Uranium reservoir sedimentary facies belt automatic identification and division method based on logging data constraint

Abstract

The invention relates to the field of mineral exploration, and discloses a uranium reservoir sedimentary facies belt automatic identification and division method based on logging data constraint. Generating initial feature vectors of the nodes; inputting the layer control well bitmap into a map attention network, and performing spatial feature fusion on the initial feature vector of the node to obtain a fused feature vector; dynamically generating a geological transition probability matrix for each well; calculating the emission probability of each depth point according to the logging data and the fusion feature vector; and decoding an optimal sedimentary facies belt sequence as a division result by adopting a Viterbi algorithm in combination with the emission probability and the geological transition probability matrix. According to the method, a layer control well position map is constructed, a well pattern relationship is introduced into a model, and adaptive weighted fusion is performed on inter-well information by using a map attention network, so that inter-well transverse geological relevance is effectively brought into a sedimentary facies identification process, and transverse continuity and geological rationality of a division result are improved.

Description

Technical Field

[0001] This invention relates to the field of mineral exploration technology, specifically to a method for automatic identification and division of sedimentary facies zones in uranium reservoirs based on well logging data constraints. Background Technology

[0002] The study of sedimentary facies is fundamental to reservoir prediction and evaluation, and plays a decisive role in understanding the distribution patterns of stratabound orebodies such as sandstone-type uranium deposits. Well logging curves continuously reflect the vertical variations in formation lithology and physical properties, and are the primary data source for sedimentary facies classification.

[0003] Traditional sedimentary facies classification mainly relies on manual interpretation by geologists. This process is not only labor-intensive and inefficient, but the interpretation results are also easily influenced by subjective experience, making it difficult to guarantee consistency. To improve interpretation efficiency and objectivity, researchers have introduced various machine learning and deep learning algorithms to achieve automatic identification of facies zones by learning the mapping relationship between well logging responses and sedimentary facies.

[0004] However, existing automatic identification methods still have limitations in application. Most methods treat the logging sequence of each well as an independent sample, focusing on exploring the vertical variation patterns of the logging curves while ignoring the lateral spatial correlation between wells. Due to the lack of consideration for the continuity of geological bodies between wells, the single-well identification results obtained by these methods may show geological unconformities when comparing multiple wells, affecting the reliability of the reservoir distribution morphology in the profile and even in three-dimensional space.

[0005] Furthermore, to ensure that the identification results conform to geological evolution patterns, some methods introduce vertical succession probabilities of sedimentary facies as constraints. These constraints are typically represented by a fixed geological transition probability matrix and applied to all wells across the entire work area. However, sedimentary environments are spatially heterogeneous, and sedimentary succession patterns may differ across regions. A globally static geological constraint is insufficient to adapt to such spatial variations, limiting the model's identification accuracy under different geological backgrounds. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an automatic identification and division method for sedimentary facies zones in uranium reservoirs based on well logging data constraints. This method solves the problem that existing automatic sedimentary facies identification methods fail to simultaneously consider the spatial correlation between wells and the spatial variability of geological patterns, resulting in geological inconsistencies and limited accuracy in the identification results.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an automatic identification and division method for uranium reservoir sedimentary facies zones based on well logging data constraints, comprising the following steps: First, a layered well location map with wells as nodes is constructed, and a one-dimensional convolutional neural network encoder is used to process the logging sequence data of each well to generate an initial feature vector that can characterize the vertical logging features of the well.

[0008] Subsequently, the constructed layered well location map and the initial feature vectors of the nodes are input into the graph attention network. This network performs spatial feature fusion on the initial feature vectors of the nodes through an information propagation mechanism, thereby generating a fused feature vector for each node that contains both its own vertical information and the spatial correlation of its neighborhood.

[0009] Next, the fused feature vector corresponding to each well is input into an independent transition matrix generation network. This network's function is to calculate and generate a uniquely corresponding geological transition probability matrix for each well based on the input fused feature vector. This geological transition probability matrix is ​​a row-normalized square matrix, whose elements quantify the probability of one sedimentary facies evolving into another in the specific geological context in which the well is located.

[0010] Subsequently, based on the original well logging data and the aforementioned fused feature vector, the emission probability of various candidate sedimentary facies at each depth sampling point is calculated.

[0011] Finally, by combining the emission probability at each depth point with the geological transfer probability matrix dynamically generated for the well, the Viterbi algorithm is used to perform a global optimal path search, decoding a sedimentary facies sequence that runs through the entire well section and has the highest joint probability, which is used as the final partitioning result.

[0012] Preferably, the one-dimensional convolutional neural network encoder is obtained through self-supervised pre-training. This pre-training process utilizes well logging data within the work area that does not require manual annotation. By applying data augmentation operations such as Gaussian noise injection, sequence masking, or temporal distortion to the well logging sequence segments, positive and negative sample pairs are constructed, and optimization is performed using a contrastive loss function. This enables the encoder to learn the inherent and universally applicable feature representations of the well logging data.

[0013] Preferably, the construction process of the layered well location map is as follows: each well in the work area is defined as a node; the average horizontal distance between any two wells in the target formation unit is calculated; when the distance is less than a preset distance threshold set based on geological understanding, an edge connection is established between the corresponding two nodes.

[0014] Preferably, when the graph attention network performs spatial feature fusion, it calculates the attention coefficient between each center node and its neighboring nodes. This calculation process includes linearly transforming and concatenating the feature vectors of the center node and its neighboring nodes, calculating the original attention score, and then normalizing the score. The network uses this attention coefficient to perform a weighted summation of the feature vectors of the neighboring nodes to update the feature vector of the center node.

[0015] Preferably, the Viterbi algorithm decoding process involves calculating an optimal path probability for each candidate sedimentary facies at each depth point. This calculation is implemented recursively: first, among all possible paths from all sedimentary facies states at the previous depth point to a candidate sedimentary facies at the current depth point, the path with the highest probability is determined; then, the probability of this highest-probability path is combined with the emission probability of that candidate sedimentary facies at the current depth point to obtain the optimal path probability for the current state.

[0016] Preferably, the process of calculating the emission probability at each depth point involves combining the logging data at the current depth point with the fused feature vector corresponding to that well, and using this combination as input to the decoder network. The decoder network outputs a probability vector, defined as the emission probability at the current depth point, where each element corresponds to the probability value of the current depth point being interpreted as one of the various candidate sedimentary facies.

[0017] This invention provides an automatic identification and division method for uranium reservoir sedimentary facies zones based on well logging data constraints. It has the following beneficial effects: 1. This invention constructs a layered well location map and introduces the well network relationship into the model. It uses a graph attention network to adaptively weight and fuse the information between wells, so that the final facies identification not only depends on the vertical data of a single well, but also considers the geological environment of neighboring wells. This effectively incorporates the lateral geological correlation between wells into the sedimentary facies identification process, improving the lateral continuity and geological rationality of the division results.

[0018] 2. This invention generates a unique geological transfer probability matrix for each well, replacing the globally uniform static constraints in traditional methods. This allows the model to follow the differentiated sedimentary succession patterns at different spatial locations, realizing the dynamic and personalized nature of geological constraints and improving the model's adaptability to reservoir heterogeneity.

[0019] 3. This invention performs self-supervised pre-training on the feature extraction encoder before supervised training, enabling it to learn universal feature representations of logging curves from a large amount of unlabeled logging data. This reduces the model's dependence on manually labeled samples, and at the same time, it improves the recognition accuracy and generalization ability of the final model by obtaining better initial parameters. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 This invention provides an automatic identification and division method for uranium reservoir sedimentary facies zones based on well logging data constraints, comprising the following steps: Step S100: Model Pre-training. This step involves self-supervised pre-training of a one-dimensional convolutional neural network (1D-CNN) encoder. This encoder is used in subsequent steps to extract vertical sequence features from well logging curves. The pre-training process utilizes well logging data within the work area that does not require manual annotation. Through contrastive learning, the encoder learns the inherent structure and patterns of the well logging data itself. After pre-training is complete, the encoder's network parameters are saved for initialization in subsequent steps.

[0023] Step S200: Graph Construction and Feature Initialization. This step first performs standardized preprocessing on the logging data of all wells within the target work area. Then, based on the geographic coordinates of each well and the division of the target stratigraphic units, a stratified well location map is constructed. In this map, each well is defined as a node, and edge connections are established between node pairs whose spatial distance between wells is less than a preset threshold. Next, the 1D-CNN encoder pre-trained in step S100 is invoked to process the logging sequence data of training wells with sedimentary facies annotations, converting the logging sequence of each well into a fixed-dimensional initial feature vector, which is assigned to the corresponding node on the well location map.

[0024] Step S300: Spatial Feature Fusion and Dynamic Constraint Generation. The well location map constructed in Step S200, with nodes having completed feature initialization, is input into a Graph Attention Network (GAT). GAT calculates and updates the attention weights of each node to its neighboring nodes through a multi-layer information propagation mechanism, and aggregates neighborhood information based on these weights, thereby generating a final feature vector that integrates vertical sequence information and lateral spatial correlation. Subsequently, this final feature vector is input into a Transition Matrix Generation Network (TMGN) to calculate and generate a corresponding personalized geological transition probability matrix for the well represented by the node. This matrix quantifies the probability of vertical succession between different sedimentary facies within the specific geological environment of the well.

[0025] Step S400: Sequence Decoding and Facies Zone Division. For each well, a decoder is used to calculate the emission probability of each candidate sedimentary facies at each depth sampling point based on the raw logging data and the final feature vector generated in step S300. Then, this emission probability is integrated with the specific geological transfer probability matrix generated for the well in step S300, and the Viterbi algorithm is used for optimal path decoding. This algorithm searches for and determines a sedimentary facies sequence with the highest joint probability that runs through the entire well section.

[0026] Step S500: Result Output. The optimal sedimentary facies sequence decoded in step S400 is matched with the corresponding depth information to generate the final sedimentary facies zoning result for the well. This result can be formatted and output to generate various geological analysis maps such as single-well sedimentary facies columnar sections and multi-well sedimentary interconnected well profiles.

[0027] In one embodiment, the encoder used to extract vertical sequence features from well logging is subjected to self-supervised pre-training. Step S1 is performed before supervised training, and its purpose is to enable the encoder network to learn the inherent, universal feature representation capabilities of the well logging curve data itself, utilizing a large amount of well logging data that does not require manual annotation. This pre-training process improves the quality of the encoder's initial parameters, thereby enabling better performance in the subsequent supervised training phase with fewer labeled samples and faster convergence.

[0028] The pre-training process can be broken down into the following steps: Step S101: Training Sample Construction. From all available well logging data within the work area, fixed-length well logging sequence segments are extracted using a sliding window method to form the pre-training base sample set. For any well logging sequence segment in this sample set, i.e., the anchor sample, two different data augmentation operators are applied to generate two new sequence segments with related content but slightly different representations; these two constitute a positive sample pair. All other sequence segments in the dataset, except for the anchor sample itself and its augmented view, are considered negative samples of the anchor sample.

[0029] Specific technical means of data augmentation include, but are not limited to: Gaussian noise injection: Random Gaussian noise with a mean of zero and a preset small variance is added to the original logging sequence values ​​to simulate instrument measurement errors or background noise.

[0030] Sequence masking: Randomly select a portion of continuous or discrete time points in a sequence and replace their corresponding logging values ​​with zero or the global mean to improve the model's robustness to local missing data.

[0031] Temporal distortion: Slight, non-linear stretching or compression of the time axis of a sequence to simulate subtle variations in formation thickness or inhomogeneities in sampling intervals.

[0032] In this way, a set of positive and negative samples for comparative learning was constructed for each anchor point sample, providing a data foundation for the model to learn and distinguish the essential features of different logging modes.

[0033] Step S102: Encoder Network Construction and Training. In this embodiment, the encoder adopts a one-dimensional convolutional neural network (1D-CNN) structure. This network consists of multiple convolutional layers, activation layers, and pooling layers stacked together. The convolutional layers are responsible for extracting local morphological features of the logging curve, such as peaks, smooth surfaces, and gradient changes; the activation layers (such as ReLU) introduce nonlinearity to enhance the model's expressive power; and the pooling layers downsample the feature map, increasing the receptive field and reducing computational cost. For the specific implementation of the 1D-CNN network, those skilled in the art can configure it according to the actual data dimensions and computational resources. Its network construction is a well-known technology in the field and will not be described in detail here.

[0034] In model training, the core principle is to maximize the similarity of positive sample pairs in the feature space while minimizing their similarity with all negative samples. This objective is achieved by optimizing the contrastive loss function. For any sample within a batch, the loss... The calculation method is as follows: ; In the formula, and These are the feature vectors output by the encoder from the two samples in a positive sample pair; For the first batch The feature vector output by the encoder from each negative sample; The cosine similarity function between two feature vectors; Temperature is a hyperparameter used to adjust the distribution of similarity scores. Its value is usually between 0.01 and 0.5. Smaller values ​​will increase the penalty for difficult-to-distinguish negative samples. This represents the total number of samples in this training batch; For indicator functions, when index Not equal to the current sample index The value is 1 if it is true, and 0 otherwise.

[0035] The gradient is calculated using the backpropagation algorithm based on the aforementioned loss function, and the weight parameters of the encoder network are updated. After the pre-training process converges, the encoder network parameters are saved for use in subsequent steps.

[0036] In one embodiment, after the model pre-training is completed, the raw data needs to be processed, and the spatial relationship of multiple wells is abstracted into graph-structured data. At the same time, each node in the graph is given an initial feature that can characterize its vertical logging sequence information.

[0037] This process can be broken down into the following steps: Step S201: Well Logging Data Preprocessing. The raw well logging data from all wells within the target area are preprocessed to eliminate systematic differences between data and improve data quality. Preprocessing operations include depth correction, data cleaning, and data standardization. Data standardization aims to unify well logging curves of different types and dimensions into a similar numerical range. A specific standardization method is Z-score standardization, which is calculated by subtracting the average value of the curve from the value of each sampling point on the original well logging curve, and then dividing by the standard deviation of the curve. Other well logging data preprocessing methods, such as outlier removal and missing value imputation, are standard techniques in this field and will not be elaborated upon here.

[0038] Step S202: Construction of Stratified Well Location Map. To characterize the lateral geological relationships between wells in the model, this embodiment constructs a stratified well location map. This map is an undirected graph, where each node represents a well within the work area. The connecting edges between nodes are established based on the spatial proximity between wells, and its construction rules are as follows: Node definition: Each node in the diagram represents a well within the work area.

[0039] Edge establishment conditions: Connection edges between nodes are established based on the spatial proximity of wells. Specifically, firstly, the vertical depth range of all wells within the target reservoir unit is determined through stratigraphic correlation. Then, the average horizontal distance between any two wells within the target unit is calculated using wellbore trajectory data. If the average horizontal distance between two wells is less than a preset distance threshold, an edge is established between these two nodes. The preset distance threshold is set based on regional geological understanding; for example, it can be set to 1.0 to 1.5 times the average width of the target sedimentary body (such as a single channel sand body) on the plane to ensure that wells within the same sedimentary unit can be effectively connected.

[0040] In the above way, the geospatial distribution of the well network is transformed into a graph topology that can be directly processed by subsequent models.

[0041] Step S203: Node Feature Initialization. An initial feature vector is assigned to each node in the graph. This vector must contain the vertical logging information of the well represented by the node. In this embodiment, the 1D-CNN encoder pre-trained in step S100 is used to perform this task. The specific process is as follows: For each well, a two-dimensional matrix is ​​constructed from the pre-processed logging curves (such as natural gamma, resistivity, neutron, and density) within the target reservoir unit. The number of rows in this matrix corresponds to the number of depth sampling points, and the number of columns corresponds to the number of logging curve types. This two-dimensional matrix is ​​used as input and fed into a pre-trained 1D-CNN encoder for forward computation. The encoder network maps this variable-length two-dimensional input matrix and outputs a fixed-dimensional vector, which is the initial node feature vector of the well on the graph. This feature vector encodes the vertical variation pattern, morphology, and numerical distribution characteristics of the well logging curve combination in a compact form.

[0042] In one embodiment, after obtaining a stratified well location map with initial features, an information fusion and constraint generation process is required to achieve in-depth exploration of the lateral correlation between wells and provide personalized geological constraints for subsequent sequence decoding.

[0043] This process can be broken down into the following steps: Step S301: Adaptive Spatial Feature Fusion. This embodiment employs a Graph Attention Network (GAT) to iteratively update and fuse the node features in the well location map. The basic principle is that the model can automatically learn the influence weights of each well (node) on its neighboring wells, rather than simply averaging all neighbor information, thus achieving adaptive weighted fusion of information. This network consists of multiple stacked graph attention layers and can employ a multi-head attention mechanism to enhance the model's expressive power and stability. In each graph attention layer, the feature vector of each node (well) is updated according to the influence of its neighboring nodes.

[0044] Specifically, in the first An attention layer, for any node in the graph. Its neighbor nodes Attention coefficient Calculated in the following way: ; ; In the formula, For nodes The feature vector at the (k-1)th layer; For the first The layer is a shared weight matrix used to perform linear transformations on the feature vectors; This is a vector concatenation operation; It is a trainable attention weight vector; For leaky modified linear unit activation functions; For nodes and The original attention score between; For nodes The set of neighboring nodes in the graph; The final attention coefficients are normalized using the Softmax function, and their values ​​quantify the neighboring nodes. For updating nodes The importance of features.

[0045] After obtaining the attention coefficient, the node eigenvectors from Updated to The process is as follows: ; In the formula, For nodes In the The updated feature vector after the layer; It is a non-linear activation function.

[0046] By executing the above process Second-rate( The total number of layers in the GAT (typically 2 to 4), and the final output feature vector of each node. It integrates its own multi-level neighborhood information. This vector not only contains the vertical sequence features of the well itself, but also implies the lateral spatial dependence of its local well network environment.

[0047] Step S302: Dynamic Geological Transition Probability Matrix Generation. To integrate geological evolution laws into the model in a data-driven manner, this embodiment designs a Transition Matrix Generation Network (TMGN) to generate a unique geological transition probability matrix for each well. The TMGN is specifically implemented as a Multilayer Perceptron (MLP), which may contain an input layer, one or two hidden layers using the ReLU activation function, and a linear output layer.

[0048] The network's workflow is as follows: The final feature vector of each node, output by GAT in step S301, is... This vector serves as the input to TMGN. After undergoing a nonlinear transformation through the hidden layer of TMGN, the output layer outputs a vector of size [value missing]. The matrix, where This represents the total number of predefined sedimentary facies types. To ensure that this output matrix conforms to the definition of a probability transition matrix, it needs to be normalized to obtain the final dynamic geological transition probability matrix. : ; In the formula, For nodes The final feature vector after GAT fusion; Network mapping function representing TMGN; well node A proprietary dynamic geological transfer probability matrix; Operation representation pair Each row of the output matrix is ​​applied independently function; This is the final dynamic geological transfer probability matrix.

[0049] This operation ensured the matrix The sum of each row of elements in the matrix is ​​1. The elements in the matrix represent the geological context of the well. This matrix is ​​based on feature vectors that incorporate spatial information. It is generated so that it can reflect the different vertical evolution laws that wells in different spatial locations should follow, realizing the dynamic and personalized nature of geological constraints.

[0050] In one embodiment, after obtaining the fused feature vector and its specific dynamic geological transfer probability matrix for each well, a sequence decoding process is performed to determine the final sedimentary facies sequence of the well within the target interval as it varies with depth. This process integrates data-driven instantaneous facies identification with model-generated geological evolution constraints.

[0051] This process can be broken down into the following steps: Step S401: Emission Probability Calculation. This embodiment uses a decoder network to calculate the probability of each candidate sedimentary facies appearing at each depth sampling point; this probability is called the emission probability. The decoder can be a recurrent neural network (such as a Long Short-Term Memory network LSTM) or a simple multilayer perceptron. The decoder input includes two parts: the logging response vector at the current depth point, and the final feature vector of the well as the global, fused spatial information.

[0052] Specifically, for wells At depth point Well logging data and the final eigenvector of the well decoder Output a dimensional probability vector .

[0053] ; In the formula, The Each element represents the well depth point Based on the logging response and the overall geological background of the well, the probability that its sedimentary facies is of type m is determined. For well At depth point Well logging data; For well The final eigenvector. This probability is purely data-driven and reflects the instantaneous correspondence between well logging data and sedimentary facies types.

[0054] Step S402: Viterbi Optimal Path Decoding. To obtain a continuous facies sequence that conforms to both well logging data response and geological evolution patterns, this embodiment employs the Viterbi algorithm for optimal path decoding. This algorithm operates within a Hidden Markov Model framework, where the emission probability is provided by the decoder in step S401, while the state transition probability is generated by the dynamic geological transition probability matrix for the well in step S302. supply.

[0055] The goal of the Viterbi algorithm is to find a sequence of sedimentary facies. This sequence maximizes the joint probability of the entire well section. In logarithmic space, this optimization objective is equivalent to maximizing a cumulative score. The algorithm uses dynamic programming to calculate and record the cumulative probability (or logarithmic probability) of the optimal path reaching the current depth point and ending with each sedimentary facies, as well as the previous state of that path, starting from the first depth point of the well section.

[0056] Specifically, at depth point For candidate sedimentary facies The log probability of its optimal path The recursive calculation method is as follows: ; In the formula, For depth point sedimentary facies The logarithmic probability of the optimal path that terminates; To obtain from sedimentary facies Transition to sedimentary facies The logarithmic transition probability; For depth point Observed as sedimentary facies The logarithmic emission probability; This represents the total number of sedimentary phases.

[0057] After calculating the last depth point After scoring all optimal paths, the sedimentary facies with the highest score is selected as the final facies zone for that point. Then, using a backtracking algorithm, based on the recorded previous state information, the optimal facies zone for the previous depth point is determined sequentially from the last depth point backward, until the first depth point, thus constructing a complete, globally optimal sedimentary facies zone sequence. The implementation details of the Viterbi algorithm are well-known in the field and will not be elaborated upon here.

[0058] In one embodiment, after completing the sequence decoding of all target wells, the calculation results need to be formatted to generate digital results and visualizations that can be directly used by geological researchers.

[0059] This process can be broken down into the following steps: Step S501: Generate a standardized digital outcome file. The sequence of discrete sedimentary facies labels decoded in the preceding steps is integrated with the original depth sampling data to generate a structured digital outcome file. Specifically, the decoded facies band sequence is processed, merging consecutive identical facies band labels into a single sedimentary facies unit, and recording its starting depth (top boundary) and ending depth (bottom boundary). Finally, a data table is generated for each well, containing at least three columns: the top depth and bottom depth of the sedimentary facies unit, and the corresponding sedimentary facies type identifier. This outcome file can be stored in a common data exchange format, such as a CSV text file or a LAS file conforming to well logging industry standards, for subsequent software calls and data analysis.

[0060] Step S502: Generate visualized geological maps. Based on the digital output file generated in step S501, various visualized maps for geological analysis and evaluation can be generated. One type of map is a single-well sedimentary facies columnar section. This map uses well depth as the vertical axis, and draws the various sedimentary facies units divided by each well sequentially along the depth axis according to preset color or fill pattern codes, intuitively displaying the vertical sedimentary facies succession sequence of a single well. Another type of map is a multi-well sedimentary interconnected well profile. The generation process of this map is as follows: the sedimentary facies columnar sections of multiple wells are arranged according to the actual relative positions between wells, and sedimentary facies zones of the same type are connected, thereby constructing a geological map that can reflect the lateral distribution and superposition relationship of reservoirs in the profile direction. For the drawing of visualized maps, those skilled in the art can use existing commercial or open-source geological mapping software to complete the task. The specific implementation is a well-known technology in the field and will not be described in detail here.

[0061] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An automatic identification and division method for sedimentary facies zones of uranium reservoirs based on well logging data constraints, characterized in that, Includes the following steps: Construct a layered well location map with wells as nodes, and use a one-dimensional convolutional neural network encoder to extract the logging sequence data of each well to generate the initial feature vector of the node; The layered well location map is input into a graph attention network, and the initial feature vectors of the nodes are fused spatially to obtain a fused feature vector. The fused feature vector is input into the transfer matrix generation network to dynamically generate a geological transfer probability matrix for each well. The emission probability at each depth point is calculated based on the well logging data and the fused feature vector. Combining the emission probability and the geological transfer probability matrix, the Viterbi algorithm is used to decode the optimal sedimentary facies sequence as the partitioning result.

2. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 1, characterized in that, The one-dimensional convolutional neural network encoder is obtained through self-supervised pre-training; The self-supervised pre-training includes the following steps: Extracting logging sequence segments from unlabeled logging data; Data augmentation is performed on the well logging sequence segments to construct positive and negative sample pairs; The contrastive loss function is used to drive the encoder to bring positive sample pairs closer together and push negative sample pairs further apart in the feature space, thus completing self-supervised pre-training. The data augmentation includes at least one of Gaussian noise injection, sequence masking, or temporal warping.

3. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 1, characterized in that, The geological transition probability matrix is ​​a row-normalized square matrix, and the elements of the geological transition probability matrix represent the probability that one sedimentary facies will evolve into another sedimentary facies in the geological context of the corresponding well.

4. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 1, characterized in that, The construction of the layered well location map specifically includes the following steps: Each well within the work area is defined as a node; Calculate the horizontal distance between any two wells within the target formation unit; When the average horizontal distance is less than a preset distance threshold, an edge connection is established between the corresponding two nodes to construct a layered well location map.

5. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 1, characterized in that, The step of inputting the layered well location map into a graph attention network and fusing spatial features of the initial feature vectors of the nodes to obtain a fused feature vector specifically includes the following steps: The graph attention network calculates the attention coefficients based on the feature vectors of the center node and its neighboring nodes. The attention coefficients are used to perform a weighted summation of the feature vectors of neighboring nodes; Update the feature vector of the central node to generate the fused feature vector.

6. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 1, characterized in that, The process of dynamically generating a geological transition probability matrix for each well specifically includes the following steps: The fused feature vector corresponding to each well is used as the input to the transfer matrix generation network; Generate a geological transfer probability matrix that uniquely corresponds to each well.

7. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 1, characterized in that, The process of decoding the optimal sedimentary facies sequence using the Viterbi algorithm specifically includes the following steps: For each candidate sedimentary facies at each depth point within the well section, calculate an optimal path probability. The calculation steps include: Determine the path with the highest probability among all possible paths from any sedimentary facies at the previous depth point to a candidate sedimentary facies at the current depth point; The probability of the highest path is combined with the emission probability of the candidate sedimentary facies at the current depth point to obtain the optimal path probability of the current candidate sedimentary facies.

8. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 4, characterized in that, The preset distance threshold is set to 1.0-1.5 times the average width of the target deposit on the plane.

9. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 7, characterized in that, The calculation of the emission probability at each depth point specifically includes: The logging data at the current depth point is combined with the fused feature vector corresponding to the well and used as the input to a decoder network; The probability vector output by the decoder network is defined as the emission probability of the current depth point, and each element of the probability vector output by the decoder network corresponds to the probability value of the current depth point belonging to each candidate sedimentary phase.

10. The method for automatic identification and division of uranium reservoir sedimentary facies zones based on well logging data constraints according to claim 5, characterized in that, The calculation of the attention coefficient specifically includes the following steps: Perform linear transformations on the feature vectors of the central node and any of its neighboring nodes respectively; The two transformed feature vectors are concatenated, and an original attention score is calculated based on the concatenated vector. The original attention scores between the central node and all its corresponding neighboring nodes are normalized to obtain the final attention coefficients.

Images