An Unsupervised Adaptive Method for Lithology Identification Models in Different Regions

Abstract

This invention discloses an unsupervised adaptive method for lithology identification models in different regions. First, feature extraction and dimensionality reduction are performed on rock images with lithology labels in the source domain and unlabeled rock images in the target domain, resulting in an initial projection matrix. A class-prototype-based and structured prediction strategy is used to predict lithology pseudo-labels for rock samples in the target domain, while reinforcement learning is used to selectively pseudo-label rock samples with lithology pseudo-labels in the target domain. Then, the projection matrix is ​​updated by using a locally preserved projection of the lithology-labeled source domain data and the lithology pseudo-labeled target domain data selected by reinforcement learning, and the lithology pseudo-labels for all data in the target domain are updated using the class-prototype-based and structured prediction strategy. Finally, by updating the projection, the domain adaptation of the lithology identification model for different regions is achieved. This method improves the accuracy of pseudo-labeling and domain alignment for unlabeled data.

Description

Technical Field

[0001] This invention relates to lithology identification models, and more particularly to an unsupervised adaptive method for lithology identification models in different regions. Background Technology

[0002] With the application of artificial intelligence in lithology identification, the domain adaptation problem exists in many practical applications and has attracted much attention as a current research hotspot and challenge. When it is necessary to identify the lithology of unlabeled samples in the target domain, but the labeled samples only come from the source domain, and the data distributions of the two domains are different, the performance of the classifier will significantly decrease if the classification model trained from the source domain is directly applied to the target domain samples.

[0003] Unsupervised domain adaptation (UDA) is an effective method for addressing the problem of unlabeled test samples and labeled training samples coming from different domains. It achieves transfer learning by reducing the distributional differences between the source and target domains. Mainstream unsupervised domain adaptation algorithms can be categorized into those based on sample importance estimation, higher-order moment matching, pseudo-labels, and adversarial training. Pseudo-label-based unsupervised domain adaptation algorithms attempt to assign pseudo-labels to each instance in the target domain and select trustworthy instances to achieve supervised training of the model.

[0004] Traditional pseudo-labeling methods struggle to balance the accuracy and representativeness of pseudo-labeled data. Pseudo-labels often contain many incorrect labels, and using incorrectly labeled samples for training can negatively impact the original model. Furthermore, inaccurate pseudo-labels can lead to catastrophic error accumulation during the learning process. Therefore, a selective pseudo-labeling strategy can be employed to select the most reliable pseudo-label samples, increasing the probability of correct pseudo-labeling in the target domain. This avoids including high-error-rate target domain samples in subsequent iterations, thereby improving the accuracy of pseudo-labels for target domain samples and enhancing adaptive robustness.

[0005] Local Preserving Projection (LPP), as a method of mapping to a subspace, has been experimentally proven to yield better results than Principal Component Analysis (PCA) and Fisher's Linear Discriminant Analysis (FSA) in many applications. Its main idea is to preserve the local structure of the original data by ensuring that pairs of points that are close together in the original space remain close in the reduced-dimensional space. This not only preserves the structure of the original data well but also achieves good results in clustering problems with multiple clusters within a class. By constructing an objective function about the projection matrix, two domains are projected onto the same subspace, and pseudo-labeled samples are selected within the subspace. The projection matrix is ​​continuously updated according to an iterative strategy, thereby reducing the constraints imposed by distributional differences and promoting class alignment between the two domains. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide an unsupervised adaptive method for lithological identification models of different regions that promotes accurate lithological pseudo-labeling through a selective pseudo-labeling strategy and improves domain alignment accuracy.

[0007] Technical solution: The unsupervised adaptive method of the present invention includes the following steps:

[0008] S1, after manually classifying the rock image dataset containing various types of lithological data, assigns lithological labels to it as the source domain dataset;

[0009] S2, use the saved core scan images and electron microscope scan images as the target domain dataset;

[0010] S3, based on the source domain dataset and the target domain dataset, after preprocessing the rock images, the feature extractor F in the CNN lithology identification model is introduced to extract features and reduce dimensions, and the initial projection P0 is obtained using the labeled rock samples in the source domain.

[0011] S4, In the learning subspace, a lithological pseudo-label is applied to all data in the target domain using a class prototype-based combined structured prediction method;

[0012] S5, a reinforcement learning-based lithology pseudo-label selection strategy is used to select the most reliable lithology pseudo-label samples in the target domain;

[0013] S6. Combine the lithological pseudo-label samples obtained in step S5 with the lithological label samples in the source domain into a sample set, and use local preservation projection as an enabling technique to complete the class alignment task of the two domains.

[0014] S7, using a prototype-based combined structured prediction method to update the target domain rock samples with lithological pseudo-labels;

[0015] S8. Repeat steps S5 to S7 until the target number of iterations is reached.

[0016] Furthermore, in step S1, according to lithological classification, coal, sandstone, coarse sandstone, medium sandstone, fine sandstone, siltstone, and mudstone are used as lithological labels; among them, sandstone particles have a diameter greater than 2 mm, coarse sandstone particles have a diameter of 1–0.5 mm, medium sandstone particles have a diameter of 0.5–0.25 mm, fine sandstone particles have a diameter of 0.25–0.05 mm, siltstone particles have a diameter of 0.05–0.005 mm, and mudstone particles have a diameter less than 0.005 mm; the color of coal is black or grayish-black.

[0017] Furthermore, in step S1, n in the source domain s Artificial lithological classification and labeling were performed on a rock image dataset containing various lithological data, serving as a labeled dataset for the source domain. in This represents the i-th labeled rock sample in the source domain. This indicates a manually labeled lithology tag.

[0018] Furthermore, in step S2, the n stored in the target domain is... t The target domain unlabeled dataset consists of 100 unlabeled core scan images and electron microscopy scan images. in This represents the i-th unlabeled rock sample in the target domain.

[0019] Furthermore, in step S3, the preprocessing steps for the rock image are as follows:

[0020] S31: Read rock images from the source and target domain datasets, select a 256×256 pixel region in the middle of the rock image and resize it to 224×224 pixels. After normalization and data augmentation, transmit it to the feature extractor F of the CNN rock identification model. Feature extractor F extracts features based on the color, shape, texture, and structure of the rocks in the image. and Feature vectors: and d represents the feature dimension;

[0021] S32, using local preservation projection to reduce the dimensionality of the feature vectors of rock samples in the source and target domains, the objective function expression is:

[0022]

[0023] Where P is the projection matrix sought by the objective function;

[0024] This is the matrix generated by concatenating the feature vectors of rock samples from the source and target domains.

[0025] D is a diagonal matrix The similarity matrix M is defined as:

[0026]

[0027] L = DM is the Laplace matrix;

[0028] The eigenvector corresponding to the smallest eigenvalue is obtained using the following formula:

[0029] XLX T p = λXDX T p

[0030] The eigenvector corresponding to the smallest eigenvalue is the optimal projection vector of the objective function;

[0031] Solving for the projection matrix yields the solution. and low-dimensional feature matrix In the formula d1≤d denotes the dimension of the feature space after dimensionality reduction, n=n s +n t ;

[0032] S33 is a matrix composed of low-dimensional feature vectors of labeled rock samples from the source domain. Substitute into the formula In the process, the initial projection matrix is ​​obtained. Where d2 represents the dimension of the projected subspace, and d2 < d1.

[0033] Furthermore, in step S4, the implementation steps for using a prototype-based combined with structured prediction method to perform lithological pseudo-labeling on all rock sample data in the target domain are as follows:

[0034] S41, Calculate the projection of the source region rock sample:

[0035]

[0036] S42, Calculate the projection of the target domain rock sample:

[0037]

[0038] S43, Projecting source domain rock samples z s Projection z of rock sample in the target domain t After centralization and L2 normalization, the average value z of the projections of all source domain rock samples and target domain rock samples is obtained;

[0039] S44, for rock sample x in the target domain t The expression for predicting lithological pseudo-labels is as follows:

[0040]

[0041] Where p(y|x) t )=max{p1(y|x t ),p2(y|x t )};

[0042]

[0043] Among them, |У 集合 | Indicates the number of rock categories; The class prototype representing lithology type y is the average vector of the projections of source domain rock samples labeled y, expressed as follows:

[0044]

[0045] When y = y i hour, otherwise,

[0046]

[0047] in, Indicates the cluster center corresponding to class y;

[0048] For rock sample x in the target domain t The formula for predicting lithological pseudo-labels is as follows:

[0049]

[0050] Where p(y|x) t )=max{p1(y|x t ),p2(y|x t )},У 集合 A set representing rock categories.

[0051] Combine the target domain rock sample set with lithological pseudo-labels to form a pseudo-label dataset:

[0052]

[0053] Furthermore, in step S5, the process of selecting the most reliable lithological pseudo-label samples in the target domain is a Markov decision process, including a state space, action space, and reward. The agent is trained to select lithological pseudo-label samples; the pseudo-label dataset D is defined. u n obtained from random sampling c The pseudo-marker lithological feature vector is composed of 10 pseudo-marker samples to be selected. and the positive set vector composed of selected lithological pseudo-marker samples Where the positive set vector Initialize to empty;

[0054] Define a state space S, the state space Each state s i Including pseudo-marker lithological eigenvectors D c , positive set vector And the target domain unlabeled lithological eigenvector D t Among them, the pseudo-marker lithological feature vector D c From rock examples low-dimensional feature vectors Composition; positive set vector From rock examples low-dimensional feature vectors Composition; Target domain without lithological marker vector D t Rock instances from the target domain low-dimensional feature vectors composition;

[0055] Define action space A, the action space Each action a i This indicates that the lithological eigenvector D from the current pseudo-marker... c Select a lithological pseudo-marker instance for each state s. i The agent takes action a i Choose D c Rock instances with reliable lithology and move them to In the middle, replace the selected rock instance with the zero vector;

[0056] Define a reward R, which reflects whether the selected lithological pseudo-marker instance is more accurate; set a metric function to measure representativeness and accuracy, as follows:

[0057]

[0058] Where: β is a hyperparameter; Lithological pseudo-label types predicted by combining prototype-based and structured prediction methods The probability represents the confidence level of the lithology prediction; Δ e The formula representing the reduction in entropy of unlabeled rock sample data in the target domain is:

[0059] Δ e =HH′

[0060] Where H is s i The entropy value of a state, H′ is the next state s. i+1 The entropy value under;

[0061] The formula for calculating entropy is:

[0062] Set the reward function as follows: Where: τ is the set threshold, and its value is (1+β)log(0.8);

[0063] For each state and action (s) i ,a i ), using the function Q(s i ,a i The cumulative reward for the discounted state and action is represented by ), and the iterative learning method for deep updating the Q-network is as follows:

[0064]

[0065] Where Ω represents the parameters of the Q-network; ε is the exploration factor; V i For Q(s) i ,a iThe target value is given by the formula:

[0066]

[0067] Where γ represents a discount factor that determines the importance of future cumulative rewards compared to current rewards, γ∈(0,1);

[0068] When the agent is invoked to select lithology pseudo-marker instances, the operation using the policy output is defined as a. t The formula is:

[0069]

[0070] Where: s t Indicates the current state.

[0071] Furthermore, in step S6, local preservative projection is used as an enabling technique for the lithology pseudo-labeled samples and the source domain rock samples with lithology labels to obtain the projection matrix and complete the class alignment task of the two domains. The specific implementation process is as follows:

[0072] The matrix consists of the low-dimensional feature vector set of lithologically labeled samples in the source domain and the set of pseudo-labeled lithological samples selected by reinforcement learning. Substitute into the formula In the process, the projection matrix P of the k-th iteration is obtained. k .

[0073] Compared with the prior art, the significant advantages of this invention are as follows:

[0074] 1. We utilize LPP-based subspace learning and two selective pseudo-labeling strategies for lithology pseudo-labeling, and improve the accuracy of lithology pseudo-labels through iteration;

[0075] 2. By combining a class prototype-based structured prediction pseudo-label method with a reinforcement learning-based lithology pseudo-label selection strategy, the accuracy of domain alignment is improved, thereby enhancing the performance of unsupervised domain adaptation. Attached Figure Description

[0076] Figure 1 This is a flowchart of the present invention;

[0077] Figure 2 This is a diagram illustrating the iterative process of learning the projection matrix in this invention.

[0078] Figure 3 This is a flowchart of the pseudo-labeling algorithm based on reinforcement learning in this invention. Detailed Implementation

[0079] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0080] The rock grain size and color characteristics of this invention are shown in Table 1, as detailed below:

[0081] The lithology is classified as follows: coal, conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone, and mudstone. Specifically, conglomerate has grain diameters greater than 2 mm, coarse sandstone has grain diameters of 1–0.5 mm, medium sandstone has grain diameters of 0.5–0.25 mm, fine sandstone has grain diameters of 0.25–0.05 mm, siltstone has grain diameters of 0.05–0.005 mm, and mudstone has grain diameters less than 0.005 mm. These types of rock can be distinguished by grain size in images of the same scale.

[0082] Table 1. Rock Grain Size and Color Characteristics

[0083]

[0084]

[0085] Since coal is generally black or grayish-black in color and has a fine particle size, this color characteristic can be used to distinguish coal from sandstone and mudstone.

[0086] Figure 1 The flowchart of the domain adaptive method for lithology identification models in different regions of this invention is shown below, with specific steps as follows:

[0087] S1, after classifying the image dataset containing various types of lithological data, assigns lithological labels to it as the source domain dataset;

[0088] For n in the source domain s Artificial lithology classification and labeling are performed on an image dataset containing various types of lithology data (i.e., n). s (This represents the number of rock images with lithological labels in the source domain), which is used as the labeled dataset in the source domain. in This represents the i-th labeled rock sample in the source domain. This represents manually labeled lithology tags; the number of iterations, N=10, is set as input.

[0089] S2 uses the saved well data, including core scan images and electron microscopy scan images, as the target domain dataset.

[0090] The target domain stores n t Unlabeled core scan images and electron microscopy scan images constitute the target domain unlabeled lithology dataset. in This represents the i-th unlabeled rock sample in the target domain.

[0091] S3. Based on the source domain dataset obtained in step S1 and the target domain dataset obtained in step S2, the rock images are preprocessed and then introduced into the feature extractor F in the CNN lithology identification model for feature extraction and dimensionality reduction. The initial projection P0 is then obtained using the source domain data.

[0092] S31: Read rock images from the source and target domain datasets, select a 256×256 pixel region in the center of the image, resize the input image to 224×224 pixels, and after normalization and data augmentation, transmit it to the feature extractor F of the CNN rock identification model. The feature extractor F extracts features based on the color, shape, texture, and structure of the rocks in the image. and Feature vectors: and d represents the feature dimension;

[0093] S32, using Local Preserving Projection (LPP) to reduce the dimensionality of rock sample feature vectors in the source and target domains, the objective function is:

[0094]

[0095] Where P is the projection matrix sought by the objective function; D is the matrix generated by concatenating the feature vectors of rock image samples from the source and target domains; D is a diagonal matrix. The similarity matrix M is defined as:

[0096]

[0097] L = DM is the Laplace matrix.

[0098] Therefore, finding the optimal projection vector of the objective function is equivalent to finding the eigenvector corresponding to the smallest eigenvalue in the characteristic equation, as shown below:

[0099] XLX T p = λXDX T p (2)

[0100] Thus, the projection matrix is ​​obtained:

[0101]

[0102] And low-dimensional feature matrix:

[0103]

[0104] In the formula d1≤d denotes the dimension of the feature space after dimensionality reduction; n=n s +n t , where is the total number of samples.

[0105] The matrix composed of the low-dimensional feature vectors of all labeled rock samples in the source domain. Substituting into the objective function of equation (1), the initial projection matrix is ​​obtained:

[0106]

[0107] Where d2 represents the dimension of the projected subspace, and d2 <d1。

[0108] S4, In the learning subspace, a lithological pseudo-label is applied to all data in the target domain using a class prototype-based structured prediction method;

[0109] Specifically as follows:

[0110] S41, Calculate the projection of the source region rock sample:

[0111]

[0112] S42, Calculate the projection of the target domain rock sample:

[0113]

[0114] S43, Projecting source domain rock samples z s Projection z of rock sample in the target domain t After centralization and L2 normalization, the average value z of the projections of all source domain rock samples and target domain rock samples is obtained;

[0115] Based on the prototype, sample x belonging to the lithological target domain of type y. t The conditional probability is:

[0116]

[0117] Among them, |У 集合 | Indicates the number of rock categories; The class prototype of class y is the average vector of source domain rock samples with lithological label y, expressed as follows:

[0118]

[0119] When y = y i hour, otherwise,

[0120] Rock samples x belonging to class y in the target domain based on structured prediction t The conditional probability is:

[0121]

[0122] in, Represents the cluster center corresponding to class y, |У 集合 | indicates the number of rock categories.

[0123] S44, for rock sample x in the target domain t The formula for predicting lithological pseudo-labels is as follows:

[0124]

[0125] Where p(y|x) t )=max{p1(y|x t ),p2(y|x t )},У 集合 A set representing rock categories.

[0126] Combine the target domain sample set with lithological pseudo-labels to form a pseudo-label dataset:

[0127]

[0128] S5, a reinforcement learning-based lithology pseudo-label selection strategy is used to select the most reliable lithology pseudo-label samples in the target domain;

[0129] A reinforcement learning-based strategy for selecting lithological pseudo-labels is used, as follows:

[0130] The problem of selecting lithology pseudo-labeled samples can be described as a Markov decision process, including a state space, an action space, and a reward. An agent (deep Q-network) is trained to select lithology pseudo-labeled samples. The pseudo-labeled dataset D is defined as follows: u n obtained from random sampling c A vector of pseudo-labeled lithology instances composed of 10 pseudo-labeled samples to be selected (i.e., the candidate set) and the positive set vector composed of selected lithological pseudo-marker samples in Initialize to empty. Set the deep Q-network to a three-layer fully connected network, with the activation function of each layer set to the ReLU function.

[0131] Define the state space Each state s i Contains pseudo-marker lithology instance vector D c , positive set vector And the target domain lithology-free marker vector D t lithological pseudo-marker instance vector D c From examples low-dimensional feature vectors Composition; positive set vector From rock examples low-dimensional feature vectors Composition; lithological marker vector D t Rock instances from the target domain low-dimensional feature vectors composition.

[0132] Define action space Each action a i This indicates that the current pseudo-marker lithology instance vector D c (i.e., as a candidate set) selects pseudo-labeled rock instances for each state s. i The agent takes action a i Select pseudo-marker lithology instance vector D c Rock instances in the vector and move them to the positive set vector In the process of executing the transfer of instances from the pseudo-labeled lithology instance vector D c Move to the positive set vector The selected instance is replaced with a zero vector after the action.

[0133] Define a reward R to reflect the accuracy of the selected lithological pseudomarker instances. Set a metric function to measure representativeness and accuracy, as follows:

[0134]

[0135] Where β is a hyperparameter; The type of lithological pseudomarker predicted by the recent prototype combined with structured prediction method. The probability of Δ can represent the confidence level of the prediction; e The formula representing the reduction in entropy of unlabeled data in the target domain is:

[0136] Δ e =HH′ (14)

[0137] Where H is s i The entropy value of a state, H′ is the next state s. i+1 The entropy value under;

[0138] The formula for calculating entropy is:

[0139]

[0140] Set the final reward function as follows:

[0141]

[0142] Where τ is the set threshold, and its value is (1+β)log(0.8).

[0143] The strategy for selecting lithological pseudo-label instances is learned using a deep Q-learning method. For each state and action (s) i ,ai ), using the Q function Q(s) i ,a i The cumulative reward for the discounted state and action is represented by ), and the iterative learning method for deep updating the Q-network is as follows:

[0144]

[0145] Where Ω represents the parameters of the Q-network; ε is the exploration factor; V i For Q(s) i ,a i The target value is given by the formula:

[0146]

[0147] Here, γ represents a discount factor that determines the importance of future cumulative rewards compared to current rewards.

[0148] When the agent (deep Q network) is invoked to select lithology pseudo-label instances, the operation using the policy output is defined as a. t The formula is:

[0149]

[0150] Among them, s t Indicates the current state.

[0151] S6. Based on the sample set consisting of the lithological pseudo-labeled rock samples obtained in step S5 and the source domain lithological labeled rock samples, local preservation projection is used as the enabling technique to complete the class alignment task of the two domains.

[0152] The matrix consists of the low-dimensional feature vector set of labeled rock samples in the source domain and the lithological pseudo-label set selected by reinforcement learning. Substitute into the formula In the process, the projection matrix P of the k-th iteration is obtained. k .

[0153] S7, based on the prototype-based structured prediction method in step S4, update the target domain rock samples with lithological pseudo-labels;

[0154] S8. Repeat steps S5 to S7 until the target number of iterations of 10 is reached.

[0155] Table 2 shows the implementation algorithm of the domain adaptation method of the present invention. The specific implementation process is as follows:

[0156] B1, a dataset of labeled rock images from the source domain. Target Domain Unlabeled Rock Image Dataset The iteration count N = 10 is set as input. The count k = 0 is initialized. After cropping, normalization, and data augmentation preprocessing of the rock images from the source and target domains, the feature extractor F in the CNN extracts the feature vectors of the rock instances from the source and target domains. and

[0157] B2, the matrix X generated by concatenating the feature vectors of the source domain rock samples and the feature vectors of the target domain rock samples is used for dimensionality reduction based on LPP. The objective function is: Find the projection matrix P and the low-dimensional matrix. A matrix composed of low-dimensional feature vectors of labeled samples in the source domain Substitute into the formula In the process, the initial projection matrix P0 is obtained.

[0158] B3, Calculate the projections of the source domain rock samples and the target domain rock samples: The source and target domain rock sample projections are centralized and L2 normalized to obtain the average value z of the sample projections. The target domain rock sample x belonging to class y is then calculated. t Conditional probability: p(y|x) t )=max{p1(y|x t ),p2(y|x t )},in

[0159]

[0160] B4 assigns lithological pseudo-labels to all data in the target domain. The target domain rock sample set and lithology pseudo-labels are combined to form a pseudo-label dataset.

[0161] B5, k←k+1, initialize the positive set vector The experience pool M = φ.

[0162] B6, When the model does not converge, from the pseudo-label dataset D u A candidate set D is formed by randomly selecting lithological pseudo-label instances. c Generate initialization state s0.

[0163] B7, when D c When ≠φ, change the current state s t As input, obtain the output action. Take action a t Update candidate set D c Sum of positive set vectors And generate the next state s t+1 Calculate the reward r for the current state transition. t. The current record (s) t ,a t ,r t ,s t+1 Write the data into the experience pool M. Extract multiple sets of records {(s) from the experience pool M. i ,a i ,r i ,s i+1 Update the parameters of the deep Q network.

[0164] B8, if r t If r < 0, stop the current loop and execute B6. t If the value is ≥0, proceed directly to the next iteration of the loop and execute B7.

[0165] B9, repeat B6-B8 until the model converges, and output the value function Q corresponding to each action a.

[0166] B10 is a matrix composed of the source domain labeled dataset and the lithology pseudo-label set selected by reinforcement learning. Update projection matrix P k .

[0167] B11, Update the lithological pseudo-labels for all target data in the target domain:

[0168] B12, repeat B5-B9 until k=N, output the projection matrix P. N Predicting lithological labels

[0169] Table 2. Algorithms for Unsupervised Domain Adaptation Based on Selective Pseudo-labeling

[0170]

[0171]

[0172]

[0173] Figure 2 The following is a diagram illustrating the iterative process of learning the projection matrix in this invention:

[0174] A matrix composed of low-dimensional feature vectors of labeled samples in the source domain Substitute into the formula

[0175] In the process, the initial projection matrix is ​​obtained.

[0176] Count k=1, and form a matrix by combining the low-dimensional feature vector set of labeled samples in the source domain with the pseudo-label set obtained from reinforcement learning. Substitute into the formula In the process, the k-th cyclic projection matrix P is obtained. k Let k = k + 1. Iterate through the projection calculation and counting process until k = N, to obtain the final projection matrix P. N .

[0177] Figure 3 The flowchart of the reinforcement learning-based lithology pseudo-labeling algorithm of this invention is as follows:

[0178] Define k as initial value 0, from pseudo-label dataset D u n obtained from random sampling c A candidate set consisting of lithological pseudomarker samples to be selected and the positive set vector composed of selected lithological pseudo-marker samples Where the positive set vector Initialize to empty. The specific implementation steps are as follows:

[0179] C1, k←k+1, initialize the positive set vector The experience pool M = φ.

[0180] C2, when the model converges, from the pseudo-label dataset D u A candidate set D is formed by randomly selecting lithological pseudo-label instances. c Generate initialization state s0.

[0181] C3, when candidate set D c When ≠φ, change the current state s t As input, obtain the output action. Take action a t : Transfer rock instances from pseudo-label dataset D c Move to After the operation, the candidate set D is updated by replacing the selected instances with the zero vector. c Sum of positive set vectors And generate the next state s t+1 .

[0182] C4, set the measurement function:

[0183] Set reward function Set the threshold τ = (1 + β)log(0.8),

[0184] Calculate the reward r for the current state transition. t .

[0185] C5 will move the current record (s) t ,a t ,r t ,s t+1 Write the data into the experience pool M, and extract multiple sets of records from the experience pool M {(s)}i ,a i ,r i ,s i+1 Update the parameters of the deep Q network.

[0186] C6, if r t If r < 0, stop the current loop and proceed to step C2. t If the value is ≥0, proceed directly to the next loop and execute steps C3-C5.

[0187] C7. Repeat steps C2-C6 until the model converges, and output the value function Q corresponding to each action a.

Claims

1. An unsupervised adaptive method for identifying lithology models in different regions, characterized in that, The steps include the following: S1, after manually classifying the rock image dataset containing various types of lithological data, assigns lithological labels to it as the source domain dataset; S2, use the saved core scan images and electron microscope scan images as the target domain dataset; S3, based on the source domain dataset and the target domain dataset, after preprocessing the rock images, the feature extractor F in the CNN-based lithology identification model is introduced to extract features and reduce dimensions, and the initial projection P0 is obtained using the labeled rock samples in the source domain. S4, In the learning subspace, a lithological pseudo-label is applied to all data in the target domain using a class prototype-based combined structured prediction method; S5, a reinforcement learning-based lithology pseudo-label selection strategy is used to select the most reliable lithology pseudo-label samples in the target domain; S6. Combine the lithological pseudo-label samples obtained in step S5 with the lithological label samples in the source domain into a sample set, and use local preservation projection as an enabling technique to complete the class alignment task of the two domains. S7, using a prototype-based combined structured prediction method to update the target domain rock samples with lithological pseudo-labels; S8. Repeat steps S5 to S7 until the target number of iterations is reached.

2. The unsupervised adaptive method for lithological identification models in different regions according to claim 1, characterized in that, In step S1, based on lithological classification, coal, conglomerate, coarse sandstone, medium sandstone, fine sandstone, siltstone, and mudstone are used as lithological labels; among them, conglomerate has a particle diameter greater than 2 mm, coarse sandstone has a particle diameter of 1~0.5 mm, medium sandstone has a particle diameter of 0.5~0.25 mm, fine sandstone has a particle diameter of 0.25~0.05 mm, siltstone has a particle diameter of 0.05~0.005 mm, and mudstone has a particle diameter less than 0.005 mm; the color of coal is black or grayish-black.

3. The unsupervised adaptive method for lithology identification models in different regions according to claim 1, characterized in that, In step S1, n in the source domain s Artificial lithological classification and labeling were performed on a rock image dataset containing various lithological data, serving as a labeled dataset for the source domain. ,in This represents the i-th labeled rock sample in the source domain. This indicates a manually labeled lithology tag.

4. The unsupervised adaptive method for lithology identification models in different regions according to claim 1, characterized in that, In step S2, n stored in the target domain is... t The target domain unlabeled dataset consists of 100 unlabeled core scan images and electron microscopy scan images. ,in This represents the i-th unlabeled rock sample in the target domain.

5. The unsupervised adaptive method for lithological identification models in different regions according to claim 1, characterized in that, In step S3, the preprocessing steps for the rock image are as follows: S31: Read rock images from the source domain dataset and the target domain dataset. Select a 256×256 pixel region in the middle of the rock image and resize it to 224×224 pixels. After normalization and data augmentation, transmit it to the feature extractor F of the CNN-based lithology recognition model. The feature extractor F extracts the i-th rock sample from the source domain based on the color, shape, texture, and structure of the rocks in the image. With the i-th rock sample in the target domain Feature vectors: and d represents the feature dimension; S32, using local preservation projection to reduce the dimensionality of the feature vectors of rock samples in the source and target domains, the objective function expression is: Where P is the projection matrix sought by the objective function; , is the matrix generated by concatenating the feature vectors of rock samples from the source and target domains; n s n represents the number of rock image sets containing various types of lithological data in the source domain. t This represents the total number of unlabeled core scan images and electron microscopy scan images in the target domain; D is a diagonal matrix The similarity matrix M is defined as: ； L=DM is the Laplace matrix; The eigenvector p corresponding to the smallest eigenvalue is obtained using the following formula: The eigenvector p corresponding to the smallest eigenvalue is the optimal projection vector of the objective function; Solving for the projection matrix yields the solution. and low-dimensional feature matrix In the formula , This represents the dimension of the feature space after dimensionality reduction. ; S33 is a matrix composed of low-dimensional feature vectors of labeled rock samples from the source domain. Substitute into the formula In the process, the initial projection matrix is ​​obtained. ,in Denotes the dimension of the projected subspace, and .

6. The unsupervised adaptive method for lithological identification models in different regions according to claim 1, characterized in that, In step S4, the implementation steps for performing lithological pseudo-labeling on all rock sample data in the target domain using a prototype-based combined structured prediction method are as follows: S41, Calculate the projection of the source region rock sample: ， in, The matrix generated by concatenating the feature vectors of rock samples from the source and target domains. X The low-dimensional feature matrix; Represents the k-th projection matrix P k The transpose of the matrix; S42, Calculate the projection of the target domain rock sample: ， S43, Projecting source region rock samples Projection of rock samples from the target domain After centralization and L2 normalization, the average value z of the projections of all source domain rock samples and target domain rock samples is obtained; S44, for rock samples in the target domain The expression for predicting lithological pseudo-labels is as follows: in, ; in, Indicates the number of rock categories; express The class prototype of lithology is the average vector of the projections of source domain rock samples with lithology label y, expressed as follows: when hour, ;otherwise, ; in, Indicates and The cluster center corresponding to the class; Rock samples in the target domain The formula for predicting lithological pseudo-labels is as follows: ， in, , A set representing rock categories; Combine the target domain rock sample set with lithological pseudo-labels to form a pseudo-label dataset: 。 7. The unsupervised adaptive method for lithological identification models in different regions according to claim 1, characterized in that, In step S5, the process of selecting the most reliable lithological pseudo-label samples in the target domain is a Markov decision process, which includes a state space, an action space, and a reward. The agent is trained to select lithological pseudo-label samples; the pseudo-label dataset D is defined. u n obtained from random sampling c The pseudo-marker lithological feature vector is composed of 10 pseudo-marker samples to be selected. and the positive set vector composed of selected lithological pseudo-marker samples , where the positive set vector Initialize to empty; Define a state space S, the state space Each state Including pseudo-marker lithological eigenvectors D c , positive set vector And the target domain unlabeled lithological eigenvector D t Among them, the pseudo-marker lithological feature vector D c From rock examples low-dimensional feature vectors Composition; positive set vector From rock examples low-dimensional feature vectors Composition; Target domain without lithological marker vector D t Rock instances from the target domain low-dimensional feature vectors composition; Define action space A, the action space Each of the actions This indicates that the lithological eigenvector D from the current pseudo-marker... c Select a lithological pseudo-marker instance for each state. The agent takes action. Choose D c Rock instances with reliable lithology and move them to In the middle, replace the selected rock instance with the zero vector; Define a reward R, which reflects whether the selected lithological pseudo-marker instance is more accurate; set a metric function to measure representativeness and accuracy, as follows: in: For hyperparameters; Lithological pseudo-label types predicted by combining prototype-based and structured prediction methods The probability represents the confidence level of the lithology prediction; The formula representing the reduction in entropy of unlabeled rock sample data in the target domain is: in, for The entropy value of the state, For the next state The entropy value under; The formula for calculating entropy is: ; Set the reward function as follows: ,in: The threshold value is set to 1. ; For each state and action Using functions To represent the cumulative reward for the discounted state and action, the iterative learning method for deep updating the Q-network is as follows: in, These represent the parameters of the Q-network; For exploration factors; for The target value is given by the formula: in, This represents a discount factor that determines the importance of future accumulated rewards compared to current rewards. ; When the agent is invoked to select lithology pseudo-marking instances, the operation using the policy output is defined as follows: The formula is: in: Indicates the current state.

8. The unsupervised adaptive method for lithological identification models in different regions according to claim 1, characterized in that, In step S6, for the lithology pseudo-labeled samples and the source domain rock samples with lithology labels, local preservation projection is used as the enabling technique to obtain the projection matrix and complete the class alignment task of the two domains. The specific implementation process is as follows: The matrix consists of the low-dimensional feature vector set of lithologically labeled samples in the source domain and the set of pseudo-labeled lithological samples selected by reinforcement learning. Substitute into the formula In the process, the projection matrix of the k-th iteration is obtained. ; For the k-th iteration, the source domain has a matrix composed of the concatenated low-dimensional feature set of lithological labeled samples and the set of lithological pseudo-label samples selected by reinforcement learning; for The transpose of the matrix; D diagonal matrix ; L=DM It is a Laplace matrix.

Images