A semi-supervised domain adaptation lithology model construction method and system

Abstract

The application relates to the technical field of lithology identification, and discloses a semi-supervised domain adaptation lithology model construction method and system, and a lithology prediction model construction and training process, which comprises the following steps: taking labeled well logging data of an interpreted well as source domain data, and taking part of the labeled well logging data and unlabeled well logging data of a target well as target domain data; a semi-supervised domain adaptation lithology prediction model with dynamic threshold adjustment is constructed, the lithology prediction model comprises a feature extractor, a domain discriminator, a class-level intra-domain discriminator designed for each lithology category, an independent discriminator and a label classifier; the lithology prediction model is trained through inter-domain confrontation, intra-domain confrontation, dynamic threshold adjustment and source domain data reweighting; and the application achieves significant alignment effect in a feature space, effectively reduces the data distribution difference between different well sites, and greatly improves the identification accuracy of each lithology category.

Description

Technical Field

[0001] This invention relates to the field of lithology identification technology, specifically to a semi-supervised domain-adaptive lithology model construction method and system. Background Technology

[0002] Well logging reservoir lithology identification, as a subsequent step in reservoir characterization and modeling, provides an intuitive representation of the physical characteristics of reservoir rocks and is of great significance in the field of oil and gas exploration. Traditional methods rely on building lithological statistical models from domain knowledge and interpreting well logging data to determine the composition and properties of underground reservoir rocks, but this requires extensive expert involvement. The application of machine learning algorithms not only reduces the data analysis work of domain experts but also significantly improves efficiency and accuracy, demonstrating extremely high practical application value. Although supervised learning methods perform well when labels are plentiful, resulting in numerous research findings and application cases, new well exploration in actual production often faces the problem of scarce label data. This is because geological logging technology is complex and costly; in the new well exploration phase, coring operations are often only performed at the depths of greatest interest, and the subsequent acquisition of lithology category labels requires complex laboratory testing and professional geological interpretation. The significant consumption of time and human resources results in most new wells having only a limited number of lithology category label data. In this case, using supervised learning methods to build well logging lithology models is prone to overfitting and poor generalization ability. In well-developed research areas, a large amount of high-quality lithology category label data has been accumulated. This data provides valuable resources for training high-performance well logging reservoir lithology models and contains extremely rich geological information as well as common information of well logging lithology mapping. By integrating well logging and lithology category label data of interpreted wells in the region, it can assist in the prediction of complete well logging reservoir lithology category labels for uninterpreted wells.

[0003] However, due to the complexity of sedimentary environments and the continuous upgrading of logging equipment, the distribution of logging data at different well locations often varies significantly. Most logging lithofacies machine learning methods typically assume that training and testing data come from the same data distribution. This discrepancy makes it difficult to obtain accurate logging lithology prediction results when directly applying a trained high-performance model to an uninterpreted well. The t-SNE technique can be used to visualize the distribution of logging data in different wells, such as... Figure 1 (a) and Figure 1As shown in (c), significant differences in the distribution of logging data across different well locations can be observed. Even though these wells are geographically close, variations in geological characteristics and sedimentary environments at different well locations lead to logging curves with the same lithology exhibiting diverse morphological response patterns. Furthermore, with the continuous evolution of logging equipment and technology over time, the combined effects of various factors such as wellbore characteristics, surrounding rock properties, formation thickness, and drilling fluids result in increasingly complex and diverse logging data quality and distribution characteristics. Therefore, when the model trained on well W1 is applied to well W2, as... Figure 1 (b) and Figure 1 As shown in (d) above, comparing the actual lithology labels (“Lithology” column) and the model prediction results (“Prediction” column), the model prediction performance is significantly reduced. Furthermore, as... Figure 1 (a) and Figure 1 As shown in (c), the sandstone in wells W1 and W2 clearly forms several distinct cluster centers, indicating a certain distributional difference in logging data at different depths within the same well location. This is because a single well traverses different formations and sedimentary environments, leading to distributional biases in samples located far apart. Since only a very small amount of labeled data is available in the target domain, an unavoidable distributional bias exists between labeled data and a large amount of unlabeled data within the target domain—i.e., intra-domain variability. Simply adding labeled data for supervised learning cannot effectively alleviate the distributional variability of unlabeled data within the target domain. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a semi-supervised domain adaptive lithology model construction method and system. The aim is to fully utilize logging data and lithology category labels from interpreted wells, combined with logging data and a small number of lithology category labels from uninterpreted wells. By reducing the distribution differences in logging data between interpreted and uninterpreted wells, effective knowledge transfer is achieved, ultimately establishing a high-precision logging reservoir lithology model in the uninterpreted wells.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A semi-supervised domain-adaptive lithology model construction method identifies reservoir lithology using a trained lithology prediction model. The lithology prediction model construction and training process includes the following steps:

[0007] Step 1: Use the labeled logging data from the interpreted wells as source domain data, and use a portion of the labeled logging data and unlabeled logging data from the target wells as target domain data; the probability distributions of the source domain data and the target domain data are different; the labeled logging data includes 1D image samples constructed from logging curve vectors and corresponding lithology category labels; the unlabeled logging data includes the 1D image samples; the logging curve vectors consist of measurement values ​​of different types of logging curves at a specific depth;

[0008] Step 2: Construct a semi-supervised domain-adaptive lithology prediction model with dynamically adjusted thresholds. The lithology prediction model includes a feature extractor F and a domain discriminator D. inter Class-level intra-domain discriminants designed for each lithology category Independent discriminator The lithology prediction model is trained using a label classifier C and inter-domain adversarial, intra-domain adversarial, dynamic threshold adjustment, and source domain data reweighting.

[0009] A feature extractor is built using a convolutional neural network that incorporates a channel attention mechanism. The inter-domain adversarial mechanism uses the feature extractor to map logging data from different well locations to a unified feature space to extract domain-invariant features, and constructs the loss function for the domain discriminator. Loss of the label classifier

[0010] The intra-domain adversarial mechanism reduces the distributional differences between labeled and unlabeled logging data within the target domain by fusing the confidence predictions of the labeled classifier with an intra-domain discriminator, and constructs the loss function of the intra-domain discriminator.

[0011] The dynamic threshold is defined and adjusted using the prediction results of the lithology prediction model, thereby optimizing the pseudo-label selection of unlabeled logging data and constructing a learning loss.

[0012] Source domain data reweighting: This involves assigning different weights to source domain data based on their contributions using a dynamic weighting mechanism. This quantitatively indicates the importance of each source domain data point in the knowledge transfer process. The weights are then used to reweight the data. and Update the data to obtain the weighted loss of the label classifier. Discriminator-weighted loss

[0013] Overall loss of lithology prediction model for:

[0014]

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

[0016] Furthermore, step one specifically includes: constructing a v-dimensional logging curve vector O = [o1, o2, ..., o] from the measured values ​​of v different types of logging curves at a specific depth. v ]∈R v ;o v Let be the v-th logging curve in o; the lithology category label y corresponding to the logging curve vector is obtained by analyzing the core or rock slice at the specified depth; based on this, construct 1D image samples x = [O1, O2, ..., O h ]∈R h×v ; Corresponding to Lithology category labels at depth h represents the number of logging curve vectors in x; source domain data is typically collected from a single interpreted well or multiple interpreted wells in the same region, while target domain data comes from geographically adjacent interpreted wells; source domain dataset Where n s yes The amount of data in the source domain, yes The i-th 1D image sample in the image. yes Corresponding lithology category labels; target domain dataset Labeled datasets in the target domain Where, n tl for The number of tagged logging data in the data. for The i2th 1D image sample in the middle, for Corresponding lithology category labels; unlabeled dataset of the target domain Where, n tul express The amount of unlabeled logging data in China for The i-th 1D image sample; where n tl <n s n tl <n tul .

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

[0018] To eliminate the data distribution discrepancy between interpreted wells and target wells, adversarial learning is introduced to reduce the domain offset between different source and target domains, utilizing the total loss function of inter-domain adversarial learning. Implement the feature extractor F and the domain discriminator D. interAdversarial training between them; feature extractor F and domain discriminator D inter During adversarial training, the domain labels of the source domain data are marked as "0", and the domain labels of the target domain data are marked as "1". The feature extractor F strives to generate feature representations for the lithology identification task and reduce the loss of the label classifier C.

[0019]

[0020] in, The loss of the label classifier C is... Denotes cross-entropy loss, x i Let y represent the i-th 1D image sample from the input. i For x i Corresponding lithological category label; F(x) i ;θ f The feature extractor generates feature representations; simultaneously, the feature extractor generates features with domain invariance as much as possible to confuse the domain discriminator, seeking the feature mapping parameter θ that maximizes the loss of the domain discriminator. f This ensures that the feature distributions of the two generated domains are as similar as possible; the domain discriminator, on the other hand, minimizes the parameter θ of the domain classification loss. d·inter F and D inter The purpose of adversarial training is: the goal of F is to generate a universal feature representation across domains, maximizing D. inter The loss, and D inter It attempts to avoid confusing different domains and minimize the loss of the domain discriminator.

[0021]

[0022] d i Let represent 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 for inter-domain adversarial competition for:

[0024]

[0025] λ1 is the weighting coefficient, θ c This represents the parameters of the label classifier.

[0026] Furthermore, the inter-domain adversarial process is accomplished by optimizing the following mini-maximum game:

[0027]

[0028] These represent the optimal parameters for the feature extractor, label classifier, and inter-domain discriminator, respectively.

[0029] During the training process of the gradient inversion layer, gradient backpropagation occurs. In the forward propagation, the gradient inversion layer does not modify the input data. In the backpropagation, the gradient inversion layer multiplies the transmitted gradient by a negative constant, adding an optimization direction to the feature extractor that is opposite to that of the domain discriminator.

[0030] D grl (x) = x;

[0031]

[0032] D grl This represents the forward propagation process of the gradient inversion layer. The backpropagation process of the gradient inversion layer is represented by I, where I represents the identity matrix and λ² represents the gradient inversion layer constant, which is not updated through backpropagation. Based on this, optimization problems in two different directions can be handled simultaneously within a single lithology prediction model, avoiding the need to train multiple lithology prediction models separately.

[0033]

[0034] Where ι represents the learning rate during model training.

[0035] Furthermore, the feature extractor constructed using a convolutional neural network with a fused channel attention mechanism specifically includes:

[0036] Integrating the channel attention module into the convolutional neural network yields a convolutional neural network with a fused channel attention mechanism, which is the feature extractor. The channel attention module combines a shared network with global average pooling and global max pooling mechanisms.

[0037] Inputting a 1D image sample x into a convolutional neural network F′ yields an intermediate feature map f:

[0038] f = F′(x)

[0039] By performing global average pooling and global max pooling operations on the intermediate feature map f, the average spatial representation f is generated respectively. GAP and the maximum space representation f GMP ; then, f GAP and f GMP After concatenation, the data is input into a 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 parameters of the shared network in the channel attention module; m will be used for the intermediate feature map f. ca Perform channel scaling calculation:

[0042] f′=m ca ⊙f;

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

[0044] Furthermore, the intra-domain confrontation specifically includes:

[0045] For each lithology category, a class-level intra-domain discriminant is designed, resulting in M ​​intra-domain discriminants. m = 1, 2, ..., M, where M is the total number of lithological categories, and the domain discriminant for each category level. Responsible for matching labeled and unlabeled logging data of the target domain associated with the m-th lithology category;

[0046] The label classifier C applies a label to each 1D image sample x. i The output p i =C(x) i ), representing x i The probability distribution of belonging to M lithological categories; x i For each domain discriminator Attention can be expressed as probability Weighted features

[0047]

[0048] This represents the loss of the discriminator within the domain. This represents the loss of the class m-domain discriminator, i.e., the log loss function. θ represents the probability that the label classifier predicts the i-th 1D image sample as class m. d·intra d represents all parameters of the in-domain discriminator. i The binary label θ represents the i-th 1D image sample, used to indicate whether this 1D image sample belongs to the source domain or the target domain; f These are the feature mapping parameters for the feature extractor F;

[0049] Intra-domain confrontation losses The definition is as follows:

[0050]

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

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

[0053] First, a global threshold is estimated using the exponential moving average of the prediction confidence of unlabeled samples in the target domain through a lithology prediction model. Then, the global threshold is adjusted by the exponential moving average of the probability of each lithology category using the lithology prediction model estimated by exponential moving average. A category-specific threshold is set for each lithology category, enabling automatic adjustment of the threshold based on the characteristics of different lithology categories and the prediction confidence of the label classifier C. This results in the learning loss for the unlabeled logging data of the target domain.

[0054]

[0055] Where B is the batch size, 1≤b≤B; This is an indicator function used to pass a global threshold. Screening should be done for unlabeled logging data that should be used for training; Represents cross-entropy loss, p represents the output probability distribution from the lithology prediction model. b The result is a hard label with only one category, p b It is the probability distribution output by the label classifier.

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

[0057] global threshold Initialized to Furthermore, the exponential moving average of the probability distribution confidence of the unlabeled logging data output in the target domain at the τth iteration of the lithology prediction model was used to evaluate the initialized global threshold. Update:

[0058]

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

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

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

[0062]

[0063] Ultimately, the dynamically adjusted category-specific threshold is obtained:

[0064]

[0065] in, represents the category-specific threshold for the m-th lithology category, and Norm represents the maximum normalization method.

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

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

[0068]

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

[0070] By probability density ratio Reflecting unbiased estimates across different domains, combined with Domain labels of the predicted well logging data samples Assign weights to the labeled logging data for each source domain:

[0071]

[0072] We calculate the weighted loss of the label classifier by weighting the contribution of labeled well logging data from the source domain to knowledge transfer. Discriminator-weighted loss

[0073]

[0074]

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

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

[0077] The data collection module uses labeled logging data from interpreted wells as source domain data and partially labeled and unlabeled logging data from target wells as target domain data. The probability distributions of the source domain data and the target domain data are different. Labeled logging data includes 1D image samples constructed from logging curve vectors and corresponding lithology category labels. Unlabeled logging data includes the 1D image samples. The logging curve vectors consist of measurement values ​​of different types of logging curves at a specific depth.

[0078] The model building module constructs a semi-supervised domain-adaptive lithology prediction model with dynamically adjusted thresholds. The lithology prediction model includes a feature extractor F and a domain discriminator D. inter Class-level intra-domain discriminants designed for each lithology category Independent discriminator The lithology prediction model is trained using a label classifier C and inter-domain adversarial, intra-domain adversarial, dynamic threshold adjustment, and source domain data reweighting.

[0079] A feature extractor is built using a convolutional neural network that incorporates a channel attention mechanism. The inter-domain adversarial mechanism uses the feature extractor to map logging data from different well locations to a unified feature space to extract domain-invariant features, and constructs the loss function for the domain discriminator. Loss of the label classifier

[0080] The intra-domain adversarial mechanism reduces the distributional differences between labeled and unlabeled logging data within the target domain by fusing the confidence predictions of the labeled classifier with an intra-domain discriminator, and constructs the loss function of the intra-domain discriminator.

[0081] The dynamic threshold is defined and adjusted using the prediction results of the lithology prediction model, thereby optimizing the pseudo-label selection of unlabeled logging data and constructing a learning loss.

[0082] Source domain data reweighting: This involves assigning different weights to source domain data based on their contributions using a dynamic weighting mechanism. This quantitatively indicates the importance of each source domain data point in the knowledge transfer process. The weights are then used to reweight the data. and Update the data to obtain the weighted loss of the label classifier. Discriminator-weighted loss

[0083] The model training module is based on the overall loss. The overall loss is calculated when training the lithology prediction model. for:

[0084]

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

[0086] The system and method of the present invention correspond to each other, and the specific technical solutions applicable to the method are also applicable to the system.

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

[0088] This invention constructs a cross-domain shared feature extractor based on adversarial learning, mapping logging data from different well locations to the same latent feature space, extracting domain-invariant features while ensuring 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 with a fused channel attention mechanism is introduced as the backbone network to enhance the model's ability to capture local morphological features of logging curves. Simultaneously, by optimizing the selection of sample image sizes, effective extraction of vertical depth correlation information is ensured, compensating for potential information loss from discarding recurrent neural networks. 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. Addressing the differences in data distribution within the target domain, an intra-domain discriminator that fuses classifier prediction confidence is developed. This design fully utilizes the limited labeled samples of the target well, achieving more accurate class alignment. Finally, this invention adds a dynamic threshold adjustment module, which flexibly adjusts the utilization of samples based on the prediction confidence level, improving the utilization rate of high-quality label predictions while reducing the negative impact of low-confidence samples, effectively reducing model confirmation bias. Particularly for the classifier, as high-confidence pseudo-label samples are continuously added to the training, the problem of model bias towards the source domain is continuously improved. Experimental results on the actual drilling dataset of Shengli Oilfield strongly confirm the superior performance of the proposed method. This method achieves significant alignment in the feature space, effectively reducing the data distribution differences between different well locations, and significantly improving the identification accuracy of each lithology category. It is particularly worth mentioning that the category-differentiated dynamic threshold adjustment mechanism enables the model to maintain relatively stable and accurate classification performance even when facing the problem of uneven reservoir lithology distribution. Attached Figure Description

[0089] Figure 1 To account for the differences in the distribution of logging data at different well locations;

[0090] Figure 2 This is a framework diagram of the semi-supervised domain adaptive lithology prediction model with dynamic threshold adjustment according to the present invention.

[0091] Figure 3 This is a schematic diagram illustrating the principle of the channel attention mechanism. Detailed Implementation

[0092] A preferred embodiment of the present invention will now be described in detail with reference to the accompanying drawings.

[0093] This invention treats logging data from the same well as a single "domain," with labeled logging data from interpreted wells considered the "source domain" for training the model and completing the learning task. Unlabeled logging data from the target well is considered the "target domain." This invention proposes a semi-supervised domain-adaptive lithology prediction method based on dynamic threshold adjustment, aiming to solve the problem of logging reservoir lithology prediction in scenarios where only a very limited number of lithology category labels are available for the target well. Specifically, this invention introduces an adversarial approach, utilizing a shared feature extractor between domains to map logging data from different well locations to a latent feature space, learning domain-invariant features of the logging data, and continuously reducing inter-domain differences while ensuring class distinguishability. Simultaneously, to enhance the feature extractor's extraction and focus on the temporal and morphological features of logging curves, a channel attention mechanism is fused to emphasize key local features. It is worth noting that recurrent neural networks require a considerable number of labeled samples for effective supervised learning. However, in scenarios with limited labels, it is difficult to fully learn the time-series patterns and is prone to overfitting, resulting in poor model generalization performance. Based on this, the feature extractor adopts a convolutional neural network with an attention mechanism. In addition to capturing the local morphological features of the logging curve, it further realizes the extraction of vertical depth correlation information through subsequent network design (optimizing the selection of sample image size), thereby compensating for the impact of modifying the recurrent neural network.

[0094] Considering the low correlation between some source domain data and target domain data, making it difficult to extract transferable information from them, this invention introduces a dynamic weighting mechanism. Weights are assigned based on the potential contribution of source domain data to knowledge transfer in the target domain, optimizing the quality of knowledge transfer. Since only a very small amount of labeled data is available in the target domain, an inevitable distribution bias exists between labeled logging data and a large amount of unlabeled logging data within the target domain—i.e., intra-domain discriminant. Simply adding labeled logging data for supervised learning cannot effectively alleviate this bias. Therefore, an intra-domain adversarial learning strategy is adopted, fully utilizing both the small amount of labeled logging data and the large amount of unlabeled logging data. An intra-domain discriminator is designed to integrate the predicted confidence of the label classifier, achieving class alignment within the target domain. This also allows a small number of labeled samples to better guide reservoir lithology model construction. Finally, a dynamic threshold adjustment mechanism is introduced to adjust the sample influence based on the predicted confidence, reducing confirmation bias. The method proposed in this invention has undergone extensive experimental verification in actual oilfield drilling data, achieving excellent results and fully demonstrating its accuracy and applicability in scenarios with limited label samples in the target well.

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

[0096] Suppose there are v logging curves of different types (including porosity curves). The measurement values ​​of each logging curve at a specific depth constitute a v-dimensional logging curve vector O = [o1, o2, ..., o...]. v ]∈R v The lithology category label corresponding to the logging curve vector O is obtained by analyzing core samples or rock slices at that depth. Based on this, a 1D image sample x = [O1, O2, ..., O] is constructed. h ]∈R h×v ; Corresponding to Lithology category labels at depth The source domain dataset is typically collected from a single well or multiple wells in the same area, while the target domain dataset comes from a geographically adjacent well, thus establishing a geographical connection between the two datasets. The source domain dataset is... Where n s yes The amount of data in the source domain, yes The i-th 1D image sample in the image. yes The corresponding lithology category label.

[0097] The target domain dataset is It contains a small amount of labeled logging data and a large amount of unlabeled logging data. Specifically, the labeled dataset... in n tl for The number of tagged logging data in the data. for The i2th 1D image sample in the middle, for The corresponding lithological category label; where n tl <<n s n tl <<n tul Unlabeled dataset in n tul express The amount of unlabeled logging data.

[0098] Source domain data and target domain data should be sampled from different marginal probability distributions, assuming that the source domain data follows a probability distribution. The target domain data follows a probability distribution but The objective of this invention is to leverage the rich annotation information of the source domain and combine it with a limited amount of labeled logging data from the target domain to construct a model capable of accurately classifying unlabeled logging data in the target domain. This method not only improves the model's prediction accuracy in the target domain but also effectively addresses the problem of scarce labeled data.

[0099] 2. Semi-supervised domain-adaptive lithology prediction model with dynamic threshold adjustment

[0100] This invention proposes a framework for a semi-supervised domain-adaptive lithology prediction model with dynamic threshold adjustment. This framework has excellent adaptive feature aggregation capabilities and can reduce the influence of low-confidence unlabeled samples according to the model's learning state, thereby achieving accurate identification of the target well's lithology category.

[0101] 2.1 Overall Network Architecture

[0102] The construction of a cross-well semi-supervised lithology prediction model in scenarios with scarce target well labels focuses on how to fully utilize the limited label samples from target wells to guide the construction of the reservoir lithology prediction model and achieve efficient alignment between the source and target domains. The proposed lithology prediction model is as follows: Figure 2 As shown, it includes four parts: 1) inter-domain adversarial; 2) intra-domain adversarial; 3) dynamic threshold adjustment; and 4) source domain data reweighting.

[0103] First, a cross-domain shared feature extractor based on adversarial learning is constructed to map logging data from different well locations to a unified feature space, extracting domain-invariant features as much as possible and continuously reducing inter-domain differences while ensuring class discriminability. Considering the scarcity of labeled data in the target scene, a convolutional neural network (CNN) incorporating channel attention mechanism is designed. 2 The backbone network of the CA (Carrying Aspect Ratio) feature extractor enhances the model's ability to capture local morphological features of well logging curves. Simultaneously, by optimizing the selection of subsequent sample image sizes, it ensures the effective extraction of temporal correlation information along the vertical depth of the well logging curves, compensating for potential information loss from discarding recurrent neural networks. Secondly, to address the distributional differences between the limited number of labeled well logging data and the large number of unlabeled well logging data in the target domain, an intra-domain discriminator based on the predicted confidence of the classifier is developed. This aims to fully utilize the limited labeled samples from the target wells, reduce data distribution differences within the target domain, and thus achieve more accurate class alignment. Next, a dynamic threshold adjustment mechanism is designed to flexibly adjust sample utilization based on the predicted confidence. Especially for the labeled classifier, as high-confidence pseudo-label samples are continuously added to the training, the bias towards the source domain is continuously improved. Finally, 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 the source domain data, the importance of each source domain data in the knowledge transfer process is quantitatively indicated, thereby optimizing the model's update performance.

[0104] 2.2 Network Module Design

[0105] 1. Inter-domain confrontation

[0106] To eliminate the impact of data distribution differences across different well locations, transferable knowledge between the source and target domains is extracted, and adversarial learning is introduced to reduce domain shifts between different domains. The total loss function of inter-domain adversarial learning is utilized. Implement the feature extractor F and the domain discriminator D. inter Adversarial training is conducted between the target and source domain data. During this training, the source domain data is labeled with "0", and the target domain data is labeled with "1". In the learning process, the feature extractor F strives to improve the feature representation of the lithology identification task and reduce the loss of the label classifier C.

[0107]

[0108] Where F(x) i ;θ f The feature representation generated by the feature extractor aims to generate features with domain invariance to confuse the domain discriminator, seeking the feature mapping parameter θ that maximizes the loss of the domain discriminator. f The generated two domain feature distributions are made as similar as possible; the domain discriminator then optimizes the selection of parameters θ to minimize the domain classification loss. d·inter F and D inter The adversarial training between them aims to generate a universal feature representation across domains and maximize D. inter The loss, and D inter It strives to avoid confusing different domains and minimizes the loss of the domain discriminator.

[0109]

[0110] In summary, the total loss function for inter-domain adversarial competition for:

[0111]

[0112] The adversarial training process is accomplished by optimizing the following mini-maximum game:

[0113]

[0114] The gradient reversal layer (GRL) is used to implement gradient backpropagation during model training. During forward propagation, the GRL does not modify the input data, acting as an identity function. During backpropagation, this layer multiplies the incoming gradient by a negative constant (usually -1), adding an optimization direction opposite to that of the domain discriminator to the feature extractor.

[0115] Dgrl (x) = x;

[0116]

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

[0118]

[0119] Where ι represents the learning rate during model training.

[0120] Well logging curves are interconnected in the vertical depth dimension, and local morphological information is crucial for subsurface observation. However, with limited labeled data, recurrent neural networks struggle to fully capture the inherent patterns of time series data, and the models are prone to overfitting, failing to generalize well to unseen data. Therefore, this invention utilizes convolutional neural networks to generate image feature representations with depth receptive fields. Furthermore, different well logging curves exhibit significant differences in sensitivity 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 consideration of the importance of different feature maps. Therefore, a convolutional neural network (CG) incorporating channel attention mechanisms is used... 2 The feature extractor F is built around the backbone structure using a channel attention module (CA), focusing on feature channels that are more important to the current task, thereby improving the network's sensitivity and response to key feature information. 2 Integrating CA into the feature extractor F architecture yields a convolutional neural network (CG) with a fused channel attention mechanism. 2 CA). Channel Attention Module (G) 2 CA effectively combines the global average pooling (GAP) and global max pooling (GMP) mechanisms, such as Figure 3 As shown, Global Average Pooling (GAP) extracts the average representation of all features within a channel by calculating the average of all elements in each channel of the feature map. This effectively identifies the overall structure and preserves the average information of the entire feature map. Global Max Pooling (GMP) extracts the maximum value in each channel, focusing on the most prominent features in that channel, making it more effective for capturing salient features in the image. The results of GAP and GMP are concatenated and further fused to generate weighted representations of the feature channels, identifying and emphasizing those feature channels most important to the current task.

[0121] Given a sample x∈R H×W and intermediate feature graph f=F(x)∈R H×W×CF represents a CNN network, where layers generate two distinct 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 compute the channel attention graph m. ca ∈R C×1×1 :

[0122] m ca =σ(W(f) GAP +f GMP ));

[0123] The original intermediate feature map f will use m ca Perform channel scaling calculation:

[0124] f′=m ca ⊙f;

[0125] Here, the symbol ⊙ represents the Hadamard product, and W represents the channel attention module (G). 2 The weight parameters of the shared network in CA). Convolutional Neural Networks (CG) with integrated channel attention mechanism. 2 CA (Carrier Algorithm) technology effectively modulates the suppression or enhancement of specific feature channels. By weighting the differences of each channel, it adaptively recalibrates the feature mapping for each channel, enhancing the model's ability to prioritize the representation of task-related features.

[0126] 2. Intra-domain confrontation

[0127] In the aforementioned inter-domain adversarial learning process, since the label classifier C is trained on source domain data, its decisions are biased towards the features of the source domain data. However, in practical applications, a single well involves depths spanning multiple formations and different sedimentary environments, resulting in distribution biases in samples from distant locations. Because only a very small amount of labeled logging data is available in the target domain, an unavoidable distribution bias exists between labeled logging data and the large amount of unlabeled logging data within the target domain—i.e., intra-domain discrepancies. Simply adding labeled logging data for supervised learning cannot effectively alleviate the distribution discrepancies of unlabeled logging data within the target domain. Therefore, intra-domain adversarial learning, such as... Figure 2 As shown, a more refined multi-modal alignment method is used to reduce the probability distribution differences between labeled and unlabeled logging data in the target domain. Specifically, a class-level intra-domain discriminator is designed for each category, i.e. m = 1, 2, ..., M, where M is the total number of lithological categories, and the domain discriminant is at the category level. It is responsible for specifically matching labeled and unlabeled logging data of the target domain associated with the m-th lithology category.

[0128] Since the lithological category of unlabeled data is completely unknown for the purpose of the study, it is impossible to directly determine which category-level discriminator should be used for unlabeled logging data in a target domain. Processing. It is worth noting that the label classifier C outputs p for each logging data x. i =C(x) i The probability distribution p represents the probability that the logging data belongs to one of M lithological categories. i This can be used to represent the logging data x i Which class domain discriminators should we focus on? And the degree of attention; specifically, x i For each domain discriminator Attention can be expressed as probability Weighted features

[0129]

[0130] In this way, even if the target domain data lacks labeling information, the soft labels of the label classifier can still effectively assign the target domain data to the class discriminator based on predicted probabilities, thereby achieving intra-domain distribution matching of the target domain. The intra-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 class by utilizing model prediction results, thereby optimizing the selection of false labels for unlabeled samples and reducing model confirmation bias. Assume the threshold for class m in the τ-th iteration is expressed as... First, a global threshold is estimated using the exponential moving average (EMA) of the confidence level, serving as an indicator of the overall model's confidence. Further, the global threshold is adjusted by estimating the EMA of the probability of each class in the model, setting a locally specific threshold for each class. In this way, the invention can automatically adjust the thresholds for different classes based on their characteristics and the prediction confidence of the label classifier C for those classes. Initially, the threshold is set low during model training, allowing more potentially correct samples to participate in training. As the model becomes more confident, the threshold is dynamically adjusted and increased to filter out samples with low confidence, thus reducing the inclusion of erroneous sample labels in the calculation.

[0134] Global Threshold: Here, the global threshold reflects the model's confidence in unlabeled samples in the target domain. This means that the threshold should be adjusted based on the model's prediction confidence in unlabeled samples in the target domain to better evaluate the overall training performance of the model. During training, the global threshold is gradually increased to ensure that low-confidence pseudo-labeled samples are filtered out, thereby improving the accuracy and reliability of learning. Specifically, the global threshold... This is set as the average confidence level of the unlabeled logging data in the target domain during the τ-th iteration of the model. Since calculating the confidence level of all unlabeled data in each iteration is very time-consuming, the exponential moving average (EMA) is used to estimate the global confidence level. Initialized to A lower initial threshold ensures that the model can learn from enough samples in the early stages of training, as shown below:

[0135]

[0136] Where B is the batch size, 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] Category-specific thresholds. To better handle characteristics within each category and relationships between different categories, different thresholds are set for different categories, rather than using a uniform global threshold. The expected value of the model's predictions for each category is calculated. By estimating the model's learning state for each category, we can gain a more accurate understanding of the model's performance across different categories, thus enabling more appropriate adjustments.

[0138]

[0139] Ultimately, the following dynamically adjusted category-specific thresholds are obtained:

[0140]

[0141] Where Norm represents the maximum normalization processing method, and based on the above, the learning loss for unlabeled logging data in the target domain is expressed as:

[0142]

[0143] Indicator function Thresholding based on confidence levels is used to determine which unlabeled logging data samples should be used for training, and is derived from the model's output probability distribution p. b The conversion resulted in a single hard label for one category. This represents the cross-entropy loss.

[0144] 4. Source domain data reweighting

[0145] Source domain data reweighting aims to reduce the impact of source domain data with low relevance to the target domain on the transfer effect. Starting with the features or attributes of the original samples, it can be found that the differences in feature distribution between different samples can reflect the degree of correlation between samples. In other words, by analyzing the differences in the distribution of sample features, the relationship or similarity between source domain data and target domain data can be understood and evaluated. Therefore, a reweighting of source domain data is designed during the learning process to further enhance the extraction of domain-invariant features and promote the transfer effect of knowledge from the source domain to the target domain. Specifically, an independent domain discriminator is first pre-trained. Based on its ability to distinguish samples from different domains, it measures the similarity between source domain data and target domain data during model training. Since the joint feature distribution can more comprehensively reflect the relationship between sample features and labels, it better captures the differences between source and target domain data. Independent Discriminator The original features X and corresponding labels Y of the sample data are used as input to determine which domain the sample belongs to.

[0146]

[0147] The output of this independent domain discriminator is also 0 for the source domain and 1 for the target domain. and These indicate the joint distribution of data in the source and target domains, respectively.

[0148] Next, the pre-trained independent domain discriminator is used to quantify the distributional differences between samples in the source and target domains, representing the degree of difference between the samples in the feature space. This is taken into account the probability density ratio. It can reflect unbiased estimates between different domains, combined with Predicted sample domain labels These quantified distribution differences are transformed, and weights are assigned to each source domain sample based on the density ratio estimation results:

[0149]

[0150] By updating the model's objective loss function and By incorporating these weights into the transfer learning process, the contribution of source domain samples to knowledge transfer will be weighted according to the weights. This ensures that source domain samples that are more similar to target domain samples and are more helpful in improving model performance receive more attention and importance during training, achieving more effective sample selection and knowledge transfer.

[0151]

[0152] 5. Overall Workflow

[0153] In summary, the overall optimization objective of the semi-supervised domain-adaptive 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 trade-off role in each loss term.

[0156] The following summarizes the process of predicting reservoir lithology using the lithology prediction model proposed in this invention.

[0157] Input: Source domain dataset from interpreted wells Labeled dataset of target wells and unlabeled datasets

[0158] Output: Lithology category prediction results for unlabeled data from the target well

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

[0160] (2): Pre-train and save an independent domain discriminator

[0161] (3): Input the well 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): D s D tl D tul Corresponding inputs to each module of the lithology prediction model;

[0163] (5): Update using weight w and

[0164] (6): By minimizing the loss function Optimize the lithology prediction model until the preset iteration cycle is reached;

[0165] (7): Adjust hyperparameters using a validation set that was not used in training and save the best-performing lithology prediction model.

[0166] Test: Predict the lithology category of unlabeled logging data for the target well using a saved lithology prediction model.

[0167] In a preferred embodiment, the logging curve vector uses six logging curves and a porosity parameter curve, namely: 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 techniques that effectively distinguish between shale and sandstone. The SP curve also reflects the type of fluid in the pores, such as oil or water, with water-bearing sands typically exhibiting higher SP amplitudes than oil-bearing sands. AC is acoustic logging, reflecting reservoir characteristics such as lithology, porosity, and fluids. CNL and GR are radiometric logging techniques, sensitive to specific rock types and fluids. GR is particularly sensitive to highly radiometric 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 suggest the presence of mudstone (MS) or coal seams due to their relatively low hardness.

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

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

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

[0172] The combination of these logging curves and porosity curves has sufficient distinguishability (i.e., the eigenvectors are discriminative) to enable effective prediction of model lithology.

[0173] In a preferred embodiment, lithological category labels are obtained by analyzing core samples or cuttings collected from the borehole. The lithological categories to be predicted include any one or more of mudstone (MS), sandstone (SS), oil shale (Os), dolomite (DM), siltstone (SI), and fine sandstone (FS). Accurate prediction of these lithological categories is crucial for reservoir modeling in the target well.

[0174] In a preferred embodiment, for each logging curve, an image sample is constructed in the depth direction. Taking a certain depth point as the center, 16 points are taken upwards and downwards. For a well, the center point is taken repeatedly in the depth direction. The image sample label corresponds to the core logging at the center depth point.

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

[0176] It is worth noting that it is difficult to ensure that logging data collected from different wells are obtained using the same type of instrument, the same standard scale, and the same operating methods. The different orders of magnitude of the logging curves can affect the model's judgment of feature importance. Therefore, a normalization method is needed to eliminate systematic errors caused by factors such as instrument performance, calibration, and personnel, as well as the impact of data size 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 three-layer convolutional neural network (CGN) with a fused channel attention mechanism. 2 The system consists of a CA (Carbon Classifier), a label classifier, a domain discriminator, and a source domain data reweighting module, each composed of a two-layer fully connected network. The intra-domain discriminators are configured according to lithology categories, and each category's intra-domain discriminator also consists of a two-layer fully connected network. The Avg-Pool and Max-Pool in the feature extractor use adaptive computation. To further mitigate the impact of class imbalance, the weights of each class's samples on the supervised loss function are adjusted based on the sample size of each class. m The definition is as follows:

[0178]

[0179] Where, N m N represents the number of samples of a specific lithology. totalThis represents the total number of training samples to ensure that the minority class receives sufficient training. The model is trained using an Adam-based optimizer with a learning rate of 0.01 and a weight decay factor of 0.00005. Annealing is used to adjust the learning rate according to the model's training progress, ensuring that the model learns quickly in the early stages of training and converges more stably in the later stages.

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

[0181] This invention, based on high-precision prediction of reservoir porosity parameters from well logging, proposes a semi-supervised domain adaptation-based well logging reservoir lithology prediction model for general scenarios where lithology is known prior. This model effectively utilizes limited target domain label data to guide the model construction process, thereby achieving accurate prediction of the complete lithology sequence of the target well. This provides a broader lithology category labeling basis for constructing a comprehensive source domain within the work area.

[0182] The system and method of the present invention correspond to each other, and the specific technical solutions applicable to the method are also applicable to the system.

[0183] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0184] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider 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 semi-supervised domain adaptation lithology model construction method, characterized in that, The lithology of the reservoir is identified by completing a trained lithology prediction model, and the construction and training process of the lithology prediction model comprises the following steps: Step one, the labeled well data of the interpreted well is taken as the source domain data, and part of the labeled well data and the unlabeled well data of the target well are taken as the target domain data; the probability distribution of the source domain data and the target domain data is different; the labeled well data includes 1D image samples constructed by well logging curve vectors and corresponding lithology category labels; the unlabeled well data includes the 1D image samples; the well logging curve vector is composed of measurement values of different types of well logging curves at a certain depth; Step two, constructing a semi-supervised domain adaptation lithology prediction model with threshold dynamic adjustment, the lithology prediction model comprising a feature extractor F, a domain discriminator , a class-level intra-domain discriminator designed for each lithology category , an independent discriminator , and a label classifier C; training the lithology prediction model through inter-domain confrontation, intra-domain confrontation, dynamic threshold adjustment, and source domain data reweighting; The feature extractor is built by a convolutional neural network fused with a channel attention mechanism, the domain-invariant features are extracted by mapping the logging data of different well sites to a unified feature space through the feature extractor, and the loss of the domain discriminator is constructed and the loss of the label classifier ​ The in-domain confrontation reduces the distribution difference of the labeled well logging data and the unlabeled well logging data in the target domain through an in-domain discriminator that fuses the label classifier prediction confidence, and constructs a loss of the in-domain discriminator ; The dynamic threshold defines and adjusts the confidence threshold of each lithology category by using the prediction result of the lithology prediction model, so as to optimize the pseudo-label selection of the unlabeled logging data and construct the learning loss ; Source domain data reweighting: This involves assigning different weights to source domain data based on their contributions using a dynamic weighting mechanism. This quantitatively indicates the importance of each source domain data point in the knowledge transfer process. The weights are then used to further refine the data. and Update the data to obtain the weighted loss of the label classifier. Discriminator-weighted loss Specifically, this includes: pre-training an independent domain discriminator. The original characteristics of well logging data samples and corresponding tags Simultaneously input independent discriminator To predict 1D image samples domain tags : ; and These represent the joint distribution of the source domain data and the joint distribution of the target domain data, respectively. Let them represent the source domain feature space and the source domain label space, respectively. Represent the target domain feature space and the target domain label space respectively; by probability density ratio Reflecting unbiased estimates across different domains, combined with Domain labels of the predicted well logging data samples Assign weights to labeled logging data for each source domain. The weighted loss of the label classifier is calculated by weighting the contribution of labeled well logging data from the source domain to knowledge transfer. Discriminator-weighted loss ; ; is a source domain dataset number of source domain data in denotes cross-entropy loss; is a labeled dataset of target domain; ； represents the number of well log data in the target domain dataset; Overall loss of lithology prediction model is: ； where the hyperparameters , , and play a trade-off role in each loss term.

2. The method of claim 1, wherein, Step one specifically includes: The measured values of the different types of well logging curves at a certain depth form a well logging curve vector of dimension ; For the i-th well logging curve in the well logging curve vector, the corresponding lithology class label is obtained by analyzing the core or rock slice at the certain depth; on this basis, a 1D image sample is constructed; the corresponding lithology class label is , the number of well logging curve vectors in the well logging curve vector , 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 adjacent to the interpreted well in geography; the source domain data set , wherein is the number of source domain data in the source domain data , is the i-th 1D image sample in the source domain data , is the corresponding lithology class label; the target domain data set , the labeled data set of the target domain , wherein is the number of labeled well logging data in the labeled data set of the target domain , is the i-th 1D image sample in the labeled data set of the target domain , is the corresponding lithology class label; the unlabeled data set of the target domain , wherein represents the number of unlabeled well logging data in the unlabeled data set of the target domain , is the i-th 1D image sample in the unlabeled data set of the target domain; wherein , , , .

3. The method of claim 2, wherein, The inter-domain confrontation specifically includes: To eliminate the data distribution discrepancy between interpreted wells and target wells, adversarial learning is introduced to reduce the domain offset between different source and target domains, utilizing the total loss function of inter-domain adversarial learning. Implement the feature extractor F and the domain discriminator. Adversarial training between them; feature extractor F and domain discriminator During adversarial training, the domain labels of the source domain data are marked as "0", and the domain labels of the target domain data are marked as "1". The feature extractor F strives to generate feature representations for the lithology identification task and reduce the loss of the label classifier C. ； wherein, is the loss of the tag classifier C, denotes the cross-entropy loss, denotes the i-th input 1D image sample, is the corresponding lithology class label; is the feature representation generated by the feature extractor; meanwhile, the feature extractor generates features with domain invariance as much as possible to confuse the domain discriminator, and seeks the feature mapping parameters that maximize the domain discriminator loss , so that the two generated domain feature distributions are as similar as possible; the domain discriminator then reduces the parameters of the domain classification loss as much as possible ; and The purpose of the adversarial training between and is: The goal of is to generate domain-universal feature representations, maximizing the loss of , while tries to avoid confusion between different domains and minimize the loss of the domain discriminator : ； a binary label representing the ith 1D image sample, indicating whether this 1D image sample belongs to the source domain or to the target domain, denotes a log loss function; In summary, the total loss function of inter-domain adversarial is: ； is a weight coefficient, denotes the label classifier parameters.

4. The method of claim 3, wherein, The process of the inter-domain confrontation is completed by optimizing the following minimax game: ； ； respectively denote the optimal parameters of the feature extractor, the label classifier and the inter-domain discriminator; Through gradient back propagation in the gradient inversion layer training process, the gradient inversion layer does not change the input data in the forward propagation process, and the gradient inversion layer will multiply the transmitted gradient by a negative constant in the back propagation process, so as to add an opposite optimization direction to the feature extractor: ； ； forward propagation process of the gradient reversal layer, backward propagation process of the gradient reversal layer, identity matrix, gradient reversal layer constant, which will not be updated through backpropagation; based on this, two optimization problems in different directions can be processed simultaneously in a single lithology prediction model, avoiding the need to train multiple lithology prediction models separately: ； ； ； wherein, represents the size of the learning rate during the model training process.

5. The method of claim 1 or 2, wherein, The convolutional neural network with fused channel attention mechanism is used to build the feature extractor, which specifically includes: The channel attention module is integrated into the convolutional neural network to obtain the convolutional neural network with fused channel attention mechanism, that is, the feature extractor; the channel attention module combines the shared network, the global average pooling mechanism and the global maximum pooling mechanism; The 1D image samples are input to a convolutional neural network to obtain intermediate feature maps : By fusing the intermediate feature maps The global average pooling operation and the global max pooling operation are performed to generate an average spatial representation and a maximum spatial representation respectively; then, the average spatial representation and the maximum spatial representation are concatenated and input into a shared network to calculate a channel attention map : ； denotes the Relu activation function; W represents the weight parameter of the shared network in the channel attention module; and will be used Channel scaling calculation is performed: ； wherein, is the scaled intermediate feature map for the channel, and the notation denotes the Hadamard product.

6. The method of claim 2, wherein, The intra-domain confrontation specifically includes: An intra-domain discriminator of a class level is designed for each lithology category, and M intra-domain discriminators are obtained , is the total number of lithology categories, and the intra-domain discriminator of the class level is responsible for matching the labeled well logging data and unlabeled well logging data of the target domain related to the first lithology category; Label classifier C for each 1D image sample Output ,express belong Probability distribution of each lithological category; For each domain discriminator Attention can be represented by probability Weighted features : ； This represents the loss of the discriminator within the domain. Indicates category The loss of the domain discriminator, i.e., the logarithmic loss function, This indicates that the label classifier predicts the i-th 1D image sample as a category. The probability, This represents all parameters of the in-domain discriminator. The binary label represents the i-th 1D image sample, used to indicate whether this 1D image sample belongs to the source domain or the target domain; For feature extractor Feature mapping parameters; Intra-domain adversarial loss is defined as follows: ； denote parameters of the feature extractor, the label classifier, and the inter-domain discriminator, respectively; denotes a cross-entropy loss; denote parameters of the label classifier; denotes a weight coefficient.

7. The method of claim 1, wherein, The dynamic threshold adjustment specifically includes: First, a global threshold is estimated by exponentially moving average of the prediction confidence of the target domain unlabeled samples through the lithology prediction model Then, the global threshold is adjusted by exponentially moving average of the lithology prediction model for each lithology class probability to set a class-specific threshold for each lithology class, so as to automatically adjust the respective threshold according to the characteristics of different lithology classes and the prediction confidence of the lithology class by the label classifier C; and further obtain the learning loss of the unlabeled logging data of the target domain : ； wherein, is a batch size, ; is an indicator function that filters unlabeled log data that should be used for training by a global threshold ; represents a cross-entropy loss, denotes a hard label of only one class converted from the output probability distribution of the lithology prediction model ; is a probability distribution output by the label classifier.

8. The method of claim 7, wherein, The estimation process of the global threshold value includes: global threshold initialized to and updated by an exponential moving average of the confidence of the probability distribution output by the lithology prediction model on the unlabeled well log data of the target domain on the ith iteration: initialized global threshold ； where B is the batch size, is the momentum decay factor for the exponential moving average; The setting process of the category-specific threshold value includes: the expectation of the lithology prediction model for the predicted value of each lithology class , : ； Finally, the dynamically adjusted category-specific threshold value is obtained: ； wherein, represents the class-specific threshold value for the mth lithology class, and Norm represents the maximum normalization method. 9.A semi-supervised domain adaptation lithology model construction system, characterized in that, Including: The data collection module takes the labeled well data of the interpreted well as the source domain data, and part of the labeled well data and the unlabeled well data of the target well as the target domain data; the probability distribution of the source domain data and the target domain data is different; the labeled well data includes 1D image samples constructed by well logging curve vectors and corresponding lithology category labels; the unlabeled well data includes the 1D image samples; the well logging curve vector is composed of measurement values of different types of well logging curves at a certain depth; A model construction module constructs a semi-supervised domain adaptation lithology prediction model with dynamic threshold adjustment, the lithology prediction model comprising a feature extractor F, a domain discriminator , a class-level intra-domain discriminator designed for each lithology category , an independent discriminator , and a label classifier C; the lithology prediction model is trained through inter-domain confrontation, intra-domain confrontation, dynamic threshold adjustment, and re-weighting of source domain data; A feature extractor is built by a convolutional neural network fused with a channel attention mechanism, the inter-domain adversarial learning extracts domain-invariant features by mapping the well logging data of different well locations to a unified feature space through the feature extractor, and a loss of a domain discriminator is constructed and a label classifier ​ The in-domain adversarial passes through an in-domain discriminator that predicts the confidence of the fusion label classifier, reduces the distribution difference of the labeled well logging data and the unlabeled well logging data in the target domain, and constructs a loss of the in-domain discriminator ; The dynamic threshold defines and adjusts the confidence threshold of each lithology category by using the prediction result of the lithology prediction model, so as to optimize the pseudo-label selection of the unlabeled logging data and construct the learning loss ; Source domain data reweighting: This involves assigning different weights to source domain data based on their contributions using a dynamic weighting mechanism. This quantitatively indicates the importance of each source domain data point in the knowledge transfer process. The weights are then used to further refine the data. and Update the data to obtain the weighted loss of the label classifier. Discriminator-weighted loss Specifically, this includes: pre-training an independent domain discriminator. The original characteristics of well logging data samples and corresponding tags Simultaneously input independent discriminator To predict 1D image samples domain tags : ; and These represent the joint distribution of the source domain data and the joint distribution of the target domain data, respectively. Let them represent the source domain feature space and the source domain label space, respectively. Represent the target domain feature space and the target domain label space respectively; by probability density ratio Reflecting unbiased estimates across different domains, combined with Domain labels of the predicted well logging data samples Assign weights to labeled logging data for each source domain. The weighted loss of the label classifier is calculated by weighting the contribution of labeled well logging data from the source domain to knowledge transfer. Discriminator-weighted loss ; ; is a source domain dataset number of source domain data in denotes cross-entropy loss; is a labeled dataset of target domain; ； represents the number of well log data in the target domain dataset; a model training module, based on an overall loss training the lithology prediction model, the overall loss is: ； where the hyperparameters , , and play a trade-off role in each loss term.

Images