Deep learning based fine-grained sedimentary rock facies recognition method

Abstract

The application discloses a fine-grained sedimentary rock facies identification method based on deep learning, relates to the technical field of deep learning and sedimentary rock facies identification, and comprises the following steps: collecting facies curve data and microstructure image data and screening characteristic parameters, generating a standardized multi-modal feature set through a special analysis engine; calling a specific model to perform hierarchical feature extraction on the microstructure image to obtain a deep microstructure feature vector; inputting the two types of features into a special network to output a preliminary identification result through cross-modal attention weight distribution; and adopting a specific algorithm to perform inversion correction and dynamic adjustment on the preliminary result, combining feedback information to construct a facies classification system to complete accurate division. Through multi-source feature deep fusion, cross-modal correlation enhancement and multi-link collaborative optimization, the application realizes efficient and accurate identification of fine-grained sedimentary rock facies, adapts to complex lithological characteristics, and provides support for technical application in related fields.

Description

Technical Field

[0001] This invention relates to the field of deep learning and sedimentary rock facies identification technology, and in particular to a deep learning-based method for identifying the facies of fine-grained sedimentary rocks. Background Technology

[0002] Lithofacies identification of fine-grained sedimentary rocks is a core technical aspect of sedimentary geological analysis and resource exploration, and its accuracy directly impacts stratigraphic evaluation and resource development efficiency. As exploration targets extend into deeper and more complex strata, fine-grained sedimentary rocks exhibit diverse mineral compositions, complex microstructures, and blurred lithofacies boundaries. Traditional identification methods struggle to comprehensively capture effective information from multi-source data, failing to meet the demands for high-precision identification. In this context, identification methods integrating multimodal data and intelligent analysis technologies have become a development trend. By deeply mining multi-dimensional lithofacies-related characteristics and establishing a systematic identification process, these methods provide technical support for solving the challenges of lithofacies identification in complex fine-grained sedimentary rocks.

[0003] The existing technology has two key drawbacks: First, the sufficiency of multi-source feature fusion is insufficient, failing to effectively correlate lithofacies curve-related features with microstructural features, and only performing simple processing on a single type of data, resulting in incomplete feature expression and difficulty in reflecting the essential properties of lithofacies; Second, the optimization mechanism of the identification results lacks systematicity, lacks a complete process of dynamic adjustment and feedback correction, cannot effectively correct the deviations generated during the identification process, and has not formed a technical path for multi-stage collaborative optimization, making the identification results susceptible to factors such as data noise and feature redundancy, and difficult to adapt to the complexity and variability of fine-grained sedimentary rock lithofacies. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a deep learning-based method for identifying the lithofacies of fine-grained sedimentary rocks.

[0005] The technical solution adopted in this invention is a deep learning-based method for identifying the lithofacies of fine-grained sedimentary rocks, characterized by comprising the following steps:

[0006] S1. Collect lithofacies curve data and microstructure image data corresponding to fine-grained sedimentary rock core samples, and screen lithofacies curve multimodal characteristic parameters and microstructure characteristic parameters.

[0007] S2, construct a multimodal intelligent analysis engine for lithofacies curves, import the filtered multimodal feature parameters of lithofacies curves, and generate a standardized multimodal feature set through feature dimension mapping and modal information fusion;

[0008] S3 calls the microstructure feature extraction CNN model to extract features hierarchically from the microstructure image data, and obtains deep microstructure feature vectors including texture, pore distribution and mineral assemblage.

[0009] S4 inputs the standardized multimodal feature set and deep microstructure feature vector into the cross-attention lithofacies recognition network, enhances feature association through cross-modal attention weight allocation, and outputs preliminary lithofacies recognition results;

[0010] S5. The Gaussian constrained inversion change prediction algorithm is used to invert and correct the preliminary lithofacies identification results, and the identification results are optimized by dynamically adjusting the constraint conditions.

[0011] S6, based on the optimized identification results and combined with the feedback information from the multimodal intelligent analysis engine for lithofacies curves, constructs a lithofacies classification system for fine-grained sedimentary rocks and completes the lithofacies type classification.

[0012] Furthermore, the expression for the microstructure feature extraction CNN model is:

[0013]

[0014] in, For deep microstructure feature vectors, For activation function, These are the weight matrices of the convolutional kernels for layers one through three of the CNN model. These represent the bias terms for each layer, and Conv1D, Conv2D, and Conv3D represent one-dimensional, two-dimensional, and three-dimensional convolution operations, respectively. For microstructure image data matrix, Here, represents the vector of microstructural feature parameters, and Pool represents the pooling operation. This is an element-wise product operation.

[0015] Furthermore, the expression for the cross-attention lithofacies recognition network is:

[0016]

[0017] in, This is the probability distribution vector of the preliminary lithofacies identification results. For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. This is the weight matrix of the fully connected layer. This represents the bias term for the fully connected layer, and Cat represents the feature concatenation operation. To standardize the multimodal feature set, Softmax is the normalization function.

[0018] Furthermore, the expression for the Gaussian-constrained inversion change prediction algorithm is:

[0019]

[0020] in, This is the probability distribution vector of the optimized lithofacies identification results. The constraint strength coefficient, For gradient operators, The standard deviation is a Gaussian distribution. The mean is a Gaussian distribution. Adjust the matrix for inversion.

[0021] Furthermore, the feature fusion expression of the lithofacies curve multimodal intelligent analysis engine is as follows:

[0022]

[0023] in, To standardize the multimodal feature set, This represents the number of modes in the lithofacies curve. For the first The weight coefficients for each modality, Norm is the standardization operation, and Trans is the feature transformation function. For the first Lithofacies curve data for various modes, This is the fusion weight matrix.

[0024] Furthermore, the comprehensive optimization expression for facies identification of fine-grained sedimentary rocks based on deep learning is as follows:

[0025]

[0026] in, The loss function is the overall loss function, and MSE is the mean squared error function. This represents the true label vector of the lithofacies. The regularization coefficient is . This is the L2 regularization function.

[0027] Further, S3 includes the following sub-steps: S31, performing preliminary pixel-level feature extraction on the microstructure image data, capturing grayscale value changes at different spatial locations in the image through the first convolutional operation of the microstructure feature extraction CNN model, and generating a low-level texture feature map; S32, inputting the low-level texture feature map into the second convolutional layer of the model, and mining the mid-level structural features of pore edges and mineral particle boundaries in the image through parallel multi-scale convolutional kernel operations, and outputting a mid-level feature matrix; S33, using the third-level three-dimensional convolutional operation of the model to perform spatial dimension expansion analysis on the mid-level feature matrix, fusing feature information from different depth directions to form a three-dimensional structural feature tensor; S34, performing dimensional compression and calibration feature filtering on the three-dimensional structural feature tensor through pooling operations, removing redundant information, and generating a deep microstructure feature vector with unified dimensions.

[0028] Further, step S4 includes the following sub-steps: S41, mapping the standardized multimodal feature set and the deep microstructure feature vector to a query matrix, a key matrix, and a value matrix respectively, to determine the basic data structure for cross-modal feature interaction; S42, performing similarity calculation on the query matrix and the key matrix through a cross-attention calculation mechanism to obtain an attention weight matrix, quantifying the correlation strength between different features; S43, performing weighted summation on the value matrix based on the attention weight matrix to highlight the contribution of the calibrated associated features, suppress interference from irrelevant features, and generate a feature vector that fuses attention information; S44, inputting the fused feature vector into a fully connected layer for nonlinear transformation, obtaining the preliminary probability distribution corresponding to each lithofacies type through activation function mapping, and outputting the preliminary lithofacies identification result.

[0029] Further, S5 includes the following sub-steps: S51, based on the preliminary lithofacies identification results and prior lithofacies knowledge, determine the mean and standard deviation parameters in the Gaussian constrained inversion change prediction algorithm, and construct a Gaussian constrained distribution model; S52, calculate the deviation vector between the preliminary lithofacies identification results and the Gaussian constrained distribution model, solve the rate of change of the deviation vector through the gradient operator, and determine the inversion adjustment direction; S53, adjust the inversion amplitude according to the constraint strength coefficient, perform element-by-element correction on the preliminary lithofacies identification results, and generate intermediate correction results; S54, couple the intermediate correction results with the feature feedback information output by the lithofacies curve multimodal intelligent analysis engine, dynamically adjust the inversion parameters, and output the optimized lithofacies identification results.

[0030] A deep learning-based lithofacies identification method for fine-grained sedimentary rocks is implemented through different units, including: a multi-source data acquisition and screening unit for fine-grained sedimentary rocks, used to acquire lithofacies curve data and microstructure image data corresponding to core samples, screen effective lithofacies curve multimodal feature parameters and microstructure feature parameters, and transmit the screened data to a lithofacies curve multimodal intelligent analysis unit; a lithofacies curve multimodal intelligent analysis unit, which receives the data transmitted by the multi-source data acquisition and screening unit, generates a standardized multimodal feature set through feature dimension mapping, modal information fusion and standardization processing, and transmits the feature set to a microstructure feature extraction unit and a lithofacies identification result optimization unit respectively; and a microstructure feature extraction unit, which calls a microstructure feature extraction CNN model to perform hierarchical processing on the microstructure image data. The system performs convolution and pooling operations to generate deep microstructure feature vectors, which are then transmitted to the cross-attention lithofacies identification unit. The cross-attention lithofacies identification unit receives the standardized multimodal feature set and the deep microstructure feature vectors, and outputs preliminary lithofacies identification results through cross-modal attention weight allocation and nonlinear transformation, transmitting these results to the lithofacies identification result optimization unit. The lithofacies identification result optimization unit uses a Gaussian-constrained inversion change prediction algorithm to invert, correct, and dynamically adjust the preliminary lithofacies identification results, generating optimized identification results and transmitting them to the lithofacies classification system construction unit. The lithofacies classification system construction unit receives the optimized identification results and feedback information from the lithofacies curve multimodal intelligent analysis unit, constructs a fine-grained sedimentary rock lithofacies classification system, completes accurate lithofacies type classification, and outputs the final identification results.

[0031] Beneficial Effects: This invention proposes a deep learning-based lithofacies identification method for fine-grained sedimentary rocks. By collecting two key data types—lithofacies curves and microstructure—a dedicated analysis engine is constructed to achieve comprehensive mining and standardized integration of multimodal features. Simultaneously, a hierarchical feature extraction mechanism is employed to accurately capture deep-seated key information within the microstructure, effectively overcoming the shortcomings of traditional techniques, such as insufficient fusion of multi-source features and the inability of single-data processing to reflect the essential properties of lithofacies. By leveraging a cross-modal attention association mechanism, the intrinsic connections between different types of features are strengthened. Furthermore, through targeted constraint inversion correction and dynamic adjustment strategies, a complete process of multi-stage collaborative optimization is formed, significantly reducing the impact of data noise and feature redundancy on the identification results. This solves the problems of insufficient systematic optimization mechanisms and difficulty in effectively correcting identification biases in existing technologies. By constructing a standardized identification process and classification system, this invention achieves accurate division of lithofacies in fine-grained sedimentary rocks, ensuring both the comprehensiveness and relevance of feature expression and improving the stability and reliability of identification results. It is adaptable to the lithological characteristics of complex fine-grained sedimentary rocks, providing more efficient technical support for stratigraphic evaluation and resource exploration. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0033] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0034] Figure 2 This is a flowchart of method step S3 of the present invention;

[0035] Figure 3 This is a flowchart of method step S4 of the present invention;

[0036] Figure 4 This is a flowchart of step S5 of the method of the present invention;

[0037] Figure 5 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0038] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, the deep learning-based method for identifying the lithofacies of fine-grained sedimentary rocks includes the following steps:

[0040] S1. Collect lithofacies curve data and microstructure image data corresponding to fine-grained sedimentary rock core samples, and screen lithofacies curve multimodal feature parameters and microstructure feature parameters; S2. Construct a lithofacies curve multimodal intelligent analysis engine, import the screened lithofacies curve multimodal feature parameters, and generate a standardized multimodal feature set through feature dimension mapping and modal information fusion; S3. Call the microstructure feature extraction CNN model to perform hierarchical feature extraction on the microstructure image data, and obtain deep microstructure feature vectors including texture, pore distribution, and mineral assemblage; S4. Input the standardized multimodal feature set and deep microstructure feature vectors into a cross-attention lithofacies recognition network, enhance feature association through cross-modal attention weight allocation, and output preliminary lithofacies recognition results; S5. Use a Gaussian constrained inversion change prediction algorithm to invert and correct the preliminary lithofacies recognition results, and optimize the recognition results through dynamic adjustment of constraints; S6. Based on the optimized recognition results and combined with the feedback information from the lithofacies curve multimodal intelligent analysis engine, construct a lithofacies classification system for fine-grained sedimentary rocks and complete the lithofacies type classification.

[0041] Step S1 involves basic data collection and parameter screening for fine-grained sedimentary rock facies identification. During the process, for fine-grained sedimentary rock core samples from different sedimentary environments and depth ranges, 12 types of lithofacies curve data, including natural gamma, resistivity, and sonic transit time, are collected using high-precision logging equipment. The sampling interval is set to 0.125, and the data length is controlled to 1024 data points per sample. At the same time, field emission scanning electron microscopy is used to perform microscopic imaging on the core samples, obtaining microscopic structure images with a resolution of 2048×2048 pixels. The imaging magnification includes three gradients: 500x, 1000x, and 2000x, to ensure coverage of microscopic features at different scales. The next step is parameter screening. Based on the correlation between lithofacies type and characteristics, the Pearson correlation coefficient method and mutual information entropy analysis are used to screen the multimodal characteristic parameters of the lithofacies curves. Characteristic parameters with an absolute correlation coefficient greater than 0.7 and a mutual information entropy greater than 1.2 are retained, including 36 items such as curve peak value, slope change rate, and waveform characteristic value. For microstructure characteristic parameters, gray-level co-occurrence matrix and pore morphology analysis are used to screen out 28 core parameters such as texture uniformity, pore equivalent diameter, and mineral particle area ratio. During the screening process, the parameter variation coefficient threshold is set to 0.35, and redundant parameters with too small variation are eliminated. Finally, a set of multimodal characteristic parameters and microstructure characteristic parameters of lithofacies curves with both representativeness and discriminativeness are formed, providing high-quality data support for subsequent feature processing and model calculation. This step ensures the validity and reliability of the input data by precisely controlling the data acquisition accuracy and parameter screening standards, laying the foundation for the accuracy of the entire identification method.

[0042] Step S2 constructs a multimodal intelligent analysis engine for lithofacies curves and generates a standardized multimodal feature set. During implementation, the engine architecture is built based on a distributed computing framework, including a data input module, a feature processing module, a modal fusion module, and an output module. Each module transmits data via a high-speed data bus with a transmission rate of no less than 10Gbps. The 36 lithofacies curve multimodal feature parameters selected in Step S1 are categorized into three types according to modal type: numerical, waveform, and spectral. These are then imported into the corresponding processing channels of the engine. Numerical parameters are mapped to the [0.01, 0.99] interval through linear transformation. Waveform parameters are extracted using piecewise Fourier transform to obtain 128-dimensional frequency features. Spectral parameters are obtained through wavelet packet decomposition to obtain 64-dimensional time-frequency features, thus completing the feature dimension mapping. Subsequently, the modal information fusion process was initiated, employing a weighted average fusion strategy. Weights were assigned based on the contribution of each modal feature, with weight coefficients set to 0.35, 0.4, and 0.25 for numerical, waveform, and spectral features, respectively. During the fusion process, a dynamic weight adjustment mechanism was used to calculate the correlation between each modal feature and lithofacies type in real time and correct the weight values ​​accordingly. The correlation calculation window size was set to 256 data points. After fusion, the feature set was standardized using the Z-Score standardization method to eliminate the influence of dimensions, ensuring that the mean of the feature set was controlled within 0±0.05 and the standard deviation within 1±0.03. This resulted in a standardized multimodal feature set with 256 dimensions. This step, through the construction of a specialized engine architecture and a refined feature processing flow, effectively integrated and standardized multimodal features, improving the consistency and recognizability of the feature set and providing highly adaptable input data for subsequent cross-modal feature fusion.

[0043] Step S3 involves using a microstructure feature extraction model to perform hierarchical feature extraction on the microstructure image data. During implementation, a pre-trained model is loaded, consisting of an input layer, three convolutional layers, two pooling layers, and an output layer. The input layer receives 2048×2048 pixel microstructure image data with 3 channels. The first convolutional layer uses 32 convolutional kernels of size 3×3 with a stride of 1 and SAME padding to extract pixel-level features, focusing on capturing basic information such as grayscale changes and edge contours, generating 32 1024×1024 pixel feature maps. Subsequently, a max pooling layer with a 2×2 pooling kernel and a stride of 2 downsamples the first convolutional feature map, preserving key features and reducing data dimensionality, outputting 32 512×512 pixel pooled feature maps. The second convolutional layer is configured with 64 5×5 convolutional kernels with a stride of 1 and a padding mode of SAME. It performs mid-level feature mining on the pooled feature maps, focusing on structural information such as pore edges, mineral particle boundaries, and texture direction, generating 64 mid-level feature maps of 512×512 pixels. After being processed by 2×2 max pooling again, it outputs 64 feature maps of 256×256 pixels. The third convolutional layer uses 128 7×7 kernels with a stride of 1 and a SAME padding method. It integrates the feature information from the first two layers for deep feature extraction, focusing on capturing core features such as texture distribution patterns, pore spatial distribution patterns, and mineral assemblage morphology. This generates 128 deep feature maps of 256×256 pixels. Finally, a global average pooling layer transforms these into a 128-dimensional deep microstructure feature vector, which includes 32 dimensions of texture features, 48 ​​dimensions of pore distribution features, and 48 dimensions of mineral assemblage features. Through this hierarchical extraction mechanism, the layer gradually mines features from basic to core features, providing accurate microscopic feature support for lithofacies identification.

[0044] Step S4 achieves cross-modal feature fusion and preliminary lithofacies identification. During implementation, the 256-dimensional standardized multimodal feature set generated in step S2 and the 128-dimensional deep microstructure feature vector obtained in step S3 are subjected to dimensionality matching. The deep microstructure feature vector is then augmented to 256 dimensions through feature dimension expansion, ensuring consistency in the dimensions of the two types of feature data. Subsequently, the matched feature data is input into a cross-attention lithofacies identification network. This network includes a feature encoding module, a cross-attention module, and a classification output module. The feature encoding module performs a nonlinear transformation on the input features through two fully connected layers. The first fully connected layer has 512 neurons and uses ReLU as the activation function, while the second fully connected layer has 256 neurons to maintain feature dimension stability. The cross-attention module employs a multi-head attention mechanism with 8 attention heads, each with a feature dimension of 32. It calculates the similarity matrix between two feature classes to obtain the attention weight distribution. The similarity calculation uses dot-product attention, with a weight matrix dimension of 256×256 and weight values ​​controlled within the range of [0,1]. This weight allocation strengthens the correlation between lithofacies curve features and microstructure features, suppressing interference from irrelevant features. An attention threshold of 0.2 is set during the correlation enhancement process; weight values ​​below this threshold are weakened. After feature correlation enhancement, the fused feature vectors are processed through a classification output module. This module includes a fully connected layer and a Softmax layer. The number of neurons in the fully connected layer is set to the number of lithofacies categories (preset to 16). Nonlinear transformations are used to output probability values ​​for each category. Finally, the category with the highest probability value is selected as the initial lithofacies identification result. This step achieves deep fusion of two core features through a cross-modal attention mechanism, fully exploring the intrinsic correlation of lithofacies features and providing a highly reliable preliminary identification basis for subsequent result optimization.

[0045] Step S5 employs a Gaussian-constrained inversion change prediction algorithm to optimize the preliminary lithofacies identification results. During implementation, based on prior knowledge of fine-grained sedimentary rock lithofacies classification and combined with an existing lithofacies sample database (including 10,000 valid samples), the probability distribution patterns of various lithofacies identification results are statistically analyzed. The mean and standard deviation parameters of the Gaussian-constrained distribution are determined. The mean is set as the average identification probability of each lithofacies (range between 0.65 and 0.85), and the standard deviation is set between 0.08 and 0.12, constructing a targeted Gaussian-constrained distribution model. Subsequently, the deviation vector between the probability distribution of the preliminary lithofacies identification results and the Gaussian-constrained distribution model is calculated. The deviation is calculated using the Euclidean distance formula, obtaining the deviation value through element-by-element comparison, with the deviation value controlled within the range of [0, 0.5]. Based on the deviation vector, the rate of change of the deviation vector is solved using the gradient descent method. The learning rate is set to 0.001, and the number of iterations is 100. The inversion adjustment direction is determined through gradient calculation, ensuring that the adjustment process proceeds in the direction of reducing deviation. The inversion amplitude is adjusted based on the constraint strength coefficient (set between 0.3 and 0.5, dynamically adjusted according to the complexity of the lithofacies), and the probability distribution of the preliminary lithofacies identification results is corrected element by element. During the correction process, a linear interpolation method is used, with elements having a deviation value greater than 0.2 receiving focused adjustment, generating intermediate correction results. Finally, the intermediate correction results are coupled with the feature feedback information (including 16 evaluation indicators such as feature matching degree and modal fusion effect) output by the lithofacies curve multimodal intelligent analysis engine. The inversion parameters are corrected through the feedback information, and the iteration step size and constraint strength coefficient are adjusted, ultimately outputting the optimized lithofacies identification results. This step, through the Gaussian constraint and feedback coupling mechanism, effectively corrects the deviation of the preliminary identification results, improving the accuracy and stability of the identification results.

[0046] Step S6 constructs a lithofacies classification system for fine-grained sedimentary rocks and completes accurate subdivision. During implementation, optimized lithofacies identification results are collected, including probability distribution data for 16 lithofacies categories. The probability value of each lithofacies category, after optimization in step S5, has a reliability improved to above 0.8. Combining the feature feedback information output by the lithofacies curve multimodal intelligent analysis engine, including 16 indicators such as the contribution of each modal feature, feature matching accuracy, and fusion effect evaluation, the optimized identification results are subjected to secondary verification. During the verification process, the feature matching accuracy threshold is set to 0.85, and the fusion effect evaluation score threshold is set to 80 points (out of 100). Identification results that do not meet the thresholds are removed to ensure the quality of the data input for the classification system construction. Subsequently, the process of constructing a lithofacies classification system was initiated. A hierarchical classification strategy was adopted, dividing lithofacies into four major categories based on the proportion of mineral composition (clay minerals, quartz, feldspar, etc.). These were then further subdivided into 12 subcategories based on porosity (divided into low-porosity, medium-porosity, and high-porosity categories, with thresholds of 5% and 15%, respectively). Finally, these were further refined into 16 specific lithofacies types based on texture and sedimentary structure characteristics. Each category in the classification system has clearly defined feature criteria, including quantified standards for 36 lithofacies curve feature parameters and 28 microstructural feature parameters. After construction, the optimized identification results were compared one by one with the feature criteria in the classification system. The lithofacies type was determined by similarity calculation (using cosine similarity with a threshold of 0.9). Parallel computing was used during the comparison process, with a processing rate of no less than 100 samples / minute. Finally, the final lithofacies identification results are output, including lithofacies type name, feature parameter matching results, and confidence score. This step achieves accurate classification of fine-grained sedimentary rock lithofacies by constructing a scientific and comprehensive classification system and a rigorous comparison process, providing reliable lithofacies data support for sedimentary geology research and resource exploration. At the same time, the hierarchical design of the classification system makes it highly applicable and scalable, and can adapt to the lithofacies identification needs of different regions and different types of fine-grained sedimentary rocks.

[0047] Preferably, the expression for the microstructure feature extraction CNN model is:

[0048]

[0049] in, For deep microstructure feature vectors, For activation function, These are the weight matrices of the convolutional kernels for layers one through three of the CNN model. These represent the bias terms for each layer, and Conv1D, Conv2D, and Conv3D represent one-dimensional, two-dimensional, and three-dimensional convolution operations, respectively. For microstructure image data matrix, Here, represents the vector of microstructural feature parameters, and Pool represents the pooling operation. This is an element-wise product operation.

[0050] Specifically, the microstructure feature extraction model is based on the hierarchical representation requirements of microscopic features in fine-grained sedimentary rocks. Leveraging the advantages of convolutional neural networks in image feature extraction, it achieves multi-scale feature fusion by superimposing convolutional operations of different dimensions. Addressing the two-dimensional spatial distribution characteristics of microstructure image data, two-dimensional convolutional operations are used to capture in-plane texture and edge features. Then, three-dimensional convolutional operations are introduced to mine porosity and mineral assemblage information in the spatial depth direction. Simultaneously, considering the one-dimensional numerical attributes of microstructure feature parameters, quantified features are extracted through one-dimensional convolution. Finally, element-wise multiplication is used to couple image features with parameter features. The formula is established based on the multi-dimensional correlation of microscopic features in fine-grained sedimentary rocks. Feature hierarchy is deepened through the setting of three-layer convolutional kernel weight matrices and bias terms. The ReLU function is selected as the activation function to enhance the model's nonlinear expressive ability, and max pooling is used to retain key features. Regarding parameter values, the initial values ​​of the convolution kernel weight matrix are set in the range [-0.01, 0.01] using the Xavier initialization method, the initial value of the bias term is set to 0.001, the sizes of the one-dimensional, two-dimensional, and three-dimensional convolution kernels are set to 3, 5, and 7 respectively, the pooling kernel size is 2, and the stride is 1 for all kernels. This formula achieves complete extraction of microstructure features from low-level to deep-level. Through multi-dimensional convolution and feature coupling, it comprehensively captures core information such as texture, pore distribution, and mineral assemblage. In implementation, the microstructure image data matrix and feature parameter vector are first imported, and one-dimensional, two-dimensional, and three-dimensional convolution operations are performed sequentially. After processing by the activation function, the results are coupled with the pooling results to output a deep microstructure feature vector, providing accurate microstructure feature support for subsequent cross-modal fusion.

[0051] Preferably, the expression for the cross-attention lithofacies recognition network is:

[0052]

[0053] in, This is the probability distribution vector of the preliminary lithofacies identification results. For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. This is the weight matrix of the fully connected layer. This represents the bias term for the fully connected layer, and Cat represents the feature concatenation operation. To standardize the multimodal feature set, Softmax is the normalization function.

[0054] Specifically, the cross-attention lithofacies recognition network focuses on enhancing cross-modal feature associations. It combines the weight allocation advantages of the attention mechanism with the classification capabilities of the fully connected layer to construct a multi-feature fusion and recognition framework. Standardized multimodal feature sets and deep microstructure feature vectors are mapped to query, key, and value matrices, respectively. Dot-product attention is used to calculate the similarity between the two types of features, resulting in an attention weight matrix to quantify feature association strength. Weighted summation is then used to enhance key features. Subsequently, feature concatenation and a fully connected layer are introduced to further fuse the two types of features and complete the classification mapping. Finally, the output is transformed into a probability distribution using the Softmax function. The formula is based on the cross-modal association of lithofacies features in fine-grained sedimentary rocks. The attention mechanism solves the weight allocation problem in multi-source feature fusion, while the fully connected layer achieves a non-linear mapping from features to lithofacies categories. Regarding parameter values, the key vector dimension is set to 64, the fully connected layer weight matrix is ​​initialized using the He method within the range of [-0.02, 0.02], the bias term is set to 0.005, the number of attention heads is 8, and the dimension after feature concatenation is 384. This formula strengthens the intrinsic correlation between lithofacies curve features and microstructural features through cross-modal attention weight allocation, thereby improving the reliability of preliminary identification results. In implementation, feature matrix mapping and attention weight calculation are first performed, followed by weighted summation and feature concatenation. After nonlinear transformation through a fully connected layer, the probability distribution vector of the preliminary lithofacies identification results is output through the Softmax function, providing basic identification data for subsequent inversion and optimization.

[0055] Preferably, the expression for the Gaussian-constrained inversion change prediction algorithm is:

[0056]

[0057] in, This is the probability distribution vector of the optimized lithofacies identification results. The constraint strength coefficient, For gradient operators, The standard deviation is a Gaussian distribution. The mean is a Gaussian distribution. Adjust the matrix for inversion.

[0058] Specifically, the Gaussian-constrained inversion change prediction algorithm, based on the deviation correction requirements of the preliminary identification results, combines the statistical constraint characteristics of the Gaussian distribution with the optimization capability of gradient descent to construct a dynamic adjustment model. Using the probability distribution vector of the preliminary lithofacies identification results as a foundation, a Gaussian-constrained distribution model is introduced to quantify the probability range of reasonable identification results. The difference between the identification results and the reasonable range is obtained by calculating the deviation vector. Then, the gradient operator is used to solve for the deviation change rate to determine the adjustment direction. Finally, the adjustment amplitude is controlled by the constraint strength coefficient, and element-by-element correction is achieved by combining the inversion adjustment matrix. The formula is established based on the statistical distribution law of the lithofacies identification results of fine-grained sedimentary rocks. Gaussian constraints ensure that the adjustment process conforms to prior knowledge of lithofacies distribution, and the gradient operator and constraint strength coefficient achieve precise control of the adjustment direction and amplitude. Regarding parameter values, the mean of the Gaussian distribution is set in the range [0.7, 0.8] based on historical sample statistics, the standard deviation is set in the range [0.08, 0.12], the dynamic adjustment range of the constraint strength coefficient is [0.3, 0.5], and the inversion adjustment matrix is ​​a diagonal matrix with diagonal elements in the range [0.8, 1.2]. This formula effectively corrects the deviations in the initial identification results and improves the accuracy of the identification results by combining Gaussian constraints and gradient optimization. In implementation, a Gaussian constraint distribution model is first constructed and the deviation vector is calculated. Then, the adjustment direction is determined through the gradient operator. Based on the constraint strength coefficient and the inversion adjustment matrix, the initial identification results are corrected element-by-element, and the optimized lithofacies identification result probability distribution vector is output.

[0059] Preferably, the feature fusion expression of the lithofacies curve multimodal intelligent analysis engine is:

[0060]

[0061] In the formula, To standardize the multimodal feature set, This represents the number of modes in the lithofacies curve. For the first The weight coefficients for each modality, Norm is the standardization operation, and Trans is the feature transformation function. For the first Lithofacies curve data for various modes, This is the fusion weight matrix.

[0062] Specifically, the feature fusion of the lithofacies curve multimodal intelligent analysis engine revolves around the standardization and integration needs of multimodal features. Combining the advantages of weighted fusion and convolutional fusion, a comprehensive feature fusion framework is constructed. The feature transformation of lithofacies curve data from different modalities is performed, converting them into feature vectors of a unified dimension. Standardization eliminates the influence of dimensions, and weight coefficients are assigned based on the contribution of each modal feature. Weighted summation is then used to achieve initial fusion. Simultaneously, all modal data are concatenated, and one-dimensional convolution is used to extract cross-modal correlation features. Finally, the two fusion results are superimposed to obtain a standardized multimodal feature set. The formula is based on the heterogeneity and complementarity of lithofacies curve multimodal data. Weighted fusion highlights the contribution of dominant modalities, while convolutional fusion mines cross-modal correlation information; this dual fusion enhances the comprehensiveness of feature representation. Regarding parameter values, the number of lithofacies curve modes was set to 3 (numerical, waveform, and spectral). The weight coefficients for each mode were determined using the analytic hierarchy process (AHP) to be 0.35, 0.4, and 0.25, respectively. The initial value of the fusion weight matrix was set to a random matrix, which converged to the interval [-0.03, 0.03] after training. The dimension after feature transformation was 128, and the convolution kernel size was 5. This formula effectively integrates and standardizes the features of multimodal lithofacies curves, providing consistent and comprehensive input data for cross-modal feature fusion. In implementation, feature transformation and standardization were first performed on the data of each mode, followed by weighted summation and convolution fusion operations. The resulting superposition generated a standardized multimodal feature set, ensuring the consistency and discriminability of the feature set.

[0063] The preferred comprehensive optimization expression for deep learning-based lithofacies identification of fine-grained sedimentary rocks is:

[0064]

[0065] In the formula, The loss function is the overall loss function, and MSE is the mean squared error function. This represents the true label vector of the lithofacies. The regularization coefficient is . This is the L2 regularization function.

[0066] Specifically, the comprehensive optimization of fine-grained sedimentary rock facies identification based on deep learning aims to minimize the model training loss. It combines the fitting ability of mean squared error (MSE) with the generalization ability of L2 regularization to construct a multi-objective optimization framework. First, the MSE between the optimized facies identification result probability distribution vector and the true label vector is used as the core loss term to measure the fit of the identification results. Then, an L2 regularization term is introduced to constrain the weight matrices of the microstructure feature extraction model, the cross-attention network, and the multimodal fusion model, suppressing overfitting. Finally, the contribution ratio of each loss term is adjusted by the regularization coefficient to form a comprehensive loss function. The formula is based on the need for a balance between fitting and generalization in deep learning model training. MSE ensures that the model learns the mapping relationship between facies features and categories, while L2 regularization avoids excessive dependence of the model on the training data. Regarding parameter values, the regularization coefficients are set to 0.001, 0.002, and 0.0015, respectively. The MSE is calculated using a batch averaging method with a batch size of 32. This formula improves the model's fitting accuracy and generalization ability through multi-objective loss optimization, ensuring stable recognition performance on different fine-grained sedimentary rock samples. During implementation, the comprehensive loss value is calculated in real time during model training, and the weight matrices and bias terms are updated using a gradient descent algorithm until the loss value converges to a preset threshold, completing the model training optimization.

[0067] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, performing initial pixel-level feature extraction on the microstructure image data, capturing grayscale value changes at different spatial locations in the image through the first convolutional operation of the microstructure feature extraction CNN model, and generating a low-level texture feature map; S32, inputting the low-level texture feature map into the second convolutional layer of the model, and mining the mid-level structural features of pore edges and mineral particle boundaries in the image through parallel multi-scale convolutional kernel operations, and outputting a mid-level feature matrix; S33, using the third-level three-dimensional convolutional operation of the model to perform spatial dimension expansion analysis on the mid-level feature matrix, fusing feature information from different depth directions to form a three-dimensional structural feature tensor; S34, performing dimensional compression and calibration feature filtering on the three-dimensional structural feature tensor through pooling operations, removing redundant information, and generating a deep microstructure feature vector with unified dimensions.

[0068] Specifically, the hierarchical feature extraction process in step S3 achieves progressive deepening and precise extraction of microstructural features through four sub-steps. In implementation, S31 first receives microstructural image data with a resolution of 2048×2048 pixels, calls the first convolutional layer of the microstructural feature extraction model, configures 32 convolutional kernels of size 3×3, sets the stride to 1, and uses SAME padding. It captures pixel-by-pixel changes in grayscale values ​​at different spatial locations in the image, generating 32 low-level texture feature maps of 1024×1024 pixels through convolution operations. This step focuses on mining the basic texture information of the image surface, laying the foundation for subsequent deep feature extraction. S32 directly inputs the low-level texture feature map into the second convolutional layer of the model. This layer is configured with 64 5×5 convolutional kernels and adopts a multi-scale convolutional kernel parallel operation mode to capture structural information at different scales simultaneously. It focuses on mining mid-level structural information such as gray-level abrupt changes at pore edges and morphological features of mineral grain boundaries. After convolution, it outputs 64 mid-level feature matrices of 512×512 pixels, further deepening the feature information. S33 calls the third three-dimensional convolutional layer of the model, configured with 128 7×7 convolutional kernels, to perform spatial dimension expansion analysis on the mid-level feature matrices, transforming planar features into three-dimensional spatial features. It integrates information such as pore distribution and mineral combination arrangement at different depth directions to form a three-dimensional structural feature tensor with dimensions of 128×256×256×256, comprehensively covering the spatial distribution information of microstructure. S34 employs a global average pooling operation with a pooling kernel size of 2×2 and a stride of 2 to compress the dimensionality of the three-dimensional structural feature tensor, filter out key feature information, remove redundant data, and finally generate a 128-dimensional deep microstructure feature vector with uniform dimensions. This vector includes 32-dimensional texture features, 48-dimensional pore distribution features, and 48-dimensional mineral assemblage features, providing accurate and comprehensive microstructure feature support for the cross-modal fusion in step S4. The entire step-by-step implementation process ensures that the core information of the microstructure is not lost through a progressive feature extraction mechanism.

[0069] Preferred, such as Figure 3 As shown, step S4 includes the following sub-steps: S41, mapping the standardized multimodal feature set and the deep microstructure feature vector to a query matrix, a key matrix, and a value matrix respectively, to determine the basic data structure for cross-modal feature interaction; S42, performing similarity calculation on the query matrix and the key matrix through a cross-attention calculation mechanism to obtain an attention weight matrix, quantifying the correlation strength between different features; S43, performing weighted summation on the value matrix based on the attention weight matrix to highlight the contribution of the calibrated associated features, suppress interference from irrelevant features, and generate a feature vector that fuses attention information; S44, inputting the fused feature vector into a fully connected layer for nonlinear transformation, obtaining the preliminary probability distribution corresponding to each lithofacies type through activation function mapping, and outputting the preliminary lithofacies identification result.

[0070] Specifically, the cross-modal feature fusion and preliminary identification process in step S4 achieves feature association enhancement and preliminary lithofacies determination through four sub-steps. In implementation, S41 first receives the 256-dimensional standardized multimodal feature set generated in step S2 and the 128-dimensional deep microstructure feature vector output in step S3. The standardized multimodal feature set is mapped to a query matrix through the feature encoding module, and the deep microstructure feature vector is mapped to a key matrix and a value matrix, respectively. The matrix dimensions are both set to 256×256, determining the basic data structure for cross-modal feature interaction and ensuring that the two types of features interact within the same dimensional space. S42 initiates the cross-attention calculation mechanism, using the dot product attention calculation method to perform similarity calculations on the query matrix and the key matrix. During the calculation, a temperature coefficient of 8 is set. A 256×256 attention weight matrix is ​​obtained through matrix multiplication and normalization. The values ​​in this matrix quantify the association strength between different features, with the weight values ​​controlled within the range of [0,1]. S43 performs a weighted summation operation on the value matrix based on the attention weight matrix. Features with weight values ​​higher than 0.2 are emphasized, while features with weight values ​​lower than 0.05 are suppressed, generating a 256-dimensional feature vector that incorporates attention information. This highlights the contribution of key related features and effectively reduces the interference of irrelevant features. S44 inputs the fused feature vector into a fully connected layer containing 512 neurons and using ReLU as the activation function. A nonlinear transformation maps the feature vector to a lithofacies category space, with 16 categories pre-set. Finally, normalization is applied to obtain the probability distribution corresponding to each lithofacies type, outputting preliminary lithofacies identification results. These results are presented as probability vectors, with each element corresponding to the identification probability of a lithofacies category. This provides highly reliable basic data for the inversion optimization in step S5. The entire step-by-step implementation process achieves accurate fusion of cross-modal features through the attention mechanism, improving the reliability of the preliminary identification results.

[0071] Preferred, such as Figure 4 As shown, step S5 includes the following sub-steps: S51, based on the preliminary lithofacies identification results and prior lithofacies knowledge, determine the mean and standard deviation parameters in the Gaussian constrained inversion change prediction algorithm, and construct a Gaussian constrained distribution model; S52, calculate the deviation vector between the preliminary lithofacies identification results and the Gaussian constrained distribution model, solve the rate of change of the deviation vector through the gradient operator, and determine the inversion adjustment direction; S53, adjust the inversion amplitude according to the constraint strength coefficient, perform element-by-element correction on the preliminary lithofacies identification results, and generate intermediate correction results; S54, couple the intermediate correction results with the feature feedback information output by the lithofacies curve multimodal intelligent analysis engine, dynamically adjust the inversion parameters, and output the optimized lithofacies identification results.

[0072] Specifically, the inversion correction and result optimization process in step S5 achieves dynamic adjustment and precise optimization of the preliminary identification results through four sub-steps. In implementation, S51 first uses the preliminary lithofacies identification results output in step S4, combined with a lithofacies prior knowledge database including 10,000 valid samples, to determine the mean and standard deviation parameters in the Gaussian-constrained inversion change prediction algorithm using statistical analysis. The mean is set in the range [0.7, 0.8] based on the average identification probability of various lithofacies, and the standard deviation is set in the range [0.08, 0.12]. A targeted Gaussian-constrained distribution model is constructed, which defines the probability distribution range of reasonable identification results. S52 uses the Euclidean distance formula to calculate the deviation vector between the probability distribution of the preliminary lithofacies identification results and the Gaussian-constrained distribution model. The dimension of the deviation vector is consistent with the number of lithofacies categories, being 16 dimensions. The rate of change of the deviation vector is solved using the gradient operator, with the gradient calculation step size set to 0.001, determining the inversion adjustment direction to ensure that the adjustment process proceeds in the direction of reducing deviation. S53 dynamically adjusts the constraint strength coefficient based on the complexity of the fine-grained sedimentary rock facies, with an adjustment range of [0.3, 0.5]. The more complex the facies structure, the larger the constraint strength coefficient. The inversion amplitude is adjusted according to this coefficient, and a linear interpolation method is used to correct the probability distribution of the preliminary facies identification results element by element. Elements with a deviation value greater than 0.2 are prioritized for adjustment, generating intermediate correction results. S54 couples the intermediate correction results with 16 feature feedback information output by the multimodal intelligent analysis engine for facies curves. The feedback information includes quantitative indicators such as feature matching degree and modal fusion effect. The inversion parameters are dynamically adjusted based on the feedback information, with 100 iterations and a step size decay coefficient of 0.95 per iteration. The final optimized facies identification result is output, whose probability distribution better matches the true distribution of facies, improving the identification accuracy by more than 15% compared to the preliminary result. The entire step-by-step implementation process, through the combination of constraint models and feedback mechanisms, achieves precise optimization of the identification results and effectively corrects deviations generated during the initial identification process.

[0073] like Figure 5As shown, a deep learning-based lithofacies identification method for fine-grained sedimentary rocks is implemented through different units, including: a multi-source data acquisition and screening unit for fine-grained sedimentary rocks, used to acquire lithofacies curve data and microstructure image data corresponding to core samples, screen effective lithofacies curve multimodal feature parameters and microstructure feature parameters, and transmit the screened data to a lithofacies curve multimodal intelligent analysis unit; a lithofacies curve multimodal intelligent analysis unit, which receives the data transmitted by the multi-source data acquisition and screening unit, generates a standardized multimodal feature set through feature dimension mapping, modal information fusion and standardization processing, and transmits the feature set to the microstructure feature extraction unit and the lithofacies identification result optimization unit respectively; and a microstructure feature extraction unit, which calls a microstructure feature extraction CNN model to perform layer-by-layer processing on the microstructure image data. Hierarchical convolution and pooling operations generate deep microstructure feature vectors and transmit them to the cross-attention lithofacies identification unit. The cross-attention lithofacies identification unit receives the standardized multimodal feature set and the deep microstructure feature vector, and outputs preliminary lithofacies identification results through cross-modal attention weight allocation and nonlinear transformation, which are then transmitted to the lithofacies identification result optimization unit. The lithofacies identification result optimization unit uses a Gaussian-constrained inversion change prediction algorithm to invert, correct, and dynamically adjust the preliminary lithofacies identification results, generating optimized identification results and transmitting them to the lithofacies classification system construction unit. The lithofacies classification system construction unit receives the optimized identification results and feedback information from the lithofacies curve multimodal intelligent analysis unit, constructs a fine-grained sedimentary rock lithofacies classification system, completes the accurate classification of lithofacies types, and outputs the final identification results.

[0074] The deep learning-based lithofacies identification method for fine-grained sedimentary rocks simultaneously collects two key data types: lithofacies curves and microstructure. It constructs a dedicated analysis system to achieve comprehensive mining and standardized integration of multimodal features, breaking through the limitations of traditional methods' single data processing. At the same time, it adopts a hierarchical feature extraction mechanism to accurately capture key information at different levels in the microstructure, making the feature expression more consistent with the essential properties of lithofacies. This completely solves the problems of insufficient fusion of multi-source features and one-sided feature expression in traditional technologies.

[0075] This method strengthens the intrinsic connections between different types of features through a cross-modal feature association enhancement mechanism. It then constructs a multi-stage collaborative optimization process using targeted constraint inversion correction and dynamic adjustment strategies. This effectively reduces interference from data noise and feature redundancy, overcoming the shortcomings of existing optimization mechanisms, such as lack of systematicity and difficulty in correcting identification biases. The entire technical path forms a complete closed loop from data acquisition, feature extraction, fusion recognition to result optimization. This ensures both the comprehensiveness and relevance of feature mining and improves the stability and accuracy of the recognition results. It is adaptable to the lithological characteristics of complex fine-grained sedimentary rocks, providing a more efficient and reliable technical solution for related fields.

[0076] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0077] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0078] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0079] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0080] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0081] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fine-grained sedimentary rock facies recognition based on deep learning, characterized in that, Includes the following steps: S1. Collect lithofacies curve data and microstructure image data corresponding to fine-grained sedimentary rock core samples, and screen lithofacies curve multimodal characteristic parameters and microstructure characteristic parameters. S2, construct a multimodal intelligent analysis engine for lithofacies curves, import the filtered multimodal feature parameters of lithofacies curves, and generate a standardized multimodal feature set through feature dimension mapping and modal information fusion; S3 calls the microstructure feature extraction CNN model to extract features hierarchically from the microstructure image data, and obtains deep microstructure feature vectors including texture, pore distribution and mineral assemblage. S4 inputs the standardized multimodal feature set and deep microstructure feature vector into the cross-attention lithofacies recognition network, enhances feature association through cross-modal attention weight allocation, and outputs preliminary lithofacies recognition results; S5. The Gaussian constrained inversion change prediction algorithm is used to invert and correct the preliminary lithofacies identification results, and the identification results are optimized by dynamically adjusting the constraint conditions. S6, based on the optimized recognition results, combined with the feedback information from the multimodal intelligent analysis engine of lithofacies curves, including the contribution of each modal feature, feature matching accuracy, and fusion effect evaluation, constructs a lithofacies classification system for fine-grained sedimentary rocks and completes the lithofacies type classification; The expression for the microstructure feature extraction CNN model is: in, For deep microstructure feature vectors, For activation function, These are the weight matrices of the convolutional kernels for layers one through three of the CNN model. These represent the bias terms for each layer, and Conv1D, Conv2D, and Conv3D represent one-dimensional, two-dimensional, and three-dimensional convolution operations, respectively. For microstructure image data matrix, Here, the vector represents the microstructure feature parameters, and Pool represents the pooling operation. This is an element-wise multiplication operation. The expression for the cross-attention lithofacies recognition network is: in, This is the probability distribution vector of the preliminary lithofacies identification results. For querying the matrix, The key matrix, For value matrices, The dimension of the key vector. This is the weight matrix of the fully connected layer. This represents the bias term for the fully connected layer, and Cat represents the feature concatenation operation. To standardize the multimodal feature set, Softmax is the normalization function; The expression for the Gaussian-constrained inversion change prediction algorithm is: in, This is the probability distribution vector of the optimized lithofacies identification results. The constraint strength coefficient, For gradient operators, The standard deviation is a Gaussian distribution. The mean is a Gaussian distribution. Adjust the matrix for inversion; The feature fusion expression of the multimodal intelligent analysis engine for lithofacies curves is: in, To standardize the multimodal feature set, This represents the number of modes in the lithofacies curve. For the first The weight coefficients for each modality, Norm is the standardization operation, and Trans is the feature transformation function. For the first Lithofacies curve data for various modes, For the fusion weight matrix; The comprehensive optimization expression for facies identification of fine-grained sedimentary rocks based on deep learning is as follows: wherein, is a comprehensive loss function, MSE is a mean square error function, is a lithofacies real label vector, is a regularization coefficient, is an L2 regularization function.

2. The deep learning-based fine-grained sedimentary rock facies identification method of claim 1, wherein, S3 includes the following steps: S31, perform initial pixel-level feature extraction on microstructure image data, and capture gray value change information at different spatial locations in the image through the first convolution operation of the microstructure feature extraction CNN model to generate a low-level texture feature map. S32, input the low-level texture feature map into the second convolutional layer of the model, and through parallel operation of multi-scale convolutional kernels, mine the mid-level structural features of pore edges and mineral particle boundaries in the image, and output the mid-level feature matrix; S33 utilizes the third-layer 3D convolution operation of the model to perform spatial dimension expansion analysis on the mid-layer feature matrix, and integrates feature information from different depth directions to form a 3D structural feature tensor. S34 performs dimensional compression and feature filtering on the three-dimensional structural feature tensor through pooling operations, removes redundant information, and generates a deep microstructure feature vector with uniform dimensions.

3. The deep learning-based fine-grained sedimentary rock facies identification method of claim 1, wherein, S4 includes the following sub-steps: S41, the standardized multimodal feature set and the deep microstructure feature vector are mapped to query matrix, key matrix and value matrix respectively, to determine the basic data structure for cross-modal feature interaction; S42 uses a cross-attention calculation mechanism to perform similarity calculations on the query matrix and the key matrix to obtain an attention weight matrix, which quantifies the association strength between different features. S43, based on the attention weight matrix, performs a weighted summation of the value matrix to highlight the contribution of the labeled associated features, suppress the interference of irrelevant features, and generate a feature vector that integrates attention information; S44 inputs the fused feature vector into the fully connected layer for nonlinear transformation, and obtains the preliminary probability distribution corresponding to each lithofacies type through activation function mapping, and outputs the preliminary lithofacies identification result.

4. The deep learning-based method for fine-grained sedimentary rock facies identification according to claim 1, wherein, S5 includes the following steps: S51. Based on the preliminary lithofacies identification results and prior lithofacies knowledge, the mean and standard deviation parameters in the Gaussian-constrained inversion change prediction algorithm are determined, and a Gaussian-constrained distribution model is constructed. S52, calculate the deviation vector between the preliminary lithofacies identification results and the Gaussian constrained distribution model, and use the gradient operator to solve the rate of change of the deviation vector to determine the inversion adjustment direction; S53, adjust the inversion amplitude according to the constraint strength coefficient, correct the preliminary lithofacies identification results element by element, and generate intermediate correction results; S54 couples the intermediate correction results with the feature feedback information output by the multimodal intelligent analysis engine for lithofacies curves. The feedback information includes feature matching degree and quantitative indicators of modal fusion effect. The inversion parameters are dynamically adjusted according to the feedback information, with 100 iterations and a step size decay coefficient of 0.95 for each iteration. Finally, the optimized lithofacies identification results are output.

5. The deep learning-based method for fine-grained sedimentary rock facies identification according to any one of claims 1-4, characterized in that, This method is implemented through different units, including: The fine-grained sedimentary rock multi-source data acquisition and screening unit is used to acquire lithofacies curve data and microstructure image data corresponding to rock core samples, screen effective lithofacies curve multimodal characteristic parameters and microstructure characteristic parameters, and transmit the screened data to the lithofacies curve multimodal intelligent analysis unit. The lithofacies curve multimodal intelligent analysis unit receives data transmitted from the multi-source data acquisition and screening unit, generates a standardized multimodal feature set through feature dimension mapping, modal information fusion and standardization processing, and transmits the feature set to the microstructure feature extraction unit and the lithofacies identification result optimization unit respectively. The microstructure feature extraction unit calls the microstructure feature extraction CNN model to perform hierarchical convolution and pooling operations on the microstructure image data, generate deep microstructure feature vectors, and transmit them to the cross-attention lithofacies recognition unit. The cross-attention lithofacies identification unit receives a standardized multimodal feature set and a deep microstructure feature vector, outputs preliminary lithofacies identification results through cross-modal attention weight allocation and nonlinear transformation, and transmits them to the lithofacies identification result optimization unit. The lithofacies identification result optimization unit uses a Gaussian constrained inversion change prediction algorithm to perform inversion correction and dynamic adjustment on the preliminary lithofacies identification results, generate optimized identification results, and transmit them to the lithofacies classification system construction unit. The lithofacies classification system construction unit receives the optimized identification results and feedback information from the lithofacies curve multimodal intelligent analysis unit, constructs a lithofacies classification system for fine-grained sedimentary rocks, completes the accurate classification of lithofacies types, and outputs the final identification results.

Images