Semi-supervised domain adaptive lithologic model construction method and system

Abstract

The invention relates to the technical field of lithology identification, and discloses a semi-supervised domain adaptive lithology model construction method and system, and a lithology prediction model construction and training process comprises the steps: taking labeled logging data of an explained well as source domain data, and taking part of labeled logging data and unlabeled logging data of a target well as target domain data; constructing a threshold dynamically adjusted semi-supervised domain adaptive lithology prediction model, wherein the lithology prediction model comprises a feature extractor, a domain discriminator, a class-level intra-domain discriminator designed for each lithology class, an independent discriminator and a label classifier; training the lithology prediction model through inter-domain confrontation, intra-domain confrontation, dynamic threshold adjustment and source domain data reweighting; according to the method, a remarkable alignment effect is realized in the feature space, the data distribution difference between different well positions is effectively reduced, and meanwhile, the identification accuracy of each lithology category is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithology identification, and particularly relates to a method and system for constructing a semi-supervised domain adaptation lithology model. Background Art

[0002] Well logging reservoir lithology identification, as a subsequent step of reservoir characterization and modeling, can visually represent the petrophysical characteristics of reservoir rocks and is of great significance in the field of oil and gas exploration. Traditional methods rely on establishing lithology statistical models from domain knowledge and determining the composition and properties of underground reservoir rocks through the interpretation of well logging data, but they require a large amount of expert work. The application of machine learning algorithms not only reduces the data analysis work of domain experts but also greatly improves the efficiency and accuracy, showing extremely high practical application value. Although supervised learning methods perform well when there is an abundance of labels and have produced many research results and application cases. However, in actual production, new well exploration often faces the problem of scarce labeled data. This is because the geological logging process is complex and costly. During the new well exploration stage, coring operations are often only carried out for the most concerned depth sections, and the subsequent acquisition of lithology category labels requires complex laboratory tests and professional geological interpretations, consuming a large amount of time and human resources, resulting in most new wells having only a limited number of lithology category label data. At this time, using supervised learning methods to construct a well logging lithology model is prone to overfitting problems and has poor generalization ability. In a relatively mature research area, a large amount of high-quality lithology category label data has often been accumulated. These data provide valuable resources for training a high-performance well logging reservoir lithology model, and also contain extremely rich geological information and common information of well logging lithology mapping. The well logging and lithology category label data of the interpreted wells in the region can be integrated to assist in predicting the complete well logging reservoir lithology category labels of the uninterpreted wells.

[0003] However, due to the complexity of the sedimentary environment and the continuous update of well logging equipment, there are often significant differences in the distribution of well logging data at different well locations. Most well logging lithofacies machine learning methods usually assume that the training data and test data come from the same data distribution when modeling. This difference will make it difficult to obtain accurate well logging lithology prediction results when directly applying a trained high-performance model to uninterpreted wells. By using the t-SNE technique to visualize the distribution of well logging data in different drillings, such as Figure 1 in (a) and Figure 1As shown in (c), obvious differences can be seen in the distribution of well logging data at different well positions. Even though these drillings are relatively close geographically, the geological characteristics and sedimentary environments at different well positions vary, resulting in various morphological response laws of well logging curves for the same lithology. Moreover, over time, with the continuous evolution of well logging equipment and technology, the combined effects of various factors such as wellbore characteristics, surrounding rock properties, formation thickness, and drilling fluid make the quality and distribution characteristics of well logging data more complex and diverse. Therefore, when the model trained on Well W1 is applied to Well W2, as Figure 1 shown in (b) and Figure 1 shown in (d), by comparing the true lithology labels ("Lithology" column) and the model prediction results ("Prediction" column), the model prediction performance significantly decreases. In addition, as Figure 1 shown in (a) and Figure 1 shown in (c), the sandstones in Well W1 and Well W2 clearly form several different cluster centers, which means that there are also certain distribution differences among the well logging data at different depths in the same well position. This is because the depth penetrated by a drilling well spans different strata and sedimentary environments, resulting in a certain distribution deviation for sample data that are far apart. Since there is only a very small amount of labeled data available in the target domain, there is inevitably a distribution deviation, that is, intra-domain difference, between the labeled data and a large amount of unlabeled data within the target domain. Simply using the method of supervised learning by adding labeled data cannot effectively alleviate the distribution differences of the unlabeled data in the target domain. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a semi-supervised domain adaptation lithology model construction method and system, aiming to make full use of the well logging data and lithology category labels of the interpreted wells, combine the well logging data of the uninterpreted wells and a small amount of lithology category labels, and realize the effective transfer of knowledge by reducing the distribution differences of the well logging data between the interpreted wells and the uninterpreted wells, and finally establish a high-precision well logging reservoir lithology model in the uninterpreted wells.

[0005] To solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A semi-supervised domain adaptation lithology model construction method, which identifies the lithology of a reservoir through a trained lithology prediction model. The construction and training process of the lithology prediction model includes the following steps:

[0007] Step 1: Use the labeled log data of the interpreted well as the source domain data, and use part of the labeled log data and unlabeled log data of the target well as the target domain data; the probability distributions of the source domain data and the target domain data are different; the labeled log data includes 1D image samples constructed from log curve vectors and corresponding lithology class labels; the unlabeled log data includes the 1D image samples; the log curve vector is composed of the measured values of different types of log curves at a specific depth.

[0008] Step 2: Construct a semi-supervised domain adaptation lithology prediction model with dynamically adjusted thresholds. The lithology prediction model includes a feature extractor F, a domain discriminator D inter , a class-level intra-domain discriminator designed for each lithology class Independent discriminator and a label classifier C; train the lithology prediction model through inter-domain adversarial, intra-domain adversarial, dynamic threshold adjustment, and source domain data reweighting;

[0009] Build the feature extractor through a convolutional neural network that integrates the channel attention mechanism. The inter-domain adversarial extracts domain-invariant features by mapping the log data of different well positions to a unified feature space through the feature extractor, and constructs the loss of the domain discriminator and the loss of the label classifier

[0010] The intra-domain adversarial reduces the distribution difference between the labeled log data and the unlabeled log data in the target domain through an intra-domain discriminator that integrates the prediction confidence of the label classifier, and constructs the loss of the intra-domain discriminator

[0011] The dynamic threshold defines and adjusts the confidence threshold for each lithology class using the prediction results of the lithology prediction model, thereby optimizing the selection of pseudo-labels for unlabeled log data, and constructing the learning loss

[0012] Source domain data reweighting: Through a dynamic weight allocation mechanism based on the contribution of source domain data, different weights are assigned to the source domain data to quantitatively indicate the importance of each source domain data in the knowledge transfer process. According to the obtained weights, and are updated to obtain the weighted loss of the label classifier and the weighted loss of the discriminator

[0013] The overall loss of the lithology prediction model is:

[0014]

[0015] Among them, the hyperparameters α, β, γ, and λ play a balancing role in each loss term.

[0016] Further, Step 1 specifically includes: constructing a v-dimensional logging curve vector O = [o1, o2, …, o v ∈ R v ; o v is the v-th logging curve in O; the lithology class label y corresponding to the logging curve vector is obtained by analyzing the core or rock slice at the specific depth; on this basis, constructing a 1D image sample x = [O1, O2, …, O h ∈ R h×v ; corresponding to the lithology class label at the depth; h is the number of logging curve vectors in x; the source domain data is usually collected from an interpreted well or multiple interpreted wells in the same area, and the target domain data comes from a target well geographically adjacent to the interpreted well; the source domain dataset where n s is the number of source domain data in is the i-th 1D image sample in is the corresponding lithology class label; the target domain dataset the labeled dataset of the target domain where n tl is the number of labeled logging data in is the i2-th 1D image sample in is the corresponding lithology class label; the unlabeled dataset of the target domain where n tul represents the number of unlabeled logging data in is the i3-th 1D image sample; where n tl < n s n tl < n tul .

[0017] Further, the inter-domain adversarial specifically includes:

[0018] To eliminate the data distribution differences between the interpreted well and the target well, by introducing adversarial learning, the domain shift between different source domains and the target domain is reduced, and the total loss function of the inter-domain adversarial is used to implement the feature extractor F and the domain discriminator D interAdversarial training between; feature extractor F and domain discriminator D inter During the adversarial training, the domain label of the source domain data is marked as "0", and the domain label of the target domain data is marked as "1". The feature extractor F strives to generate a feature representation for the lithology identification task and reduce the loss of the label classifier C:

[0019]

[0020] where, is the loss of the label classifier C, represents the cross-entropy loss, x i represents the i-th 1D image sample of the input, y i is x i corresponding lithology category label; F(x i ; θ f ) is the feature representation generated by the feature extractor; meanwhile, the feature extractor tries to generate features with domain invariance to confuse the domain discriminator and seeks the feature mapping parameter θ f that maximizes the loss of the domain discriminator, making the two domain feature distributions generated as similar as possible; the domain discriminator tries to minimize the parameter θ d·inter of the domain classification loss; the purpose of the adversarial training between F and D inter is: the goal of F is to generate a domain-general feature representation and maximize the loss of D inter , while D inter tries to avoid confusion between different domains and minimize the loss of the domain discriminator

[0021]

[0022] d i represents the binary label of the i-th 1D image sample, used to indicate whether this 1D image sample belongs to the source domain or the target domain, represents the logarithmic loss function;

[0023] In summary, the total loss function of the inter-domain adversarial is:

[0024]

[0025] λ1 is the weight coefficient, θ c represents the label classifier parameter.

[0026] Furthermore, the process of the inter-domain adversarial is completed by optimizing the following min-max game:

[0027]

[0028] respectively represent the optimal parameters of the feature extractor, label classifier, and inter-domain discriminator;

[0029] During the gradient backpropagation in the training process of the gradient reversal layer, in the forward propagation process, the gradient reversal layer does not make any changes to the input data. In the backward propagation process, the gradient reversal layer multiplies the passed gradient by a negative constant, adding an optimization direction opposite to that of the domain discriminator to the feature extractor:

[0030] D grl (x) = x;

[0031]

[0032] D grl represents the forward propagation process of the gradient reversal layer, represents the backward propagation process of the gradient reversal layer, I represents the identity matrix, λ2 represents the gradient reversal layer constant, which will not be updated through backpropagation; based on this, it is possible to simultaneously handle two optimization problems in different directions in a single lithology prediction model, avoiding the need to train multiple lithology prediction models separately:

[0033]

[0034] where ι represents the magnitude of the learning rate during the model training process.

[0035] Furthermore, the feature extractor is built by integrating a convolutional neural network with a channel attention mechanism, specifically including:

[0036] Integrate the channel attention module into the convolutional neural network to obtain a convolutional neural network with a fused channel attention mechanism, that is, obtain the feature extractor; the channel attention module combines a shared network, as well as a global average pooling mechanism and a global max pooling mechanism;

[0037] Input the 1D image sample x into the convolutional neural network F′ to obtain the intermediate feature map f:

[0038] f = F′(x)

[0039] By performing global average pooling operation and global max pooling operation on the intermediate feature map f, generate the average spatial representation f GAP and the maximum spatial representation f GMP ; subsequently, concatenate f GAP and f GMP and input them into the shared network to calculate the channel attention map m ca :

[0040] m ca = σ(W(f GAP + f GMP ));

[0041] σ represents the Relu activation function; W represents the weight parameter of the shared network in the channel attention module; for the intermediate feature map f, m will be used ca to perform channel scaling calculation:

[0042] f′ = m ca ⊙ f;

[0043] where f′ is the intermediate feature map after channel scaling, and the symbol ⊙ here represents the Hadamard product.

[0044] Furthermore, the in - domain adversarial training specifically includes:

[0045] Design a class - level in - domain discriminator for each lithology category, obtaining M in - domain discriminators m = 1, 2,..., M, where M is the total number of lithology categories, and the class - level in - domain discriminator is responsible for matching the labeled well - logging data and unlabeled well - logging data related to the m - th lithology category in the target domain;

[0046] The label classifier C outputs p i for each 1D image sample x i = C(x i ), representing the probability distribution of x i belonging to M lithology categories; x i The attention to each in - domain discriminator can be expressed as the weighted feature with probability

[0047]

[0048] represents the loss of the in - domain discriminator, represents the loss of the discriminator for the m - th category, that is, the logarithmic loss function, represents the probability that the label classifier predicts the i - th 1D image sample as the m - th category, θ d·intra represents all the parameters of the in - domain discriminator, d i represents the binary label of the i - th 1D image sample, used to indicate whether this 1D image sample belongs to the source domain or the target domain; θ f is the feature mapping parameter of the feature extractor F;

[0049] The in - domain adversarial loss is defined as follows:

[0050]

[0051] θ f , θ c , θd·intra respectively represent the parameters of the feature extractor, the label classifier, and the inter-domain discriminator; represents the cross-entropy loss; θ c represents the label classifier parameters; λ3 represents the weight coefficient.

[0052] Furthermore, the dynamic threshold adjustment specifically includes:

[0053] First, estimate a global threshold through the exponential moving average of the prediction confidence of the unlabeled samples in the target domain by the lithology prediction model Then, adjust the global threshold through the exponential moving average of the probability of each lithology category estimated by the exponential moving average of the lithology prediction model, set a category-specific threshold for each lithology category, and automatically adjust their respective thresholds according to the characteristics of different lithology categories and the prediction confidence of the lithology category by the label classifier C; furthermore, obtain the learning loss of the unlabeled logging data in the target domain

[0054]

[0055] where B is the batch size, 1 ≤ b ≤ B; is an indicator function used to screen the unlabeled logging data that should be used for training through the global threshold represents the cross-entropy loss, represents the output probability distribution p of the lithology prediction model b transformed into a hard label with only one category, p b is the probability distribution output by the label classifier.

[0056] Furthermore, the estimation process of the global threshold includes:

[0057] The global threshold is initialized to and updated through the exponential moving average of the confidence of the probability distribution output by the lithology prediction model for the unlabeled logging data in the target domain at the τ-th iteration of the lithology prediction model for the initialized global threshold as follows:

[0058]

[0059] where B is the batch size, κ ∈ (0, 1) is the momentum decay factor of the exponential moving average;

[0060] The process of setting the category-specific threshold includes:

[0061] Calculate the expectation of the prediction value of the lithology prediction model for each lithology category m = 1, 2,..., M:

[0062]

[0063] Finally, a dynamically adjusted class-specific threshold is obtained:

[0064]

[0065] wherein, represents the class-specific threshold of the m-th lithology class, and Norm represents the maximum normalization method.

[0066] Furthermore, the reweighting of the source domain data specifically includes:

[0067] Pre-train an independent domain discriminator Input the original features X of the well logging data samples and the corresponding labels Y into the independent discriminator simultaneously to predict the domain label of the 1D image sample x

[0068]

[0069] and represent the joint distribution of the source domain data and the joint distribution of the target domain data respectively, X s , Y s represent the source domain feature space and the source domain label space respectively, X t , Y t represent the target domain feature space and the target domain label space respectively;

[0070] Through the probability density ratio reflect the unbiased estimation between different domains, and combine the predicted domain label of the well logging data sample to assign weights to the labeled well logging data of each source domain:

[0071]

[0072] Weight the contribution of the source domain labeled well logging data to knowledge transfer to calculate the weighted loss of the label classifier and the weighted loss of the discriminator

[0073]

[0074]

[0075] n t represents the number of well logging data in the target domain dataset.

[0076] A semi-supervised domain adaptation lithology model construction system includes:

[0077] A data collection module that uses the labeled well logging data of the interpreted wells as source domain data, and uses part of the labeled well logging data and unlabeled well logging data of the target well as target domain data; the probability distributions of the source domain data and the target domain data are different; the labeled well logging data includes 1D image samples constructed from well logging curve vectors and corresponding lithology class labels; the unlabeled well logging data includes the 1D image samples; the well logging curve vectors are composed of the measurement values of different types of well logging curves at a specific depth.

[0078] A model construction module that constructs a semi-supervised domain adaptation lithology prediction model with dynamically adjusted thresholds. The lithology prediction model includes a feature extractor F, a domain discriminator D inter and a class-level intra-domain discriminator designed for each lithology class Independent discriminator and a label classifier C; the lithology prediction model is trained through inter-domain adversarial, intra-domain adversarial, dynamic threshold adjustment, and source domain data reweighting.

[0079] The feature extractor is built by a convolutional neural network that integrates a channel attention mechanism. The inter-domain adversarial extracts domain-invariant features by mapping the well logging data of different well positions to a unified feature space through the feature extractor, and constructs the loss of the domain discriminator and the loss of the label classifier

[0080] The intra-domain adversarial reduces the distribution difference between the labeled well logging data and the unlabeled well logging data in the target domain through an intra-domain discriminator that integrates the prediction confidence of the label classifier, and constructs the loss of the intra-domain discriminator

[0081] The dynamic threshold defines and adjusts the confidence threshold for each lithology class using the prediction results of the lithology prediction model, thereby optimizing the selection of pseudo-labels for the unlabeled well logging data, and constructing a learning loss

[0082] Source domain data reweighting: Through a dynamic weight allocation mechanism based on the contribution of source domain data, different weights are assigned to the source domain data to quantitatively indicate the importance of each source domain data in the knowledge transfer process. According to the obtained weights, and are updated to obtain the weighted loss of the label classifier and the weighted loss of the discriminator

[0083] A model training module that trains the lithology prediction model based on the overall loss The overall loss is:

[0084]

[0085] Among them, the hyperparameters α, β, γ, and λ play a trade - off role in each loss term.

[0086] The system of the present invention corresponds to the method and is applicable to the specific technical solutions of the method and equally applicable to the system.

[0087] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0088] The present invention constructs a cross - domain shared feature extractor based on adversarial learning, maps well - logging data at different well positions to the same latent feature space, and while extracting domain - invariant features, ensures the class separability of features as much as possible. To address the scarcity of labeled data in the target scenario, the feature extractor is optimized. A convolutional neural network incorporating a fused channel attention mechanism is introduced as the backbone network to enhance the model's ability to capture local morphological features of well - logging curves. At the same time, by optimizing the selection of sample image sizes, the effective extraction of vertically - related depth information is ensured, compensating for the possible information loss caused by abandoning the recurrent neural network. To improve the efficiency of knowledge transfer, a dynamic weight allocation mechanism based on the contribution of source - domain data is designed. By assigning different weights to source - domain data, the importance of each sample in the knowledge transfer process is quantitatively indicated, thereby optimizing the model update effect. For the data distribution differences within the target domain, an in - domain discriminator that fuses the prediction confidence of the classifier is developed. This design makes full use of the limited labeled samples of the target well and achieves more accurate class alignment. Finally, the present invention adds a dynamic threshold adjustment module, which flexibly adjusts the utilization of samples according to the prediction confidence, improves the utilization rate of high - quality label predictions, while reducing the possible negative impacts of low - confidence samples, and effectively reduces the confirmation bias of the model. Especially for the classifier, as high - confidence pseudo - label samples are continuously added to the training, the problem of the model biasing towards the source domain has been continuously improved. The experimental results on the actual drilling dataset of Shengli Oilfield strongly confirm the superior performance of the method proposed in the present invention. The method achieves a significant alignment effect in the feature space, effectively reduces the data distribution differences between different well positions, and at the same time greatly improves the recognition accuracy of each lithology category. It is particularly worth mentioning that the dynamic threshold adjustment mechanism for class differentiation enables the model to maintain a relatively stable and accurate classification effect when facing the problem of uneven reservoir lithology distribution. Description of the Drawings

[0089] Figure 1 For the data distribution differences of well - logging data at different well positions;

[0090] Figure 2 For the framework diagram of the semi - supervised domain - adaptation lithology prediction model with dynamic threshold adjustment of the present invention;

[0091] Figure 3 It is a schematic diagram of the principle of the channel attention mechanism. Specific implementation manner

[0092] The following will make a detailed description of a preferred implementation manner of the present invention with reference to the accompanying drawings.

[0093] The present invention regards the logging data from the same exploration well as a "domain", then the labeled logging data of the interpreted well is regarded as the "source domain" for training the model to complete the learning task; the unlabeled logging data of the target well is regarded as the "target domain". The present invention proposes a semi-supervised domain adaptation lithology prediction method based on dynamic threshold adjustment, aiming to solve the problem of logging reservoir lithology prediction in the scenario where only a very small number of lithology category labels can be obtained for the target well. Specifically, the present invention introduces the adversarial idea, uses the feature extractor shared between domains to map the logging data at different well positions to the latent feature space, learns the domain-invariant features of the logging data, and continuously reduces the domain difference while ensuring the class distinguishability. At the same time, in order to enhance the extraction and attention of the feature extractor to the temporal features and morphological features of the logging curves, the channel attention mechanism is fused to emphasize the key local features. It should be noted that the recurrent neural network requires a considerable number of labeled samples for effective supervised learning. In the scenario of limited labels, it is difficult to fully learn the patterns of the time series, and it is very easy to have the overfitting problem, resulting in poor generalization performance of the model. Based on this, the feature extractor adopts a convolutional neural network integrated with the attention mechanism. In addition to being able to capture the local morphological features of the logging curves, through the subsequent network specific design (optimizing the selection of the sample picture size), the extraction of the vertically depth-related information is further realized to make up for the impact brought by modifying the recurrent neural network.

[0094] Considering that the correlation between some source domain data and target domain data is relatively low, and it is difficult to extract effective information that can be transferred to the target domain from them, the present invention introduces a dynamic weighting mechanism to allocate weights according to the potential contribution of the source domain data to the target domain knowledge transfer, and optimize the quality of knowledge transfer. Since there are only a very small number of labeled data available for the target domain, there is inevitably a distribution deviation, that is, the intra-domain difference, between the labeled logging data and the large amount of unlabeled logging data within the target domain. Simply using the method of adding labeled logging data for supervised learning cannot effectively alleviate this difference. Therefore, the intra-domain adversarial learning strategy is adopted, making full use of a small amount of labeled logging data and a large amount of unlabeled logging data, designing an intra-domain discriminator that integrates the prediction confidence of the label classifier to achieve class alignment within the target domain. This also enables a small number of labeled samples to better guide the construction of the reservoir lithology model. Finally, a dynamic threshold adjustment mechanism is introduced to adjust the sample influence according to the prediction confidence and reduce the confirmation bias. The method proposed by the present invention has completed a large number of experimental verifications in the actual drilling data of the oilfield and achieved excellent results, fully demonstrating its accuracy and applicability in the scenario of limited sample labels for the target well.

[0095] 1. Construct source domain data and target domain data

[0096] Suppose there are v different types of logging curves (including porosity curves), and the measured values of each logging curve at a specific depth form a v-dimensional logging curve vector O = [o1, o2, …, o v ∈ R v . The lithology class label corresponding to the logging curve vector O is obtained by analyzing the core or rock slice at this depth. On this basis, construct the 1D image sample x = [O1, O2, …, O h ∈ R h×v ; corresponding to the lithology class label at the depth The source domain dataset is usually collected from one well or multiple wells in the same area, while the target domain dataset comes from another well in the geographical vicinity, making the two datasets geographically related. The source domain dataset is where n s is the number of source domain data in is the i-th 1D image sample in is the corresponding lithology class label.

[0097] The target domain dataset is composed of a small amount of labeled logging data and a large amount of unlabeled logging data. Specifically, the labeled dataset where n tl is the number of labeled logging data in is the i2-th 1D image sample in is the corresponding lithology class label; where n tl << n s , n tl << n tul . The unlabeled dataset where n tul represents the number of unlabeled logging data in

[0098] The source domain data and target domain data should be sampled from different marginal probability distributions. Suppose the source domain data follows the probability distribution The target domain data follows the probability distribution Then The objective of the present invention is to construct a model that can accurately classify unlabeled well logging data in the target domain by leveraging the rich annotation information in the source domain and combining a small amount of labeled well logging data in the target domain. This method can not only improve the prediction accuracy of the model in the target domain but also effectively solve the problem of scarce labeled data.

[0099] 2. Semi-supervised Domain Adaptation Lithology Prediction Model with Dynamic Threshold Adjustment

[0100] The present invention proposes a framework for a semi-supervised domain adaptation lithology prediction model with dynamic threshold adjustment. This framework has excellent adaptive feature aggregation capabilities and can weaken the influence of low-confidence unlabeled samples according to the learning state of the model to achieve accurate identification of the lithology categories of target wells.

[0101] 2.1, Overall Network Architecture

[0102] The construction of a cross-well semi-supervised lithology prediction model in the scenario of scarce target well labels focuses on how to make full use of the limited labeled samples of target wells to guide the construction of the reservoir lithology prediction model and achieve efficient alignment between the source domain and the target domain. The proposed lithology prediction model is as Figure 2 shown, including four parts, namely: 1) inter-domain adversarial; 2) intra-domain adversarial; 3) dynamic threshold adjustment; 4) reweighting of source domain data.

[0103] First, construct a cross-domain shared feature extractor based on adversarial learning to map well logging data at different well positions to a unified feature space, extract domain-invariant features as much as possible, and continuously reduce the inter-domain difference while ensuring class distinguishability. Considering the scarcity of labeled data in the target scenario, design a convolutional neural network (CG 2 CA) with channel attention mechanism as the backbone network of the feature extractor to enhance the model's ability to capture local morphological features of well logging curves. At the same time, through the optimized selection of the subsequent sample image size, ensure the effective extraction of the temporal correlation information in the vertical depth of the well logging curve to make up for the information loss that may be caused by abandoning the recurrent neural network. Second, for the distribution difference between a small amount of labeled well logging data and a large amount of unlabeled well logging data in the target domain, develop an intra-domain discriminator that fuses the prediction confidence of the classifier, aiming to make full use of the limited labeled samples of target wells, reduce the data distribution difference within the target domain, and thus achieve more accurate class alignment. Then, design a dynamic threshold adjustment mechanism to flexibly adjust the utilization of samples according to the prediction confidence. Especially for the label classifier, as high-confidence pseudo-labeled samples are continuously added to the training, the problem of bias towards the source domain can be continuously improved. Finally, to improve the efficiency of knowledge transfer, design a dynamic weight allocation mechanism based on the contribution of source domain data. By assigning different weights to source domain data to quantitatively indicate the importance of each source domain data in the process of knowledge transfer, the update effect of the model is optimized.

[0104] 2.2, Network Module Design

[0105] 1. Inter-Domain Adversarial

[0106] To eliminate the influence of the difference in data distribution at different well positions and extract the transferable knowledge between the source domain and the target domain, an adversarial learning approach is introduced to reduce the domain shift between different domains. The total loss function of inter-domain adversarial is used to implement the adversarial training between the feature extractor F and the domain discriminator D inter During this period, the domain label of the source domain data is marked as "0", and the domain label of the target domain data is marked as "1". During the learning process, the feature extractor F strives for the feature expression of the lithology identification task and reduces the loss of the label classifier C:

[0107]

[0108] where F(x i ; θ f ) is the feature expression generated by the feature extractor. At the same time, it tries to generate features with domain invariance to confuse the domain discriminator and seeks the feature mapping parameter θ that maximizes the loss of the domain discriminator f to make the two domain feature distributions generated as similar as possible; the domain discriminator then optimizes and selects the parameter θ that can reduce the domain classification loss as much as possible d·inter ; the adversarial training between F and D inter lies in that the goal of F is to generate a general feature expression between domains and maximize the loss of D inter , while D inter tries to avoid confusing different domains and minimizes the loss of the domain discriminator

[0109]

[0110] In summary, the total loss function of inter-domain adversarial is:

[0111]

[0112] The adversarial training process is completed by optimizing the following min-max game:

[0113]

[0114] The gradient reverse propagation in the model training process is realized through the Gradient Reversal Layer (GRL). During the forward propagation process, GRL does not make any changes to the input data, that is, it works like an identity function. During the reverse propagation process, this layer multiplies the passed gradient by a negative constant (usually -1), adding an optimization direction opposite to that of the domain discriminator to the feature extractor:

[0115] Dgrl ι(x) = x;

[0116]

[0117] Based on this, two optimization problems in different directions can be processed simultaneously in a single model, avoiding the need to train multiple models separately:

[0118]

[0119] Among them, ι represents the magnitude of the learning rate during model training.

[0120] Well logging curves are correlated vertically in depth, and local morphological information is important for underground observations. However, in the case of limited labeled data, it is difficult for recurrent neural networks to fully capture the internal laws of time series, and the model is prone to overfitting and difficult to generalize well to unseen data. Therefore, the present invention uses a convolutional neural network to generate image feature representations with deep receptive fields. In addition, there are significant differences in the sensitivity of different well logging curves to different lithologies. For example, natural gamma ray logging (GR) is sensitive to rocks containing radioactive elements. Inspired by this, the expectation of feature representation also includes considerations of the importance of different feature maps. Thus, a feature extractor F is built with a convolutional neural network (CG 2 CA) with a fused channel attention mechanism as the backbone structure, focusing on the feature channels that are more important for the current task and enhancing the sensitivity and response of the network to key feature information. The channel attention module (G 2 CA) is integrated into the architecture of the feature extractor F to obtain a convolutional neural network (CG 2 CA) with a fused channel attention mechanism. The channel attention module (G 2 CA) effectively combines the global average pooling (GAP) and global max pooling (GMP) mechanisms, as Figure 3 shown. Global average pooling (GAP) extracts the average representation of all features within a channel by calculating the average value of all elements on each channel of the feature map, and can effectively identify the overall structure and retain the average information of the entire feature map. Global max pooling (GMP) extracts the maximum value on each channel, focusing on the most prominent features in the channel, and is more effective for capturing significant features in images. The results of global average pooling (GAP) and global max pooling (GMP) are concatenated and calculated to further fuse and generate the weight representation of the feature channels, identifying and emphasizing the feature channels that are most important for the current task.

[0121] Given a sample x ∈ R H×W and an intermediate feature map f = F(x) ∈ R H×W×C, where \(F\) represents the CNN network, and the layer generates two different spatial representations \(f\) through global average pooling (GAP) and global max pooling (GMP) operations GAP and \(f\) GMP Subsequently, the concatenated representation is input into the shared network to calculate the channel attention map \(m\) ca \(\in \mathbb{R}\) C×1×1 :

[0122] \(m\) ca \(=\sigma(W(f\) GAP +f\) GMP ));

[0123] The original intermediate feature map \(f\) will use \(m\) ca for channel scaling calculation:

[0124] \(f' = m\) ca \(\odot f\);

[0125] Here, the symbol \(\odot\) represents the Hadamard product, and \(W\) represents the weight parameters of the shared network in the channel attention module (\(G\) 2 CA). The convolutional neural network (\(CG\) 2 CA) technology that integrates the channel attention mechanism effectively modulates the suppression or enhancement of specific feature channels, adaptively recalibrates the feature map for each channel by weighting the differences between channels, and enhances the model's ability to preferentially represent task-related features.

[0126] 2. Intra-domain adversarial

[0127] In the above inter-domain adversarial learning process, since the label classifier \(C\) is trained with source domain data, this causes the decision of the label classifier to be biased towards the data features of the source domain. However, in practical applications, the depth involved in a well spans multiple strata and different sedimentary environments, so there will be a certain distribution bias in the sample data that is far apart. Since there is only a very small amount of labeled logging data available in the target domain, there is an inevitable distribution bias between the labeled logging data and a large amount of unlabeled logging data in the target domain, that is, the intra-domain difference. Simply adding labeled logging data for supervised learning cannot effectively alleviate the distribution difference of the unlabeled logging data in the target domain. Therefore, through intra-domain adversarial, as Figure 2 shown, a more refined multi-modal alignment method is used to reduce the probability distribution difference between the labeled and unlabeled logging data in the target domain. Specifically, a class-level intra-domain discriminator is designed for each category, that is, \(m = 1, 2, \ldots, M\), where \(M\) is the total number of lithology categories, and the class-level intra-domain discriminator is responsible for specifically matching the labeled and unlabeled logging data in the target domain related to the \(m\)-th lithology category.

[0128] Since the lithology categories to which the unlabeled data belong are completely unknown in the study, it is impossible to directly know which class-level domain discriminator should process a target-domain unlabeled logging data. It should be noted that the output p i = C(x i ) of the label classifier C for each logging data x represents the probability that the logging data belongs to M lithology categories. Here, the probability distribution p i can be used to indicate which class-domain discriminators i the logging data x should focus on and the degree of attention; specifically, the attention of x i to each domain discriminator can be expressed as the feature weighted by the probability

[0129]

[0130] In this way, even if the target-domain data has no labeled information, the soft labels of the label classifier can still be used to effectively assign the target-domain data to the class-domain discriminators according to the prediction probability, so as to achieve the in-domain distribution matching of the target domain. The in-domain adversarial loss is defined as follows:

[0131]

[0132] 3. Dynamic Threshold Adjustment

[0133] The dynamic threshold adjustment mechanism automatically defines and adjusts the confidence threshold for each category by using the model prediction results, so as to optimize the selection of pseudo-labels for unlabeled samples and reduce the confirmation bias of the model. Suppose the threshold of category m in the τ-th iteration is denoted as First, a global threshold is estimated through the exponential moving average (EMA) of the confidence, as an indicator of the overall model confidence. Further, the global threshold is adjusted by estimating the EMA of the probability of each category in the model to set a locally specific threshold for each category. In this way, the present invention can automatically adjust the respective thresholds according to the characteristics of different categories and the prediction confidence of the label classifier C for these categories. At the initial stage of model training, the threshold is set low to allow more potentially correct samples to participate in the training. As the model becomes more confident, the threshold will be dynamically adjusted upward to filter out samples with low confidence, so as to reduce the participation of incorrect sample labels in the calculation.

[0134] Global Threshold: Here, the global threshold reflects the confidence of the model in the unlabeled samples of the target domain, which means it should be adjusted according to the prediction confidence of the model in the unlabeled samples of the target domain to better evaluate the overall training of the model. During the training process, the global threshold gradually increases to ensure that low-confidence sample pseudo-labels are filtered out, thereby improving the accuracy and reliability of learning. Specifically, the global threshold is set to the average confidence of the model in the unlabeled well logging data of the target domain at the τ-th iteration. Since it is very time-consuming to calculate the confidence of all unlabeled data in each iteration, the exponential moving average (EMA) is used to estimate the global confidence. is initialized as a relatively low initial threshold to ensure that the model can accept enough samples for learning at the initial stage of training, as follows:

[0135]

[0136] where B is the batch size, and p b = C(y|x b ) is the output probability of the label classifier, and (κ ∈ (0, 1)) is the momentum decay factor of EMA.

[0137] Class-Specific Threshold. To better handle the characteristics within each class and the relationships between different classes, different thresholds are set for different classes instead of using a unified global threshold. Calculate the expectation of the model's prediction values for each class to estimate the learning state of the model for each class, and to more accurately understand the performance of the model on each class, so as to make more reasonable adjustments:

[0138]

[0139] Finally, the following dynamically adjusted class-specific thresholds are obtained:

[0140]

[0141] where Norm represents the maximum normalization method. Based on the above, the learning loss of the unlabeled well logging data in the target domain is expressed as:

[0142]

[0143] where the indicator function is used for confidence-based threshold screening to determine which unlabeled well logging data samples should be used for training, and it is a hard label with only one class converted from the output probability distribution p b of the model, representing the cross-entropy loss.

[0144] 4. Reweighting of Source Domain Data

[0145] The reweighting of source domain data aims to reduce the impact of source domain data with less relevance to the target domain on the migration effect. Starting from the features or attributes of the original samples, it can be found that the difference in feature distribution among different samples can reflect the degree of correlation between samples. In other words, by analyzing the distribution differences of sample features, the relationship or similarity between source domain data and target domain data can be understood and evaluated. Therefore, it is designed to reweight the source domain data during the learning process to further enhance the extraction of domain-invariant features and promote the migration effect of knowledge from the source domain to the target domain. Specifically, first, an independent domain discriminator is pre-trained According to its discrimination of samples from different domains, the similarity degree of each source domain data relative to the target domain data during the model training process is measured. Since the joint feature distribution can more comprehensively reflect the relationship between the features and labels of samples, it can better capture the differences between source domain and target domain data. The independent discriminator takes both the original features X of the sample data and the corresponding label Y as inputs to discriminate which domain the sample belongs to.

[0146]

[0147] Among them, an output of 0 from the independent domain discriminator represents the source domain, and 1 represents the target domain, and respectively indicate the joint distributions of source domain and target domain data.

[0148] Next, use the pre-trained independent domain discriminator to quantify the distribution differences between each source domain and target domain sample. These distribution differences represent the degree of difference of samples in the feature space. Considering that the probability density ratio can reflect the unbiased estimate between different domains, combined with the predicted sample domain labels transform these quantified distribution differences. According to the results of density ratio estimation, weights are assigned to each source domain sample:

[0149]

[0150] By updating the target loss function of the model and introducing these weights into the process of transfer learning, the contribution of source domain samples to knowledge transfer will be weighted according to the weights, so as to ensure that those source domain samples that are more similar to target domain samples and more helpful for improving the model performance receive more attention and emphasis during training, and more effective sample screening and knowledge transfer are achieved:

[0151]

[0152] 5. Overall Workflow

[0153] In summary, the overall optimization objective of the semi-supervised domain adaptation lithology prediction model with dynamic threshold adjustment is clearly defined as follows:

[0154]

[0155] Among them, the hyperparameters α, β, γ, and λ play a balancing role in each loss term.

[0156] The following summarizes the process of predicting the lithology category of logging reservoirs using the lithology prediction model proposed in the present invention.

[0157] Input: Source domain dataset from the interpreted wells Labeled dataset of the target well And unlabeled dataset

[0158] Output: Prediction results of the lithology category of the unlabeled data of the target well

[0159] (1): Normalize the logging data of the source domain and the target domain, complete the prediction work of porosity parameters, construct 1D image samples, and divide the dataset into a training set and a validation set;

[0160] (2): Pretrain and save an independent domain discriminator

[0161] (3): Input the logging data and labels of the source domain into the independent domain discriminator And obtain the weight w of the source domain data;

[0162] (4): Input D s 、D tl 、D tul into each module of the lithology prediction model correspondingly;

[0163] (5): Update and

[0164] (6): Optimize the lithology prediction model by minimizing the loss function until the preset number of iteration cycles is reached;

[0165] (7): Adjust the hyperparameters using the validation set not involved in training, and save the lithology prediction model with the best performance.

[0166] Testing: Use the saved lithology prediction model to predict the lithology category of the unlabeled logging data of the target well.

[0167] In a preferred embodiment, the logging curve vector uses six logging curves and a porosity parameter curve. The logging curves are respectively: Acoustic Logging Curve (AC), Caliper Logging Curve (CAL), Compensated Neutron Logging Curve (CNL), Gamma Ray Logging Curve (GR), 2.5m Bottom Gradient Resistivity Logging Curve (R25), and Spontaneous Potential Logging Curve (SP).

[0168] SP and R25 are electrochemical logging, which can effectively distinguish shale and sandstone. The SP curve also reflects the fluid type in the pores, such as oil or water. The SP amplitude of water-bearing sand is usually higher than that of oil-bearing sand. AC is an acoustic property logging, which reflects reservoir characteristics such as lithology, porosity, and fluid. CNL and GR are radioactive logging, which are sensitive to specific types of rocks and fluids. GR is especially sensitive to highly radioactive lithologies (such as igneous rocks) and petroleum. CAL, as an auxiliary logging curve, is closely related to lithology identification. For example, an increase in CAL indicates borehole enlargement, which may mean the presence of mudstone (MS) or coal seam because they have relatively low hardness characteristics.

[0169] The porosity curve can help identify different types of reservoirs and plays a key role in lithology prediction. High porosity is usually associated with good reservoir quality, while low porosity may be associated with tight rocks or non-reservoirs.

[0170] By combining with other logging curves (such as SP, GR, CNL, etc.), different lithologies can be more accurately distinguished.

[0171] For example, high porosity combined with a low GR value may indicate sandstone, while low porosity combined with a high GR value may indicate shale or mudstone.

[0172] The combination of these logging curves and the porosity curve has sufficient distinguishability (i.e., the eigenvector has discriminability) to effectively predict the lithology of the model.

[0173] In a preferred embodiment, the lithology class labels are obtained by analyzing cores or cuttings collected from a borehole. The lithology classes to be predicted include any one or more of Mudstone (MS), Siltstone (SS), Oil Shale (Os), Dolomite (DM), Siltstone (SI), and Fine Sandstone (FS). Accurately predicting these lithology classes is very important for reservoir modeling in the target well.

[0174] In a preferred embodiment, for each logging curve, image samples are constructed in the depth direction. Taking a certain depth point as the center, 16 points are taken upward and downward respectively. The center points are successively repeated in depth for a well, and the image sample labels correspond to the core logging at the center depth point.

[0175] In a preferred embodiment, the source domain data is divided into an 80% training set and a 20% validation set, and the target domain unlabeled data is completely invisible to the model.

[0176] It should be noted that since it is difficult to ensure that the same type of instrument, the same standard scale, and the same operation method are used for the logging data collected from different wells. The orders of magnitude of different logging curves are different, which will affect the model's judgment of the importance of features. Therefore, a normalization method is needed to eliminate the systematic errors caused by factors such as instrument performance, calibration, and personnel, as well as the impact of data scale on model training. In a preferred embodiment, the normalization method uses the min-max method.

[0177] In a preferred embodiment, the feature extractor consists of a convolutional neural network (CG 2 CA) with three layers of fused channel attention mechanism. The label classifier, domain discriminator, and source domain data reweighting module are all composed of two layers of fully connected networks. The within-domain discriminator is set separately according to the lithology classes, and each within-domain discriminator related to a class is also composed of two layers of fully connected networks. Avg-Pool and Max-Pool in the feature extractor use adaptive calculation. To further reduce the impact of data class imbalance, the weights weight of various samples on the supervised loss function are adjusted according to the sample sizes of different classes m , defined as follows:

[0178]

[0179] where N m represents the number of samples of a specific lithology type, N totalDenote the total number of training samples to ensure that the minority class receives sufficient training. The model is trained using an Adam-based optimizer. During this training process, the learning rate is set to 0.01, and the weight decay coefficient is set to 0.00005. The annealing algorithm is used to adjust the learning rate according to the model training process to ensure that the model learns quickly in the initial stage of training and converges more stably in the later stage of training.

[0180] In a preferred embodiment, the hyperparameters α, β, γ, λ of the target loss function are respectively set to 1, 0.5, 0.5, 0.1, and the initial threshold of the threshold dynamic adjustment module is set to A relatively low initial threshold to ensure that in the initial stage of training, the model can receive a sufficient number of samples for training.

[0181] Based on the high-precision prediction of logging reservoir porosity parameters, for the general scenario where the lithology prior is known, the present invention proposes a logging reservoir lithology prediction model based on semi-supervised domain adaptation, which can effectively utilize limited target domain label data to guide the model construction process, thereby achieving accurate prediction of the complete lithology sequence of the target well. Thus, it provides a broader lithology category label basis for constructing a comprehensive source domain in the work area.

[0182] The system of the present invention corresponds to the method and is applicable to the specific technical solutions of the method and also to the system.

[0183] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention, and any reference signs in the claims should not be regarded as limiting the claims involved.

[0184] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative way of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for constructing a semi-supervised domain adaptation lithology model, characterized in that Identify the lithology of the reservoir through the trained lithology prediction model. The construction and training process of the lithology prediction model includes the following steps: Step 1: Use the labeled well logging data of the interpreted wells as the source domain data, and use part of the labeled well logging data and the unlabeled well logging data of the target wells as the target domain data. The probability distributions of the source domain data and the target domain data are different. The labeled well logging data includes 1D image samples constructed from well logging curve vectors and corresponding lithology class labels. The unlabeled well logging data includes the 1D image samples. The well logging curve vectors are composed of the measured values of different types of well logging curves at a specific depth. Step 2: Construct a semi-supervised domain adaptation lithology prediction model with dynamically adjusted thresholds. The lithology prediction model includes a feature extractor F, a domain discriminator D inter , a class-level intra-domain discriminator designed for each lithology category Independent discriminator and a label classifier C; train the lithology prediction model through inter-domain adversarial, intra-domain adversarial, dynamic threshold adjustment, and source domain data reweighting; Build a feature extractor through a convolutional neural network that integrates a channel attention mechanism. The inter-domain adversarial training extracts domain-invariant features by mapping well logging data at different well positions to a unified feature space through the feature extractor, and constructs the loss of the domain discriminator and the loss of the label classifier The in-domain adversarial approach reduces the distribution difference between the labeled and unlabeled logging data in the target domain through an in-domain discriminator that fuses the prediction confidence of the label classifier, and constructs the loss of the in-domain discriminator. The dynamic threshold defines and adjusts the confidence threshold for each lithology category using the prediction results of the lithology prediction model, thereby optimizing the selection of pseudo-labels for unlabeled log data and constructing a learning loss Source domain data reweighting: By means of a dynamic weight assignment mechanism based on the contribution of source domain data, different weights are assigned to the source domain data to quantitatively indicate the importance of each source domain data in the knowledge transfer process. According to the obtained weights, and are updated to obtain the weighted loss of the label classifier and the weighted loss of the discriminator Overall loss of the lithology prediction model is as follows: Among them, the hyperparameters α, β, γ, and λ play a balancing role in each loss term.

2. The semi-supervised domain adaptation lithology model construction method according to claim 1, wherein, Step 1 specifically includes: The measurement values of v different types of logging curves at a specific depth form a v-dimensional logging curve vector O = [o1, o2, …, o v ∈ R v ; o v is the v-th logging curve in O; the lithology class label y corresponding to the logging curve vector is obtained by analyzing the core or rock slice at the specific depth; on this basis, a 1D image sample x = [O1, O2, …, O h ∈ R h×v ; corresponding to the lithology class label at depth h is the number of logging curve vectors in x; the source domain data is usually collected from an interpreted well or multiple interpreted wells in the same area, and the target domain data comes from a target well geographically adjacent to the interpreted well; the source domain dataset where n s is the number of source domain data in is the i-th 1D image sample in is the corresponding lithology class label; the target domain dataset the labeled dataset of the target domain where n tl is the number of labeled logging data in is the i2-th 1D image sample in is the corresponding lithology class label; the unlabeled dataset of the target domain where n tul represents the number of unlabeled logging data in is the i3-th 1D image sample; where, n tl < n s n tl < n tul .

3. The semi-supervised domain adaptation lithology model construction method according to claim 2, wherein The inter-domain adversarial specifically includes: To eliminate the data distribution differences between the interpreted wells and the target wells, the domain shift between different source domains and the target domain is reduced by introducing adversarial learning, and the total loss function of inter-domain adversarial is used to implement the adversarial training between the feature extractor F and the domain discriminator D inter During the adversarial training between the feature extractor F and the domain discriminator D inter the domain label of the source domain data is marked as "0", and the domain label of the target domain data is marked as "1". The feature extractor F tries to generate the feature expression of the lithology identification task and reduce the loss of the label classifier C: Among them, is the loss of the label classifier C, represents the cross-entropy loss, x i represents the i-th 1D image sample of the input, y i is the i corresponding lithology category label of x i ; θ f ) is the feature representation generated by the feature extractor; at the same time, the feature extractor tries to generate features with domain invariance to confuse the domain discriminator and seeks the feature mapping parameter θ f that maximizes the loss of the domain discriminator, so that the two domain feature distributions generated are as similar as possible; the domain discriminator is the parameter θ d·inter that minimizes the domain classification loss; the purpose of the adversarial training between F and D inter is as follows: the goal of F is to generate a domain-general feature representation and maximize the loss of D inter , while D inter tries to avoid confusion between different domains and minimize the loss of the domain discriminator d i Indicates the binary label of the i-th 1D image sample, used to represent whether this 1D image sample belongs to the source domain or the target domain. Indicates the logarithmic loss function; In summary, the total loss function for inter-domain confrontation is as follows: λ1 is the weight coefficient, and θ c represents the label classifier parameter.

4. The semi-supervised domain adaptation lithology model construction method according to claim 3, wherein The process of inter-domain adversarial is completed by optimizing the following min-max game: respectively represent the optimal parameters of the feature extractor, label classifier, and inter-domain discriminator; Through the gradient backpropagation in the training process of the gradient reversal layer, in the forward propagation process, the gradient reversal layer does not change the input data. In the backpropagation process, the gradient reversal layer multiplies the transmitted gradient by a negative constant, adding an optimization direction opposite to that of the domain discriminator to the feature extractor: D grl f(x) = x; D grl Represents the forward propagation process of the gradient reversal layer, represents the backward propagation process of the gradient reversal layer, I represents the identity matrix, and λ2 represents the gradient reversal layer constant, which will not be updated through backpropagation; based on this, it is possible to simultaneously handle two optimization problems in different directions in a single lithology prediction model, avoiding the need to train multiple lithology prediction models separately: Among them, ι represents the magnitude of the learning rate during the model training process.

5. The semi-supervised domain adaptation lithology model construction method according to claim 1 or 2, characterized in that The construction of the feature extractor by integrating the convolutional neural network with the channel attention mechanism specifically includes: Integrate the channel attention module into the convolutional neural network to obtain the convolutional neural network with the integrated channel attention mechanism, that is, obtain the feature extractor. The channel attention module combines the shared network, as well as the global average pooling mechanism and the global max pooling mechanism. Input the 1D image sample x into the convolutional neural network F ′ , and obtain the intermediate feature map f: f = F ′ (x) By performing global average pooling operation and global maximum pooling operation on the intermediate feature map f, an average spatial representation f GAP and a maximum spatial representation f GMP are generated respectively; subsequently, f GAP and f GMP are concatenated and input into the shared network to calculate the channel attention map m ca : m ca = σ(W(f GAP + f GMP )); σ represents the Relu activation function; W represents the weight parameter of the shared network in the channel attention module; for the intermediate feature map f, m will be used ca to perform channel scaling calculation: f ′ = m ca ⊙f; Among them, f ′ is the intermediate feature map after channel scaling, and the symbol ⊙ here represents the Hadamard product.

6. The semi-supervised domain adaptation lithology model construction method according to claim 2, wherein The intra-domain adversarial specifically includes: Design a class-level intra-domain discriminator for each lithology category to obtain M intra-domain discriminators M is the total number of lithology categories, and the class-level intra-domain discriminator is responsible for matching the labeled and unlabeled well logging data in the target domain related to the m-th lithology category; The label classifier C outputs p i for each 1D image sample x i = C(x i ), representing the probability distribution of x i belonging to M lithology categories; The attention of x i to each in-domain discriminator can be expressed as the feature weighted by the probability Denotes the loss of the in - domain discriminator, Denotes the loss of the discriminator for class m domain, i.e., the logarithmic loss function, Denotes the probability that the label classifier predicts the i - th 1D image sample as class m, θ d·intra Denotes all the parameters of the in - domain discriminator, d i Denotes the binary label of the i - th 1D image sample, used to indicate whether this 1D image sample belongs to the source domain or the target domain; θ f Are the feature mapping parameters of the feature extractor F; Intra-domain adversarial loss is defined as follows: θ f , θ c , θ d·intra represent the parameters of the feature extractor, the label classifier, and the inter-domain discriminator, respectively; represents the cross-entropy loss; θ c represents the label classifier parameter; λ3 represents the weight coefficient.

7. The semi-supervised domain adaptation lithology model construction method according to claim 1, wherein The dynamic threshold adjustment specifically includes: First, estimate a global threshold through the exponential moving average of the prediction confidence of the unlabeled samples in the target domain by the lithology prediction model Then, adjust the global threshold through the exponential moving average of the probability of each lithology category by the lithology prediction model estimated by the exponential moving average, set a category-specific threshold for each lithology category, and realize automatic adjustment of their respective thresholds according to the characteristics of different lithology categories and the prediction confidence of the lithology category by the label classifier C; furthermore, obtain the learning loss of the unlabeled logging data in the target domain Among them, B is the batch size, where 1 ≤ b ≤ B; is an indicator function used to screen unlabeled logging data that should be used for training through a global threshold ; represents the cross-entropy loss, and b denotes the hard label with only one class converted from the output probability distribution p b of the lithology prediction model, where p is the probability distribution output by the label classifier.

8. The semi-supervised domain adaptation lithology model construction method according to claim 7, characterized in that The estimation process of the global threshold includes: Global threshold is initialized to and updated by the exponential moving average of the confidence of the probability distribution output by the lithology prediction model for the unlabeled logging data in the target domain at the τ-th iteration: Update is performed as follows: Among them, B is the batch size, and κ∈(0,1) is the momentum decay factor of the exponential moving average; The setting process of the class-specific threshold includes: Calculate the expectation of the predicted values of the lithology prediction model for each lithology category Finally, obtain the dynamically adjusted class-specific threshold: Among them, represents the class-specific threshold of the m-th lithology category, and Norm represents the maximum normalization method.

9. The semi-supervised domain adaptation lithology model construction method according to claim 2, wherein The reweighting of the source domain data specifically includes: Pre-train an independent domain discriminator Simultaneously input the original features X of the well logging data samples and the corresponding labels Y into the independent discriminator to predict the domain labels of the 1D image samples x and respectively represent the joint distribution of source domain data and the joint distribution of target domain data, where X s , Y s respectively represent the source domain feature space and the source domain label space, and X t , Y t respectively represent the target domain feature space and the target domain label space; Through probability density ratio reflects the unbiased estimation between different domains, combined with the domain labels of the predicted log data samples assign weights to the labeled log data for each source domain: Weight the contribution of the labeled logging data in the source domain to knowledge transfer to calculate the weighted loss of the label classifier and the weighted loss of the discriminator n t represents the number of logging data in the target domain dataset.

10. A semi-supervised domain adaptation lithology model construction system, characterized in that, Includes: Data collection module: Use the labeled well logging data of the interpreted wells as the source domain data, and use part of the labeled well logging data and the unlabeled well logging data of the target wells as the target domain data. The probability distributions of the source domain data and the target domain data are different. The labeled well logging data includes 1D image samples constructed from well logging curve vectors and corresponding lithology class labels. The unlabeled well logging data includes the 1D image samples. The well logging curve vectors are composed of the measured values of different types of well logging curves at a specific depth. Model construction module, which constructs a semi-supervised domain adaptation lithology prediction model with dynamically adjusted thresholds. The lithology prediction model includes a feature extractor F, a domain discriminator D inter , a class-level intra-domain discriminator designed for each lithology category Independent discriminator and a label classifier C; the lithology prediction model is trained through inter-domain adversarial, intra-domain adversarial, dynamic threshold adjustment, and source domain data reweighting; Build a feature extractor through a convolutional neural network that integrates a channel attention mechanism. The inter-domain adversarial training uses the feature extractor to map the logging data of different well positions into a unified feature space to extract domain-invariant features, and constructs the loss of the domain discriminator and the loss of the label classifier The in-domain adversarial approach reduces the distribution difference between the labeled well logging data and the unlabeled well logging data in the target domain through an in-domain discriminator that fuses the prediction confidence of the label classifier, and constructs the loss of the in-domain discriminator. The dynamic threshold defines and adjusts the confidence threshold for each lithology category using the prediction results of the lithology prediction model, thereby optimizing the pseudo-label selection of unlabeled logging data and constructing a learning loss Source domain data reweighting: By means of a dynamic weight allocation mechanism based on the contribution of source domain data, different weights are assigned to the source domain data to quantitatively indicate the importance of each source domain data in the knowledge transfer process. According to the obtained weights, and are updated to obtain the weighted loss of the label classifier and the weighted loss of the discriminator Model training module, based on the overall loss Train the lithology prediction model, and the overall loss is as follows: Among them, the hyperparameters α, β, γ, and λ play a balancing role in each loss term.

Images