ​​and used as supplementary inputs to the multimodal deep learning model.

[0030] Optionally, it also includes a step of uncertainty quantification:

[0031] The model decoding output module simultaneously outputs the mean and variance of the predicted values;

[0032] The confidence intervals of the spectral resolution results are estimated using Monte Carlo Dropout or deep ensemble methods.

[0033] The reliability of the spectral resolution results is evaluated based on the confidence intervals.

[0034] Optionally, it also includes a transfer learning step:

[0035] The multimodal deep learning model is pre-trained on a large-scale simulated dataset in the source domain;

[0036] Freeze the bottom feature extraction layer and train only the high-level feature fusion layer and decoding layer;

[0037] The model is fine-tuned using small sample measured data in the target domain to achieve adaptive spectral resolution under small sample conditions.

[0038] A second aspect of the present invention provides a system for use in the above-described method, comprising:

[0039] The data acquisition module is used to acquire raw energy spectrum data and time spectrum data from downhole instruments;

[0040] The preprocessing module is used to correct and normalize the raw data;

[0041] The model storage module is used to store the multimodal deep learning model trained according to the above method;

[0042] The inference and computation module is used to call a multimodal deep learning model to automatically perform spectral analysis on the preprocessed data.

[0043] The saturation calculation module is used to calculate the formation oil saturation based on the spectral results.

[0044] The output module is used to display and store the spectral results and saturation curves.

[0045] Optionally, the inference computing module includes a GPU-accelerated computing unit that supports parallel processing of multiple depth points with a processing speed of no less than 500 points / second.

[0046] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method proposed in the first aspect of the present invention.

[0047] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. High-precision spectral resolution: By utilizing both local feature information of the energy spectrum and global decay information of the time spectrum through a multimodal deep learning network, complementary information is achieved, reducing the average relative error of element yield to 6.8% and the average relative error of Σ value to 5.2%, which is significantly better than traditional methods.

[0049] 2. Strong anti-interference capability: By adding various noises and spectral distortions to the training data, the model has strong robustness to statistical fluctuations in the measured data and environmental changes, and the spectral stability under low count rate conditions is improved by more than 3 times.

[0050] 3. High-efficiency processing: The model's spectral resolution process is a single forward propagation calculation, with a processing speed of up to 950 points / second, which is 190 times faster than traditional iterative methods, meeting the real-time requirements for continuous processing of massive depth points.

[0051] 4. Physical consistency guarantee: A physical constraint loss function is introduced to ensure that the spectral results conform to the physical laws of the formation, avoid unreasonable results such as negative element yields or Σ values ​​exceeding the physical range, and improve the reliability of the interpretation results.

[0052] 5. Uncertainty Quantification: Provides confidence intervals for spectral results, offering a risk assessment basis for subsequent geological decisions and avoiding over-reliance on unreliable results.

[0053] 6. Good generalization ability: Through training on large-scale simulated datasets and transfer learning strategies, the model can be applied to spectral interpretation tasks under different blocks, different lithologies, and different well conditions, and has a strong ability to be extended across wells and regions.

[0054] 7. High interpretability: Through SHAP value analysis and feature visualization, the basis for model decision-making can be traced, the contribution of each input feature to the spectral interpretation results can be explained, and the geologists' trust in the artificial intelligence spectral interpretation results can be enhanced. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating an automatic spectral interpretation method for post-jacket saturation logging based on multimodal deep learning fusion according to the present invention.

[0056] Figure 2 This is a structural diagram of an automatic spectrum interpretation system according to an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram illustrating the data enhancement effect according to an embodiment of the present invention;

[0058] Figure 4 This is a visualization of the feature extraction process of the energy spectrum processing branch according to an embodiment of the present invention;

[0059] Figure 5 This is a schematic diagram of BiLSTM feature extraction in a time spectrum processing branch according to an embodiment of the present invention;

[0060] Figure 6 This is a comparison diagram of the spectral interpretation results of the method of the present invention and the conventional method, as shown in an embodiment of the present invention;

[0061] Figure 7 This is a hardware system structure diagram of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] An automatic spectral interpretation method for post-jacket saturation logging based on multimodal deep learning fusion, such as... Figure 1 As shown, it includes the following steps:

[0064] (1) Obtain gamma spectrum data and time spectrum data of the stratum to be tested.

[0065] (2) Preprocess the energy spectrum data and time spectrum data to obtain standardized input features; the preprocessing includes one or more combinations of energy scale correction, time scale correction, background subtraction, normalization and smoothing noise reduction.

[0066] Furthermore, the training dataset is constructed, specifically including:

[0067] Using standard spectral libraries or Monte Carlo simulation methods, we construct mixed spectral data and their corresponding real component labels under different formation and wellbore parameter conditions to form a training dataset;

[0068] Data augmentation of the training dataset includes one or more combinations of adding random noise (such as Gaussian white noise to simulate statistical fluctuations), simulating spectral drift (such as energy scale drift), and time axis scaling (simulating different time window settings) to improve the model's generalization ability. The effects of data augmentation can be referenced. Figure 3 The figure shows a comparison of the original spectrum, the spectrum after adding noise, and the spectrum after time-axis scaling, which intuitively illustrates the role of enhancement methods in expanding data diversity.

[0069] (3) Input the standardized input features into the pre-trained multimodal deep learning model.

[0070] Multimodal deep learning models include:

[0071] The energy spectrum processing branch is used to extract the characteristic peak information and spectral shape features of the gamma energy spectrum data. The energy spectrum processing branch adopts a one-dimensional convolutional neural network combined with a residual connection structure, and extracts spectral features at different scales in parallel through multi-scale convolutional kernels. Figure 4 This section presents visualizations of the features learned by different convolutional layers in the energy spectrum processing branch. From... Figure 4 It can be seen that shallow networks mainly respond to the counting of single channels (original spectral morphology), while deep networks can combine features with physical meaning, such as identifying characteristic peak combinations of specific elements (such as complex peak regions of Si and Ca).

[0072] The time spectrum processing branch is used to extract the decay characteristics and temporal dependencies of the time spectrum data. The time spectrum processing branch includes a variational mode decomposition preprocessing module and a bidirectional long short-term memory network. The variational mode decomposition decomposes the original time series into intrinsic mode functions of different frequencies, and the bidirectional long short-term memory network extracts the bidirectional temporal features of each intrinsic mode function. Figure 5 This diagram illustrates how the BiLSTM network captures contextual information within the temporal spectrum. The lines of different colors represent the hidden states of the BiLSTM at different time steps. It is clear that the network has different response patterns to the decay trends in the early (high count rate) and late (low count rate) stages, thus effectively extracting key features such as the decay time constant.

[0073] The multimodal feature fusion module includes an explicit tensor interaction layer, which constructs a binary interaction plane and a ternary interaction core for energy spectrum features and time spectrum features; specifically, it includes: Specific implementation method: Let the energy spectrum eigenvector be... The time spectrum eigenvector is ; Binary interaction plane: calculated by outer product Get one Interaction matrix Each element of the matrix Indicates the energy spectrum of the first The first feature and time spectrum The pairwise interactions of each feature, due to The size is usually large, so a linear layer is subsequently used to reduce its dimensionality. dimensional vector ; The core of ternary interaction: To capture more complex intermodal relationships, a higher-order interaction tensor is constructed. A simplified implementation is to first concatenate two feature vectors to obtain... Then calculate its interaction with the outer product, for example, through The low-rank approximation can be used to achieve this, or a multi-head attention mechanism can be used to simulate high-order interactions between features. Finally, the results of binary and ternary interactions are fused to obtain a joint representation.

[0074] The decoding output module is used to map the fused features to the target spectral parameters.

[0075] The training of the multimodal deep learning model adopts a joint loss function, which includes a combination of element yield prediction loss, capture cross section prediction loss, physical constraint loss and temporal spectrum reconstruction loss. The physical constraint loss includes element yield normalization constraint and capture cross section physical range constraint, and the temporal spectrum reconstruction loss calculates the similarity between the reconstructed spectrum and the original spectrum based on the dynamic time warping distance.

[0076] (4) The multimodal deep learning model outputs the spectrum results, including the relative yield of formation elements and the macroscopic capture section Σ.

[0077] Furthermore, it also includes the step of uncertainty quantification:

[0078] The model decoding output module simultaneously outputs the mean and variance of the predicted values;

[0079] The confidence intervals of the spectral resolution results are estimated using Monte Carlo Dropout or deep ensemble methods.

[0080] The reliability of the spectral resolution results is evaluated based on the confidence intervals.

[0081] For example, if the confidence interval for the predicted Σ value at a certain depth point is too wide, it indicates that the result for that point is greatly affected by noise and has low reliability. The mathematical basis for this is the assumption about the prediction target. Obey the model output For the mean, For a Gaussian distribution with standard deviation, the negative log-likelihood loss for model optimization is:

[0082] ;

[0083] in, The input features are the preprocessed gamma-ray spectral data and time-spectral data. The target for prediction could be the relative yield of a certain element in a formation, or the true value of the macroscopic capture section Σ. The mean prediction of the model output corresponds to the expected value of the target variable; The standard deviation prediction of the model output represents the model's estimate of the uncertainty of the prediction result. (Usually, positive numbers are guaranteed through exponential activation functions or Softplus). To give input Below, target value The conditional probability density; The negative log-likelihood loss function is used to minimize this loss, enabling the model to output both accurate predictions and reasonable confidence intervals. This is a constant term independent of the model parameters and can be ignored during optimization. It is the first term in the loss function. The penalty for excessively large uncertainty estimates encourages the model to produce smaller variances where possible; the second term It is the weighted mean square error, which is used when the uncertainty of the model prediction... When the value is large, the weight of this error term decreases, allowing the model to assign higher uncertainty to samples with higher noise, thereby avoiding overfitting.

[0084] (5) The formation oil saturation So is calculated based on the spectral results.

[0085] This is typically based on a volumetric model and a specific saturation formula. For the macroscopic capture section Σ, the following formula is used:

[0086] ;

[0087] Among them, water saturation Calculated using the following formula:

[0088] ;

[0089] Solve :

[0090] ;

[0091] in, This is the macroscopic capture section of the formation output by the model. The content of clay (which can be obtained from conventional logging curves) is the mud content. Porosity (which can be obtained from other well logging data) , , , These are macroscopic capture sections for the rock skeleton, mudstone, formation water, and oil, respectively, and are usually known constants or can be determined by other methods.

[0092] In addition, saturation can also be calculated using elemental yields. For example, the ratio method based on the carbon-to-oxygen ratio (C / O) can be used by comparing the measured C / O ratio with that of a pure oil layer. and pure water layer Linear interpolation is used to obtain the oil saturation:

[0093] ;

[0094] in, This represents the ratio of carbon to oxygen production output from the model.

[0095] Example 1: Joint Energy-Time Spectrum Decomposition Based on CNN-BiLSTM

[0096] Taking pulsed neutron logging data from a certain oilfield XX block as an example, the block has complex lithology, mainly consisting of interbedded sandstone, limestone and dolomite, with a large range of formation water salinity (5000~150000ppm).

[0097] First, core analysis data from three cored wells in the block were collected, including elemental content, saturation, and porosity, as true references for model validation. A refined formation-wellbore coupling model was established using the MCNP Monte Carlo simulation program. Parameters such as different lithological combinations, porosity range of 2%–35%, water saturation range of 0%–100%, clay content of 0%–40%, formation water salinity, wellbore size, casing thickness, cement sheath thickness, and well fluid properties were set, generating approximately 500,000 sets of simulation data. Each set includes corresponding gamma-ray spectra (512 channels), time spectra (256 channels), elemental yields (8 elements including C, O, Si, Ca, Fe, Cl, H, and S), and labels for the macroscopic trapping cross section Σ.

[0098] Data preprocessing includes: energy / time scale correction, background subtraction, normalization, and Savitzky-Golay smoothing and noise reduction. Data augmentation includes: adding Gaussian white noise (signal-to-noise ratio 20dB~40dB), random energy drift (±0.5 channels), random time axis scaling (0.95~1.05), and random combination of spectral fragments under different simulation conditions.

[0099] Build as Figure 2 The multimodal deep learning model shown:

[0100] Energy spectrum processing branch: Input a 512-dimensional energy spectrum, process it through Conv1D(64,k=7)+BN+ReLU+MaxPool, Conv1D(128,k=5)+BN+ReLU+MaxPool, Conv1D(256,k=3)+BN+ReLU, 3 residual blocks, global average pooling, and output a 256-dimensional feature vector;

[0101] Temporal spectrum processing branch: The input 256-dimensional temporal spectrum is decomposed into 5 intrinsic mode functions by VMD, which are then input into BiLSTM (128 units, return sequences) and BiLSTM (64 units, return sequences) respectively. Finally, a 128-dimensional context vector is obtained by weighted summation through a self-attention layer.

[0102] Multimodal feature fusion module: The energy spectrum features (256 dimensions) and time spectrum features (128 dimensions) are concatenated into 384 dimensions, and a binary interaction plane (256×128 outer product followed by dimensionality reduction) and a ternary interaction core (384×384 interaction matrix low-rank decomposition) are constructed. After fusion, a 512-dimensional joint representation is obtained.

[0103] Decoding output module: The three branches output the element yield (fully connected + Softmax, 8-dimensional), Σ value (fully connected + Linear, 1-dimensional), and uncertainty (fully connected 2-dimensional, outputting mean and variance).

[0104] The model training uses a joint loss function: .

[0105] The training of the multimodal deep learning model employs a joint loss function. The joint loss function includes a combination of various loss functions, such as element yield prediction loss, capture section prediction loss, physical constraint loss, and time spectrum reconstruction loss. The element yield prediction loss is typically measured using mean squared error (MSE) or mean absolute error (MAE) to measure the difference between the predicted element yield vector and the true label. It is the capture section prediction loss, which also uses MSE or MAE to measure the difference between the predicted Σ value and the true value. The loss is a physical constraint, which includes element yield normalization constraint (forcing the sum of all predicted element yields to be close to 1, using L1 or L2 penalty terms) and trap section physical range constraint (if the predicted Σ value exceeds the reasonable range defined by prior information such as stratigraphic lithology, porosity, and mineralization). Then a punishment will be imposed, for example ). This is the temporal spectrum reconstruction loss. The model should not only predict parameters but also reconstruct the input temporal spectrum based on hidden layer features. This reconstruction loss is based on the dynamic temporal warping distance. Computational reconstruction spectrum Compared with the original spectrum Similarity, i.e. The DTW distance is a better measure of the similarity in shape between two time series, and it can provide meaningful gradients even when there are slight deformations in the time axis.

[0106] The Adam optimizer was used with an initial learning rate of 0.001, a batch size of 128, 100 training epochs, and an early stopping mechanism.

[0107] Evaluation on the test set showed that the method of this invention had an average relative error of 6.8% for element yield and 5.2% for Σ value, with a processing speed of 950 points / second. Its noise resistance was significantly better than traditional methods and single-modal deep learning methods. When applied to five production wells in Block XX and compared with core analysis results, the average absolute error of water saturation was 4.2%, which is better than the 8.7% of the traditional method.

[0108] Figure 6 This paper presents a comparison of the spectral resolution results of the proposed method and the traditional least squares stripping spectroscopy (LSF) at a certain depth point. The figures clearly show that for the elements Si and Ca, whose characteristic peaks overlap significantly, the prediction results of the proposed method are closer to the reference values ​​from core analysis, while the traditional method exhibits a large deviation. This intuitively demonstrates the superiority of multimodal deep learning fusion methods in handling complex spectral information and high-order interactive features.

[0109] Example 2: Input Feature Optimization Using XGBoost-SHAP

[0110] Building upon Example 1, an input feature optimization module was added. All available logging data from the target well (including conventional logging curves such as GR, DEN, CNL, RT, and AC) were collected, and derived features (peak area, peak-to-column ratio, early / late decay rate, etc.) were extracted from the energy spectrum and time spectrum. An XGBoost regression model was constructed, and the SHAP value of each feature was calculated. Important features with a cumulative contribution rate >95% were selected (ultimately reducing the dimensions from 896 to 128). The multimodal model input structure was redesigned, reducing training time by 65% ​​and improving prediction accuracy by approximately 3%.

[0111] Example 3: Application of transfer learning under small sample conditions

[0112] In the new exploration area, only eight depth points with core markings were available. Using the multimodal deep learning model pre-trained in the XX block as described in Example 1, the bottom-level feature extraction layer was frozen, and only the high-level feature fusion and decoding layers were trained. Fine-tuning was performed for 20-30 rounds using small sample data from the target domain, and adaptive batch normalization was introduced. The final test set error was reduced to 9.3%, meeting the requirements for engineering applications.

[0113] The system used in the method described in the above embodiments, such as Figure 7 As shown, it includes:

[0114] The data acquisition module is used to acquire raw energy spectrum data and time spectrum data from downhole instruments;

[0115] The preprocessing module is used to correct and normalize the raw data;

[0116] The model storage module is used to store the multimodal deep learning model trained according to the above method;

[0117] The inference and computation module is used to call a multimodal deep learning model to automatically perform spectral analysis on the preprocessed data.

[0118] The saturation calculation module is used to calculate the formation oil saturation based on the spectral results.

[0119] The output module is used to display and store the spectral results and saturation curves.

[0120] System workflow: Data acquired by downhole instruments is transmitted to the surface in real time. After preprocessing, a multimodal deep learning model is loaded by the GPU server to automatically interpret the spectrum. Based on the spectrum interpretation results, the formation oil saturation is calculated, and the spectrum interpretation results and saturation curves are displayed in real time. All data is stored in the database.

[0121] In one embodiment, a computer-readable storage medium is provided on which a computer program is stored, which, when executed by a processor, implements the steps of the method described above.

[0122] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0123] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A post-casing saturation logging automatic spectral analysis method based on multi-modal deep learning fusion, characterized in that, Includes the following steps: Acquire gamma-ray spectral data and time-spectral data of the formation to be tested; The energy spectrum data and time spectrum data are preprocessed to obtain standardized input features; Standardized input features are fed into a pre-trained multimodal deep learning model; The multimodal deep learning model outputs spectral results, including relative yields of formation elements and macroscopic capture sections Σ; The formation oil saturation was calculated based on the spectral results.

2. The method of claim 1, wherein the method is characterized by, The step of inputting standardized input features into a pre-trained multimodal deep learning model includes: The energy spectrum processing branch is used to extract characteristic peak information and spectral shape features from gamma energy spectrum data; The time spectrum processing branch is used to extract the decay features and temporal dependencies of time spectrum data; The multimodal feature fusion module is used for high-order interactive fusion of energy spectrum features and time spectrum features; The decoding output module is used to map the fused features to the target spectral parameters.

3. The method of claim 2, wherein the method is characterized by, The energy spectrum processing branch uses a one-dimensional convolutional neural network combined with a residual connection structure to extract spectral features at different scales in parallel through multi-scale convolutional kernels. The time spectrum processing branch includes a variational mode decomposition preprocessing module and a bidirectional long short-term memory network. The variational mode decomposition decomposes the original time series into intrinsic mode functions of different frequencies, and the bidirectional long short-term memory network extracts the bidirectional time series features of each intrinsic mode function. The multimodal feature fusion module includes an explicit tensor interaction layer, which constructs a binary interaction plane and a ternary interaction core for energy spectrum features and time spectrum features, enabling high-order feature interactions between modes.

4. The method of claim 1, wherein the method is characterized by, It also includes the step of building the training dataset: Using standard spectral libraries or Monte Carlo simulation methods, we construct mixed spectral data and their corresponding real component labels under different formation and wellbore parameter conditions to form a training dataset; Data augmentation of the training dataset includes one or more combinations of adding random noise, simulating spectral drift, and time axis scaling.

5. The method of claim 1, wherein the method is characterized by, It also includes the step of input feature optimization: Collect multi-source logging data, including conventional logging curves and geological stratification information; Construct an XGBoost model and calculate the SHAP value of each feature; Important features are selected based on the cumulative contribution rate of SHAP values ​​and used as supplementary inputs to the multimodal deep learning model.

6. The automatic spectral interpretation method for post-jacket saturation logging based on multimodal deep learning fusion as described in claim 1, characterized in that, It also includes the step of uncertainty quantification: The model decoding output module simultaneously outputs the mean and variance of the predicted values; The confidence intervals of the spectral resolution results are estimated using Monte Carlo Dropout or deep ensemble methods. The reliability of the spectral resolution results is evaluated based on the confidence intervals.

7. The method of claim 1, wherein the method is based on a multi-modal deep learning fusion-based automatic post-cementation-saturation log interpretation method, characterized in that, It also includes the steps of transfer learning: The multimodal deep learning model is pre-trained on a large-scale simulated dataset in the source domain; Freeze the bottom feature extraction layer and train only the high-level feature fusion layer and decoding layer; The model is fine-tuned using small sample measured data in the target domain to achieve adaptive spectral resolution under small sample conditions.

8. A system for use in the method of any one of claims 1 to 7, characterized in that include: The data acquisition module is used to acquire raw energy spectrum data and time spectrum data from downhole instruments; The preprocessing module is used to correct and normalize the raw data; The model storage module is used to store the multimodal deep learning model trained by the method according to any one of claims 1-7; The inference and computation module is used to call a multimodal deep learning model to automatically perform spectral analysis on the preprocessed data. The saturation calculation module is used to calculate the formation oil saturation based on the spectral results. The output module is used to display and store the spectral results and saturation curves.

9. The system of claim 8, wherein, The inference computing module includes a GPU-accelerated computing unit that supports parallel processing of multiple depth points with a processing speed of no less than 500 points / second.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images