A nuclear magnetic logging data inversion system and method based on semi-supervised adversarial learning

Abstract

This application relates to the field of geophysical exploration and presents a nuclear magnetic resonance logging data inversion system and method based on semi-supervised adversarial learning. It includes: T The 2-spectrum reconstructor, trained using a labeled logging dataset, obtains the spin echo signal and preliminary... T The nonlinear mapping relationship of the two spectra is used to process the unlabeled well logging dataset and generate predictions. T 2. Spectrum; The pseudo-tag selector selects the actual spin echo signal and the predicted... T The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset are used to generate predictions. T 2. Spectrum confidence levels, and select predictions based on confidence levels. T 2. The spectrum is used as a pseudo-label, and the pseudo-label is combined with the prediction. T The two spectra were mixed and used with labeled and unlabeled logging datasets as... T 2. Training set for the spectral reconstructor; T The 2-spectrum reconstructor, after being trained on the training set, obtains the actual spin echo signal and the predicted signal. T The nonlinear mapping relationship of the two spectra. This application realizes high-precision and efficient intelligent nuclear magnetic resonance logging data under finite label conditions. T 2. Spectral inversion.

Description

Technical Field

[0001] This application relates to the field of geophysical exploration, specifically to a nuclear magnetic resonance logging data inversion system and method based on semi-supervised adversarial learning. Background Technology

[0002] Nuclear magnetic resonance (NMR) logging technology excites hydrogen protons in the formation to induce Larmor precession by applying an alternating current through a CPMG pulse sequence, and then acquires the spin echo signal released during the relaxation process. The spin echo signal is then inverted to obtain... Spectrum, for Analyzing the peak morphology and component area of ​​the spectral data yields petrophysical parameters such as formation porosity and fluid saturation, providing a basis for oil and gas reservoir evaluation. However, the traditional multi-exponential relaxation inversion process is ill-conditioned, resulting in multiple solutions and reducing the reliability of the inversion interpretation results. Furthermore, the regularization parameter needs to be selected through laboratory calculations, making the inversion process complex. Therefore, this paper aims to develop a high-precision and efficient intelligent NMR logging data inversion method to provide high-quality data for subsequent oil and gas reservoir evaluation. Spectral data is of great significance.

[0003] Chinese patent publication number CN114994776A discloses "a method, apparatus and medium for determining rock physical parameters", which constructs a structure containing residuals. Norms and Regularization The dual-sparse constraint inversion objective function based on norms iteratively solves for inversion solutions corresponding to different regularization parameters, and then selects the optimal parameters to obtain the optimal inversion solution. Compared with traditional inversion methods, this method achieves better sparsity and higher spectral resolution in the inversion solution. However, the inversion process in this method is ill-conditioned, and the solution is not unique, leading to errors in the inversion results; furthermore, the selection of the optimal regularization parameter relies on complex laboratory calculations, resulting in low inversion efficiency. Summary of the Invention

[0004] This application provides a nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning. It addresses the problems of ill-conditioned solutions lacking uniqueness and the need for additional laboratory calculations to select regularization parameters.

[0005] This application also provides a method for inverting nuclear magnetic resonance logging data based on semi-supervised adversarial learning.

[0006] A nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning, according to an embodiment of the first aspect of this application, includes:

[0007] The data processing module is used to match the spin echo signal generated by forward modeling with the corresponding spin echo signal. The spectrum consists of labeled logging datasets and unlabeled logging datasets composed of actual spin echo signals.

[0008] The spectral reconstructor, trained using a labeled logging dataset, obtains the spin echo signal and preliminary... The nonlinear mapping relationship of the spectrum is used to process the unlabeled logging dataset and generate predictions. Spectrum;

[0009] A pseudo-tag selector is used to compare the actual spin echo signal with the predicted one. The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset generate predictions. The confidence level of the spectrum is used to select predictions. The spectrum serves as a pseudo-label, linking the pseudo-label with the prediction. After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor;

[0010] The After training on the training set, the spectral reconstructor obtains the actual spin echo signal and the predicted signal. Nonlinear mapping relationship of the spectrum.

[0011] Furthermore, the aforementioned The spectral reconstructor comprises a feature learning module and a detail enhancement module. Skip connections are used to directly pass features extracted by the feature learning module at different scales to the detail enhancement module. The feature learning module extracts features from the labeled well logging dataset. The initial feature information of the spectrum is obtained by feature fusion and input to the detail enhancement module to recover the feature detail information. Spectrum, establishing spin echo signal and preliminary Nonlinear mapping relationship of the spectrum.

[0012] Furthermore, the feature learning module includes sequentially configured... Convolutional layers, ReLU activation layers, and multiple feature fusion blocks, wherein the feature fusion blocks contain sequentially set... Convolutional layer, batch normalization layer, ReLU activation layer, and parallel first and second branches, the first branch containing sequentially configured... Convolutional layer, dilated convolutional layer, batch normalization layer, and ReLU activation layer; the second branch contains layers set sequentially. The convolutional layer, batch normalization layer, and ReLU activation layer are combined and the two branches are merged as the output.

[0013] Furthermore, The inversion loss function of the spectral reconstructor includes cross-entropy loss, pseudo-matching loss, and pseudo-label loss;

[0014] The cross-entropy loss is used for training with a labeled well logging dataset. When evaluating the spin echo signal via the spectral reconstructor, Preliminary results obtained after spectral reconstruction Spectrum and The degree of similarity of the spectra;

[0015] The pseudo-matching loss is used to train the dataset simultaneously using both labeled and unlabeled logging datasets. When evaluating the spectral reconstructor, the prediction is evaluated. Spectrum and Spectral differences;

[0016] The pseudo-label loss is used for training with an unlabeled logging dataset. When evaluating the prediction of the actual spin echo signal in the spectral reconstructor... The degree of similarity between the spectrum and the pseudo-label.

[0017] Furthermore, the pseudo-label selector determines the corresponding label logging dataset by... Predictions corresponding to spectral and pseudo-matching datasets The confidence level is obtained from the similarity of the spectra.

[0018] A method for inverting nuclear magnetic resonance logging data based on semi-supervised adversarial learning, according to a second aspect of this application, includes:

[0019] The spin echo signal generated by forward modeling is compared with the corresponding spin echo signal. The spectrum consists of labeled logging datasets and unlabeled logging datasets composed of actual spin echo signals.

[0020] Training was performed using a tagged logging dataset to obtain spin echo signals and preliminary data. The nonlinear mapping relationship of the spectrum is used to process the unlabeled logging dataset and generate predictions. Spectrum;

[0021] Based on actual spin echo signals and predictions The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset generate predictions. The confidence level of the spectrum is used to select predictions. Spectrum as a pseudo-label;

[0022] Combining pseudo-labels with prediction After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor;

[0023] Training with training set The actual spin echo signal obtained after the spectral reconstructor is compared with the predicted signal. Nonlinear mapping relationships of the spectrum;

[0024] Based on actual spin echo signals and predictions The nonlinear mapping relationship of the spectrum is used to process the real-time acquired spin echo signal to obtain the prediction. Spectrum.

[0025] Furthermore, the inversion loss function is iteratively optimized by minimizing iterative optimization. The prediction performance of the spectral reconstructor, wherein the inversion loss function includes: cross-entropy loss, pseudo-matching loss, and pseudo-label loss;

[0026] The cross-entropy loss is used for training with a labeled well logging dataset. When evaluating the spin echo signal via the spectral reconstructor, Preliminary results obtained after spectral reconstruction Spectrum and The degree of similarity of the spectra;

[0027] The pseudo-matching loss is used to train the dataset simultaneously using both labeled and unlabeled logging datasets. When evaluating the spectral reconstructor, the prediction is evaluated. Spectrum and Spectral differences;

[0028] The pseudo-label loss is used for training with an unlabeled logging dataset. When evaluating the prediction of the actual spin echo signal in the spectral reconstructor... The degree of similarity between the spectrum and the pseudo-label.

[0029] Furthermore, the aforementioned The spectral reconstructor comprises a feature learning module and a detail enhancement module. Skip connections are used to directly pass features extracted by the feature learning module at different scales to the detail enhancement module. The feature learning module extracts features from the labeled well logging dataset. The initial feature information of the spectrum is obtained by feature fusion and input to the detail enhancement module to recover the feature detail information. Spectrum, establishing spin echo signal and preliminary Nonlinear mapping relationship of the spectrum.

[0030] Furthermore, the feature learning module includes sequentially configured... Convolutional layers, ReLU activation layers, and multiple feature fusion blocks, wherein the feature fusion blocks contain sequentially set... Convolutional layer, batch normalization layer, ReLU activation layer, and parallel first and second branches, the first branch containing sequentially configured... Convolutional layer, dilated convolutional layer, batch normalization layer, and ReLU activation layer; the second branch contains layers set sequentially. The convolutional layer, batch normalization layer, and ReLU activation layer are combined and the two branches are merged as the output.

[0031] Furthermore, a pseudo-label selector is used to determine the corresponding labeled logging dataset. Predictions corresponding to spectral and pseudo-matching datasets The confidence level is obtained from the similarity of the spectra.

[0032] This application has the following advantages and beneficial effects:

[0033] This application addresses the problems of traditional methods lacking unique solutions due to ill-conditioned approaches and requiring additional laboratory calculations for regularization parameters. It establishes a method for predicting spin echo signals based on an adversarial learning strategy. The nonlinear mapping relationship of the spectrum yields reliable inversion results, and the operation process is intelligent and efficient, enabling high-precision and high-efficiency intelligent nuclear magnetic resonance logging data under limited label conditions. Spectral inversion: To address the issue of supervised inversion methods being highly dependent on data scale and labeling, a semi-supervised learning approach is adopted to cleverly combine the advantages of supervised learning's precise mapping and unsupervised learning's weak data dependence, thereby reducing the data dependence of the inversion method and improving inversion accuracy. Attached Figure Description

[0034] Figure 1 A block diagram of a nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning provided in an embodiment of this application;

[0035] Figure 2 A block diagram of a feature fusion block in a nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning, provided in an embodiment of this application;

[0036] Figure 3 A flowchart illustrating the construction method of a nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning provided in this application embodiment;

[0037] In the diagram, A represents Convolutional layer, B represents ReLU activation layer, C represents feature fusion block, D represents transposed convolutional layer, E represents Leaky ReLU activation layer, F represents Dropout layer, G represents fully connected layer, H represents Sigmoid function layer, I represents batch normalization layer, J represents Convolutional layer, K indicates a dilated convolutional layer, and the labels here only represent layers with the same function. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0039] See Figure 1 This embodiment presents a nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning, comprising:

[0040] The data processing module is used to match the spin echo signal generated by forward modeling with the corresponding spin echo signal. The spectrum consists of labeled logging datasets and unlabeled logging datasets composed of actual spin echo signals.

[0041] The spectral reconstructor, trained using a labeled logging dataset, obtains the spin echo signal and preliminary... The nonlinear mapping relationship of the spectrum is used to process the unlabeled logging dataset and generate predictions. Spectrum;

[0042] A pseudo-tag selector is used to compare the actual spin echo signal with the predicted one. The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset generate predictions. The confidence level of the spectrum is used to select predictions. The spectrum serves as a pseudo-label, linking the pseudo-label with the prediction. After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor;

[0043] After training on the training set, the spectral reconstructor obtains the actual spin echo signal and the predicted signal. Nonlinear mapping relationship of the spectrum.

[0044] Among them, through forward modeling Group spin echo signals, represented as , And the depth of group M corresponding to the spin echo signal Spectrum, represented as , ,form Group data pairs This constitutes a labeled well logging dataset. The actual spin echo signal acquired by the nuclear magnetic resonance spectrometer is represented by a scaling transformation as follows: and expanded to Group, , To form an unlabeled well logging dataset Preprocessing was performed on both labeled and unlabeled well logging datasets, including normalization, exponential thinning, and PCA dimensionality reduction. Here, labeled well logging datasets refer to datasets where the spin echo signal has a corresponding... The spectrum, while unlabeled logging datasets only contain actual spin echo signals, without corresponding... Spectrum.

[0045] The spectral reconstructor and the pseudo-label selector form a semi-supervised adversarial system. Spectral inversion network express, Semi-supervised adversarial Network parameters of spectral inversion network: semi-supervised adversarial Spectral inversion network includes Spectrum reconstructor, using express, for The network parameters of the spectral reconstructor, and the pseudo-label selector, are used... express, These are the network parameters for the pseudo-label selector;

[0046] After training, the spectral reconstructor finally obtains the actual spin echo signal and the predicted signal. The nonlinear mapping relationship of the spectrum is used to process the actual acquired spin echo signal, and the prediction can be directly obtained. Spectrum.

[0047] The spectral reconstructor is trained in a supervised manner using labeled well logging datasets, learning to convert spin echo signals into corresponding... The nonlinear mapping relationship of the spectrum is expressed as: Initialize network parameters; then, using unsupervised training, calculate the corresponding values ​​based on the spin echo signal. Nonlinear mapping relationship of spectrum Generate predictions corresponding to the unlabeled logging dataset. Spectrum representation Compare the actual spin echo signal with the predicted one. Spectral matching, each match Group data, resulting in a pseudo-matching dataset, represented as ;

[0048] The labeled logging dataset and the pseudo-matching dataset are input into the pseudo-label selector, which then... Adversarial training of the spectral reconstructor selects predictions with confidence scores higher than a threshold. The spectrum can be used as a pseudo-label, and the threshold can be set to 95%; the pseudo-label can be combined with... After spectral mixing, as The training set of the spectral reconstructor is expanded by increasing the size of the training samples and iteratively optimized by minimizing the inversion loss function. Spectrum Reconstructor Spectral prediction results; finally, the actual spin echo signal and the predicted signal are obtained. The nonlinear mapping relationship of the spectrum is expressed as: .

[0049] Data collected using nuclear magnetic resonance imaging (NMR) Group of actual spin echo signals , To form a test set Testing semi-supervised adversarial Spectral inversion network input after training semi-supervised adversarial Spectral inversion network to obtain predictions The spectral inversion results are used to evaluate the accuracy of the inversion results and the performance of the inversion network.

[0050] In one embodiment, The spectral reconstructor comprises a feature learning module and a detail enhancement module. Skip connections are used to directly pass features extracted by the feature learning module at different scales to the detail enhancement module. The feature learning module extracts features from the labeled well logging dataset. The initial feature information of the spectrum is processed by feature fusion and then input into the detail enhancement module to recover the feature detail information, thus obtaining the preliminary result. Spectrum, establishing spin echo signal and preliminary Nonlinear mapping relationship of the spectrum.

[0051] In one embodiment, the feature learning module includes modules arranged sequentially. Convolutional layer A, ReLU activation layer B, and multiple feature fusion blocks C, where, see [link to relevant documentation] Figure 2 As shown, the feature fusion block C contains sequentially arranged... Convolutional layer A, batch normalization layer I, ReLU activation layer B, and parallel first and second branches, the first branch containing sequentially configured... Convolutional layer A, dilated convolutional layer K, batch normalization layer I, and ReLU activation layer B, the second branch contains layers set sequentially. The convolutional layer J, the batch normalization layer I, and the ReLU activation layer B are fused together to form the output.

[0052] The detail enhancement module consists of multiple transposed convolutional layers D and multiple ReLU activation layers B. Each transposed convolutional layer D is followed by a ReLU activation layer B. Skip connections are used to directly pass the features extracted by the feature learning module at different scales to the detail enhancement module, thereby achieving multi-scale feature fusion.

[0053] In one embodiment, the pseudo-tag selector includes a series of combination units, each combination unit comprising, in sequence, a combination of, a combination of, and, a combination of ... A convolutional layer A, a LeakyReLU activation layer E, and a Dropout layer F are chained together, and the outputs of these combined units are then connected to a fully connected layer G and a sigmoid function layer H. A pseudo-label selector determines the labeling of the well logging dataset. Predictions corresponding to spectral and pseudo-matching datasets The confidence level is obtained from the similarity of the spectra.

[0054] In one embodiment, see Figure 3 As shown, the construction method of the nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning includes:

[0055] Construct labeled and unlabeled well logging datasets;

[0056] Build including Semi-supervised adversarial approach for spectral reconstructors and pseudo-label selectors Spectral inversion network;

[0057] Both labeled and unlabeled logging datasets will be input. The spectral reconstructor uses labeled well logging datasets to... The spectral reconstructor undergoes supervised training to obtain the spin echo signal and the initial... Nonlinear mapping relationships of the spectrum;

[0058] Using unlabeled logging datasets to... The spectral reconstructor undergoes unsupervised training.

[0059] Based on spin echo signals and preliminary The nonlinear mapping relationship of the spectrum is predicted Spectrum;

[0060] Actual spin echo signal and prediction The pseudo-matching dataset obtained from spectral matching, along with the labeled well logging dataset, is input into the pseudo-label selector to calculate predictions. Confidence level of the spectrum;

[0061] Determine if the confidence level is higher than a threshold; if it is lower than the threshold, return to the initial method based on the spin echo signal. The nonlinear mapping relationship of the spectrum is predicted The steps of spectrogramming;

[0062] When it exceeds the threshold, a prediction will be made. The spectrum serves as a pseudo-label, linking the pseudo-label with the prediction. After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor;

[0063] After training on the training set, the spectral reconstructor obtains the actual spin echo signal and the predicted signal. Nonlinear mapping relationship of the spectrum.

[0064] On the other hand, embodiments of this application provide a semi-supervised adversarial learning-based nuclear magnetic resonance (NMR) logging data inversion method using the aforementioned semi-supervised adversarial learning-based NMR logging data inversion system, comprising:

[0065] The spin echo signal generated by forward modeling is compared with the corresponding spin echo signal. The spectrum consists of labeled logging datasets and unlabeled logging datasets composed of actual spin echo signals.

[0066] Training was performed using a tagged logging dataset to obtain spin echo signals and preliminary data. The nonlinear mapping relationship of the spectrum, based on the spin echo signal and the preliminary Nonlinear mapping of spectra to generate predictions from unlabeled well logging datasets Spectrum;

[0067] Based on actual spin echo signals and predictions The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset generate predictions. The confidence level of the spectrum is used to select predictions. The spectrum serves as a pseudo-label, linking the pseudo-label with the prediction. After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor;

[0068] Training with training set The actual spin echo signal obtained after the spectral reconstructor is compared with the predicted signal. Nonlinear mapping relationships of the spectrum;

[0069] Based on actual spin echo signals and predictions The nonlinear mapping relationship of the spectrum is used to process the real-time acquired spin echo signal to obtain the prediction. Spectrum.

[0070] In this process, both labeled and unlabeled logging datasets will be input into the semi-supervised adversarial algorithm. Spectral inversion network;

[0071] The spectral reconstructor utilizes a feature learning module, first through... Convolutional layer A and ReLU activation layer B extract the data from the labeled well logging dataset. Initial feature information of the spectrum :

[0072] ,

[0073] In the formula, It is the convolution kernel of the convolutional layer. '*' represents the bias of the convolutional layer, and '*' represents the convolution operation. It is a ReLU activation layer operation. for Spectrum;

[0074] Initial feature information Input feature fusion block C, after The number of channels is transformed by convolutional layer A, and then extracted by dilated convolutional layer K in the first branch. The overall decay trend of the spectrum, after the second branch Convolutional layer J narrow field of view extraction Information on short relaxation components of the spectrum; fusing features extracted from the two branches to obtain fused features, while simultaneously grasping... The global structure and local details of the spectrum are enhanced. Accuracy of spectral feature extraction;

[0075] The feature fusion block C operation is as follows:

[0076] ,

[0077] ,

[0078] In the formula, , , These are the output, kernel, and bias of the first convolutional layer of the feature fusion block. , , , These are the convolution kernels and biases of the two convolutional layers in the first branch of the feature fusion block. , These are the convolution kernel and bias of the convolutional layer in the second branch of the feature fusion block, respectively. It is a feature of fusion. It is a batch normalization operation;

[0079] Fusion characteristics After the layer feature fusion block is processed, it is input into the detail supplementation module. Recovering detailed feature information from transposed convolutional layers:

[0080] ,

[0081] In the formula, This represents the number of transposed convolutional layers. and The first The output and input of the transposed convolutional layer. and The first The kernel and bias of the transposed convolutional layer; when When it is 1, for ;when When supplementing the last layer of the module with details, Recorded as preliminary Spectrum;

[0082] Labeled well logging datasets were respectively processed The feature learning module and detail enhancement module in the spectral reconstructor handle initialization. The parameters of the spectral reconstructor network are Establishing a preliminary understanding from spin echo signals. Nonlinear mapping relationship of spectrum :

[0083] ,

[0084] Unlabeled logging datasets based on nonlinear mapping relationships Generate predictions Musical score:

[0085] ,

[0086] Compare the actual spin echo signal with the prediction Spectral matching yields pseudo-matching data pairs, each... A pseudo-matching dataset is composed of several pseudo-matching data pairs. ; This represents a pseudo-matching dataset.

[0087] Both the labeled logging dataset and the pseudo-matching dataset are input into the pseudo-label selector; the pseudo-matching dataset is then processed... The confidence score is obtained by processing the layer combination unit, one fully connected layer, and one Sigmoid function layer:

[0088] ,

[0089] In the formula, For Sigmoid function layer operations, For fully connected layer operations, For LeakyReLU activation layer operations, and Here, represents the convolution kernel and bias of the convolutional layer, and V is the mask matrix of the Dropout layer. Let V be the probability that an element in the mask matrix V is 0. Indicates confidence level;

[0090] A pre-set confidence threshold is used to select predictions with a confidence level higher than the threshold. Spectrum as a pseudo-label, and Spectral mixing, expanding semi-supervised adversarial The learning sample size of the spectral inversion network is iteratively optimized to obtain the optimal semi-supervised adversarial network. Spectral inversion network;

[0091] Iterative optimization by minimizing the inversion loss function The prediction performance of the spectral reconstructor, wherein the inversion loss function includes: cross-entropy loss, pseudo-matching loss, and pseudo-label loss;

[0092] Cross-entropy loss was used for training on a labeled well logging dataset. When evaluating the spin echo signal via the spectral reconstructor, Preliminary results obtained after spectral reconstruction Spectrum and Spectral similarity:

[0093] ,

[0094] For the initial Spectrum for Spectrum This represents the cross-entropy loss.

[0095] The pseudo-matching loss is used to train on both labeled and unlabeled logging datasets simultaneously. When evaluating the spectral reconstructor, the prediction is evaluated. Spectrum and Spectral differences

[0096] Pseudo-matching loss while training on labeled well logging datasets and unlabeled logging datasets Assessment and prediction spectrum and spectrum Differences:

[0097] ,

[0098] In the formula, Let F be the norm, and E be the expected value. and These are labeled well logging datasets with centralized data pairs. And the actual spin echo signal in the unlabeled logging data The probability distribution, For pseudo-matching loss, To process data pairs pseudo-tag selector, To process data pairs pseudo-tag selector, Indicates prediction Spectrum This is the actual spin echo signal;

[0099] The pseudo-label loss is used for training with an unlabeled logging dataset. When evaluating the prediction of the actual spin echo signal in the spectral reconstructor... Similarity between the spectrum and the pseudo-label:

[0100] ,

[0101] ,

[0102] In the formula, For the threshold, For the new round of training Spectrum reconstructor network parameters For pseudo-label loss, It is a pseudo-tag. For prediction The score is pseudo-tag selector, Indicates the actual spin echo signal and a new round of training Spectrum Reconstructor Network Parameters of Spectrum reconstructor;

[0103] The inversion loss function is defined as follows:

[0104] ,

[0105] In the formula, This is the balance coefficient for the spurious matching loss. This is the balancing coefficient for the pseudo-label loss. The inversion loss function;

[0106] Spectral reconstructor minimizes inversion loss function update Network parameters of the spectral reconstructor ;

[0107] Pseudo-tag selector and Adversarial training of the spectral reconstructor updates the network parameters of the pseudo-label selector by maximizing the selection loss function, further constraining the network. The inversion accuracy of the spectral reconstructor; the loss function is defined as follows:

[0108] ,

[0109] To select a loss function.

[0110] This application's embodiments combine the precise mapping of supervised learning with the weak data dependency of unsupervised learning through a semi-supervised mechanism, and employ an adversarial learning strategy to continuously optimize inversion accuracy, achieving high-precision and efficient intelligent NMR logging data under limited label conditions. Spectral inversion.

[0111] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning, characterized in that, The system includes: The data processing module is used to match the spin echo signal generated by forward modeling with the corresponding spin echo signal. The spectrum consists of labeled logging datasets and unlabeled logging datasets composed of actual spin echo signals. The spectral reconstructor, trained using a labeled logging dataset, obtains the spin echo signal and preliminary... The nonlinear mapping relationship of the spectrum is used to process the unlabeled logging dataset and generate predictions. Spectrum; A pseudo-tag selector is used to compare the actual spin echo signal with the predicted one. The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset generate predictions. The confidence level of the spectrum is used to select predictions. The spectrum serves as a pseudo-label, linking the pseudo-label with the prediction. After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor; The After training on the training set, the spectral reconstructor obtains the actual spin echo signal and the predicted signal. Nonlinear mapping relationship of the spectrum.

2. The nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning according to claim 1, characterized in that, The The spectral reconstructor comprises a feature learning module and a detail enhancement module. Skip connections are used to directly pass features extracted by the feature learning module at different scales to the detail enhancement module. The feature learning module extracts features from the labeled well logging dataset. The initial feature information of the spectrum is obtained by feature fusion and input to the detail enhancement module to recover the feature detail information. Spectrum, establishing spin echo signal and preliminary Nonlinear mapping relationship of the spectrum.

3. The nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning according to claim 2, characterized in that, The feature learning module includes sequentially configured... Convolutional layers, ReLU activation layers, and multiple feature fusion blocks, The feature fusion block contains sequentially set features. Convolutional layer, batch normalization layer, ReLU activation layer, and parallel first and second branches, the first branch containing sequentially configured... Convolutional layer, dilated convolutional layer, batch normalization layer, and ReLU activation layer; the second branch contains layers set sequentially. The convolutional layer, batch normalization layer, and ReLU activation layer are combined and the two branches are merged as the output.

4. The nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning according to claim 1, characterized in that, The inversion loss function of the spectral reconstructor includes cross-entropy loss, pseudo-matching loss, and pseudo-label loss; The cross-entropy loss is used for training with a labeled well logging dataset. When evaluating the spin echo signal via the spectral reconstructor, the spin echo signal is... Preliminary results obtained after spectral reconstruction Spectrum and The degree of similarity of the spectra; The pseudo-matching loss is used to train the dataset simultaneously using both labeled and unlabeled logging datasets. When evaluating the spectral reconstructor, the prediction is evaluated. Spectrum and Spectral differences; The pseudo-label loss is used for training with an unlabeled logging dataset. When evaluating the prediction of the actual spin echo signal in the spectral reconstructor... The degree of similarity between the spectrum and the pseudo-label.

5. The nuclear magnetic resonance logging data inversion system based on semi-supervised adversarial learning according to claim 1, characterized in that, The pseudo-label selector determines the corresponding tagged logging dataset by... Predictions corresponding to spectral and pseudo-matching datasets The confidence level is obtained from the similarity of the spectra.

6. A method for inverting nuclear magnetic resonance logging data based on semi-supervised adversarial learning, characterized in that, include: The spin echo signal generated by forward modeling is compared with the corresponding spin echo signal. The spectrum consists of labeled logging datasets and unlabeled logging datasets composed of actual spin echo signals. Training was performed using a tagged logging dataset to obtain spin echo signals and preliminary data. The nonlinear mapping relationship of the spectrum is used to process the unlabeled logging dataset and generate predictions. Spectrum; Based on actual spin echo signals and predictions The pseudo-matching dataset obtained from spectral matching and the labeled well logging dataset generate predictions. The confidence level of the spectrum is used to select predictions. Spectrum as a pseudo-label; Combining pseudo-labels with prediction After spectral mixing, it is used as a label-based logging dataset and an unlabeled logging dataset. Training set for the spectral reconstructor; Training with training set The actual spin echo signal obtained after the spectral reconstructor is compared with the predicted signal. Nonlinear mapping relationships of the spectrum; Based on actual spin echo signals and predictions The nonlinear mapping relationship of the spectrum is used to process the real-time acquired spin echo signal to obtain the prediction. Spectrum.

7. The nuclear magnetic resonance logging data inversion method based on semi-supervised adversarial learning according to claim 6, characterized in that, Iterative optimization by minimizing the inversion loss function The prediction performance of the spectral reconstructor, wherein the inversion loss function includes: cross-entropy loss, pseudo-matching loss, and pseudo-label loss; The cross-entropy loss is used for training with a labeled well logging dataset. When evaluating the spin echo signal via the spectral reconstructor, the spin echo signal is... Preliminary results obtained after spectral reconstruction Spectrum and The degree of similarity of the spectra; The pseudo-matching loss is used to train the dataset simultaneously using both labeled and unlabeled logging datasets. When evaluating the spectral reconstructor, the prediction is evaluated. Spectrum and Spectral differences; The pseudo-label loss is used for training with an unlabeled logging dataset. When evaluating the prediction of the actual spin echo signal in the spectral reconstructor... The degree of similarity between the spectrum and the pseudo-label.

8. The nuclear magnetic resonance logging data inversion method based on semi-supervised adversarial learning according to claim 6, characterized in that, The The spectral reconstructor comprises a feature learning module and a detail enhancement module. Skip connections are used to directly pass features extracted by the feature learning module at different scales to the detail enhancement module. The feature learning module extracts features from the labeled well logging dataset. The initial feature information of the spectrum is obtained by feature fusion and input to the detail enhancement module to recover the feature detail information. Spectrum, establishing spin echo signal and preliminary Nonlinear mapping relationship of the spectrum.

9. The nuclear magnetic resonance logging data inversion method based on semi-supervised adversarial learning according to claim 8, characterized in that, The feature learning module includes sequentially configured... Convolutional layers, ReLU activation layers, and multiple feature fusion blocks, The feature fusion block contains sequentially set features. Convolutional layer, batch normalization layer, ReLU activation layer, and parallel first and second branches, the first branch containing sequentially configured... Convolutional layer, dilated convolutional layer, batch normalization layer, and ReLU activation layer; the second branch contains layers set sequentially. The convolutional layer, batch normalization layer, and ReLU activation layer are combined and the two branches are merged as the output.

10. The nuclear magnetic resonance logging data inversion method based on semi-supervised adversarial learning according to claim 6, characterized in that, The pseudo-label selector is used to determine which data sets correspond to the labeled logging datasets. Predictions corresponding to spectral and pseudo-matching datasets The confidence level is obtained from the similarity of the spectra.

Images