Multi-component mineral content prediction method and device, electronic equipment and storage medium

Abstract

The invention relates to the technical field of oil and gas reservoir exploration, in particular to a multi-component mineral content prediction method and device, electronic equipment and a storage medium. The input features are input to a multi-component mineral content prediction model, corresponding mineral content prediction results of all the components are output, the multi-component mineral content prediction model is obtained by training a preset deep learning model through multiple samples, and the deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion. The model constructed by the invention not only can efficiently capture the global long-term trend and the local sudden change fluctuation of the logging curve, but also can automatically learn the mutual restriction relationship between minerals, so that the accurate mapping between the multi-scale logging characteristics and the specific mineral category is realized; therefore, cooperative efficient prediction of the multi-component minerals can be realized by utilizing the multi-component mineral content prediction model, and an effective data basis is provided for exploration and development of petroleum and natural gas, reservoir evaluation and productivity prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas reservoir exploration, and is a multi-component mineral content prediction method and device, an electronic device, and a storage medium. BACKGROUND

[0002] Accurate identification and quantitative characterization of reservoir mineral components are fundamental and key technical links in oil and gas exploration and development, reservoir evaluation, and productivity prediction. In particular, the development of multi-component minerals such as illite, sylvite, and witherite in alkaline lake sedimentary reservoirs directly affects the physical properties and fluid flow characteristics of the reservoir.

[0003] The existing traditional mineral content determination method mainly relies on core laboratory analysis (such as X-ray diffraction and thin section identification). Although this method is accurate, it faces the following challenges: (1) Non-continuity and high cost: core sampling is expensive and time-consuming, and can only provide information at discrete depth points, making it difficult to achieve continuous evaluation of the entire well section; (2) Complex nonlinear mapping: although logging technology can provide continuous observation, there is a high nonlinear relationship between conventional logging curves (such as AC, GR, DEN, etc.) and mineral content. A single curve cannot indicate a specific mineral, and when multiple minerals coexist, the logging response is often a superposition of multiple sources, making it difficult to decouple using conventional linear regression or empirical formulas.

[0004] Furthermore, with the development of machine learning technology, machine learning has also been applied to reservoir mineral content determination, for example: Existing patent literature one, with publication number CN118246579A, discloses a method and device for predicting the volume content of mixed sedimentary rocks based on deep learning, which relates to the technical field of oil reservoir exploration. The method includes: collecting measured data and establishing a sample data set based on the measured data; wherein the measured data at least includes conventional logging curves, ECS logging, and X-ray diffraction whole rock mineral analysis data; establishing a training set, a test set, and a validation set based on the sample data set; establishing a machine learning model, training the machine learning model using the training set, and performing input and output parameter nonlinear relationship mapping learning on the machine learning model; testing and verifying the machine learning model based on the test set and the validation set to obtain a mixed sedimentary rock multi-mineral volume content prediction model; calculating and generating a multi-mineral volume content curve using the mixed sedimentary rock multi-mineral volume content prediction model, and performing reservoir evaluation based on the multi-mineral volume content curve.

[0005] Existing patent document 2, publication number CN117951509B, discloses a method and system for predicting mineral composition content, belonging to the field of oil and gas field exploration and development. The method includes collecting measured rock mineral composition content data and well logging data within a preset depth range, and performing preprocessing; selecting and reducing features using the correlation coefficient method and principal component analysis; then constructing a multi-source fusion dataset and training a hybrid neural network model; establishing a loss function for mineral content prediction with custom constraints; and finally using the hybrid neural network model to predict mineral composition content. This invention fully considers the correlation between well logging data and well logging data and rock mineral composition, constructs a custom loss function based on mineral composition content characteristics, improves prediction accuracy, and solves the defects of difficult formation core sampling, inconvenient indoor experiments, and discontinuous rock mineral content data.

[0006] However, the following problems exist: (1) Limitations of traditional machine learning: Although traditional algorithms such as support vector machine and random forest introduce nonlinear processing capabilities, they rely heavily on manual feature engineering, making it difficult to automatically extract deep features from the original well logging curves. They also usually ignore the symbiotic relationship between minerals, resulting in poor geological consistency of the prediction results.

[0007] (2) In the application of Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and Transformer architectures, most models based on single or conventional deep learning architectures still have performance bottlenecks in the special task of multi-component mineral prediction, for example: The inherent limitations of CNN and RNN architectures: Although CNN models can effectively extract local fluctuation features of logging curves, they are limited by the local receptive field of the convolution kernel, making it difficult to capture long-distance dependencies and global geological trends in logging sequences; although RNNs are good at processing sequential information, they are prone to gradient vanishing problems when faced with high-resolution data in long well sections, and the serial computing mode leads to low training efficiency, making it difficult to meet the needs of rapid processing of large-scale logging data.

[0008] The computational efficiency and lack of correlation of the pure Transformer architecture: Although the Transformer solves the long-range dependency problem through the self-attention mechanism, its computational complexity increases quadratically with the length of the logging sequence (O(N)). 2 When processing data from long well sections, the computational resources consumed are enormous, making efficient inference difficult. More importantly, existing models typically treat each mineral as an independent prediction task, ignoring the geological symbiotic relationships and mutual constraints between minerals (such as the waxing and waning relationship between two minerals), resulting in a lack of consistency in the geological logic of multi-mineral joint prediction results.

[0009] Geological interpretation deficiency of loss function design: the existing deep learning methods mostly use point-to-point loss functions such as mean square error (MSE), and excessively focus on the minimization of numerical absolute error, while ignoring the consistency of the predicted curve and the real curve in the overall waveform trend, shape and phase. This leads to the model having a small error in numerical statistics, but the trend deviation, phase misplacement or shape distortion often occur in the geological interpretation level, which seriously affects the credibility and practical application value of the prediction results. SUMMARY

[0010] The present application provides a multi-component mineral content prediction method, device, electronic equipment and storage medium, which overcomes the shortcomings of the prior art. It can effectively solve the problem that the existing mineral content determination method cannot consider both the global long-range dependence and the local fine-grained features of the well logging curve, and effectively handle the internal correlation between multiple minerals.

[0011] One of the technical solutions of the present application is realized by the following measures: a multi-component mineral content prediction method, comprising: obtaining input features of a to-be-predicted well section, wherein the input features include multiple conventional well logging curves of the to-be-predicted well section; inputting the input features into a multi-component mineral content prediction model, outputting corresponding multi-component mineral content prediction results, and performing inverse normalization processing to restore the real mineral content percentage, wherein the multi-component mineral content prediction model is obtained by training a preset deep learning model using multiple samples, the deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion, and each sample includes multiple conventional well logging curves of a historical well section and corresponding multi-component mineral content label information.

[0012] The following is a further optimization or / and improvement of the above-mentioned technical solutions of the application: The construction process of the above-mentioned multi-component mineral content prediction model comprises: obtaining historical data sets of multiple historical wells, and dividing them into a training set, a validation set and a test set in proportion, wherein the historical data set of each historical well includes multiple conventional well logging curves and corresponding multi-component mineral content; performing data preprocessing on the training set, the validation set and the test set respectively; The preset deep learning model is trained by using the training set, and a joint loss function is introduced during training. In combination with a random weight average strategy and a progressive learning rate scheduling mechanism, the deep learning model is iteratively optimized and trained until convergence, to obtain a component mineral content prediction model, wherein the deep learning model comprises a feature embedding layer, a 4Mamba-attention encoder, a mineral guided Transformer decoder, a prediction head, and a hybrid loss module; the feature embedding layer performs high-dimensional mapping and adds position encoding; the 4Mamba-attention encoder introduces a bidirectional state space model Mamba, combines a parallel large / small convolution kernel strategy, captures the global long-term trend and local sudden fluctuation of the well logging curve, and extracts encoded features; the mineral guided Transformer decoder maps the multi-scale logging features extracted by the encoder to a specific mineral category to obtain decoded features; and the prediction head and the hybrid loss module obtain a component mineral content prediction result according to the decoded features. The component mineral content prediction model is verified and tested by using the verification set and the test set respectively, the model parameters of the component mineral content prediction model are optimized, and the component mineral content prediction model meeting the test evaluation requirements is output.

[0013] The above data preprocessing is performed on the training set, the verification set, and the test set respectively, including: Each historical data set in the training set is intercepted according to the window size and the batch size to obtain corresponding training samples; Each historical data set in the verification set is intercepted according to the window size to obtain corresponding verification samples; Each historical data set in the test set is intercepted according to the window size to obtain corresponding test samples; The training set, the verification set, and the test set are standardized by using the statistics of the training set, and the label values of the component mineral content are normalized to the interval [0, 1].

[0014] The above 4Mamba-attention encoder is stacked by multiple Mamba blocks to extract features from the input logging features, wherein the internal processing mechanism of each Mamba block includes: Four branches are processed in parallel, and the logging features are input into four independent state space model branches in parallel for feature extraction, wherein the four independent state space model branches include: The forward large-scale Mamba branch uses a Mamba module with a convolution kernel size of 4 to process the original sequence to capture the long-distance geological evolution trend in the well logging curve; The forward small-scale Mamba branch uses a Mamba module with a convolution kernel size of 2 to focus on capturing local lithology mutation details; The reverse large-scale Mamba branch and the reverse small-scale Mamba branch input the large-scale and small-scale Mamba modules respectively by reversing the input sequence, and reverse the output back to the original order; The cross-scale feature fusion introduces a cross-scale cross-attention fusion module to integrate the output features of the state space model branches and generate encoding features containing global long-range dependencies and local fine-grained information.

[0015] The mineral-guided Transformer decoder includes learnable mineral query vectors corresponding to each multi-component mineral and multiple Transformer decoding layers.

[0016] The joint loss function calculation formula is as follows: Wherein, is the joint loss value; is the consistency correlation coefficient loss value; is the L1 loss value; is the balance coefficient.

[0017] The second technical solution of the present application is realized by the following measures: a multi-component mineral content prediction device comprises: A data acquisition unit acquires input features of a to-be-predicted well section, wherein the input features include multiple conventional logging curves of the to-be-predicted well section. A prediction unit inputs the input features into a multi-component mineral content prediction model to output corresponding multi-component mineral content prediction results, and the prediction results are subjected to inverse normalization processing to restore the real mineral content percentage, wherein the multi-component mineral content prediction model is obtained by training a preset deep learning model using multiple samples, and the deep learning model adopts a bidirectional multi-scale feature extraction and cross-scale attention fusion strategy, and each sample includes multiple conventional logging curves of a historical well section and corresponding multi-component mineral content label information.

[0018] The following is a further optimization or / and improvement of the above-mentioned technical solution: The model construction unit comprises: A dataset construction module acquires historical datasets of multiple historical wells and divides the historical datasets into a training set, a validation set and a test set in proportion, wherein each historical dataset of each historical well includes multiple conventional logging curves and corresponding multi-component mineral content. A preprocessing module performs data preprocessing on the training set, the validation set and the test set respectively. The model training module trains a preset deep learning model by using a training set, introduces a joint loss function during training, and combines a random weight average strategy and a gradual learning rate scheduling mechanism to iteratively optimize and train the deep learning model until convergence, thereby obtaining a component mineral content prediction model, wherein the deep learning model comprises a feature embedding layer, a 4Mamba-attention encoder, a mineral-guided Transformer decoder, a prediction head, and a hybrid loss module; the feature embedding layer performs high-dimensional mapping and adds position encoding; the 4Mamba-attention encoder introduces a bidirectional state space model Mamba, combines a parallel large / small convolution kernel strategy, captures global long-term trends and local sudden fluctuations of well logging curves, and extracts encoded features; the mineral-guided Transformer decoder maps the multi-scale logging features extracted by the encoder to specific mineral categories to obtain decoded features; and the prediction head and the hybrid loss module obtain component mineral content prediction results according to the decoded features. The model test optimization module verifies and tests the component mineral content prediction model by using a verification set and a test set, respectively, optimizes model parameters of the component mineral content prediction model, and outputs a component mineral content prediction model that meets test evaluation requirements.

[0019] The third technical solution of the present application is realized by the following measures: an electronic device comprising a processor and a memory, the memory storing a computer program, the computer program being loaded and executed by the processor to realize the steps in the multi-component mineral content prediction method.

[0020] The fourth technical solution of the present application is realized by the following measures: a storage medium storing a computer program readable by a computer, the computer program being set to execute the steps in the multi-component mineral content prediction method when running.

[0021] The present application has the following advantages: Strong non-linear mapping capability: thanks to the dual capture of macro-geological trends and micro-lithology anomalies by the 4Mamba architecture, which takes into account long program sequence dependencies and fine-grained local features, the multi-component mineral content prediction model constructed by the present application exhibits excellent feature extraction performance when processing complex reservoir logging data, effectively solving the gradient vanishing problem.

[0022] Multi-mineral synergistic efficient prediction: In the decoding stage, specific learnable queries are designed to accurately map the multi-scale logging features extracted in the encoding stage to specific mineral categories through the cross-attention mechanism of the Transformer, solving the technical problem that the traditional method is difficult to effectively represent the complex response relationship between minerals and logging curves; and the content of three or more minerals can be output at a time through a single forward propagation, which can automatically learn the mutual restraint relationship between minerals, realizing the dual improvement of prediction efficiency and accuracy.

[0023] Temporal and spatial consistency hybrid loss strategy: A hybrid loss function composed of consistency correlation coefficient loss and L1 loss is constructed, the consistency correlation coefficient loss is used to constrain the consistency of the prediction results and the real curves in the overall waveform trend and shape, and the L1 loss is used to control the point-to-point numerical absolute error, which takes into account the global waveform rationality and local numerical accuracy, so that the model obtained by training is more accurate.

[0024] Robust generalization under small sample: In the actual scene where logging samples are often scarce, the SWA strategy effectively prevents the model from falling into a sharp local optimal solution, so that the model can maintain certain prediction performance when facing new well section data (blind well) that has not been seen before. BRIEF DESCRIPTION OF DRAWINGS

[0025] FIG. 1 is a schematic diagram of an implementation environment provided by an embodiment of the present application. Figure 1 FIG. 2 is a schematic diagram of a multi-component mineral content prediction method provided by an embodiment of the present application.

[0026] Figure 2 FIG. 3 is a schematic diagram of a multi-component mineral content prediction model construction method provided by an embodiment of the present application.

[0027] FIG. 4 is a schematic diagram of a deep learning model main framework structure provided by an embodiment of the present application. Figure 3 FIG. 5 is a schematic diagram of a 4Mamba-attention encoder structure provided by an embodiment of the present application.

[0028] Figure 4 FIG. 6 is a schematic diagram of a Mamba module structure provided by an embodiment of the present application.

[0029] FIG. 7 is a schematic diagram of a prediction head structure provided by an embodiment of the present application. Figure 5 FIG. 8 is a comparison visualization diagram of real values and predicted values of the mh54 well provided by an embodiment of the present application.

[0030] Figure 6 FIG. 9 is a comparison visualization diagram of real values and predicted values of the mh54 well provided by an embodiment of the present application.

[0031] FIG. 10 is a comparison visualization diagram of real values and predicted values of the mh54 well provided by an embodiment of the present application. Figure 7 FIG. 11 is a comparison visualization diagram of real values and predicted values of the mh54 well provided by an embodiment of the present application.

[0032] Figure 8 FIG. 12 is a comparison visualization diagram of real values and predicted values of the mh54 well provided by an embodiment of the present application. ​​​​

[0033] Figure 2 shows a test well diagram of the multi-component mineral content prediction model provided by an embodiment of the present application in a well mh54. Figure 9 Figure 3 shows a test well diagram of the multi-component mineral content prediction model provided by an embodiment of the present application in a blind well mh61.

[0034] Figure 10 Figure 4 shows a structure schematic diagram of a multi-component mineral content prediction device provided by an embodiment of the present application.

[0035] Figure 5 shows a structure schematic diagram of another multi-component mineral content prediction device provided by an embodiment of the present application. Figure 11 Figure 6 shows a structure schematic diagram of a multi-component mineral content prediction device provided by an embodiment of the present application.

[0036] Figure 12 Figure 7 shows a structure schematic diagram of another multi-component mineral content prediction device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0037] The present application is not limited by the following embodiments, and the specific implementation can be determined according to the technical solutions of the present application and the actual situation.

[0038] Those skilled in the art can understand that, unless specifically stated, the "module" or "unit" in the embodiments of the present application refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory) or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0039] In addition, "multiple" in the embodiments of the present application refers to two or more, and "first" and "second" and the like are used for differentiation and description, and cannot be understood as implying relative importance.

[0040] The embodiments of the present application provide a multi-component mineral content prediction method, including: obtaining input features of a to-be-predicted well section, wherein the input features include multiple conventional logging curves of the to-be-predicted well section; inputting the input features into a multi-component mineral content prediction model, outputting corresponding multi-component mineral content prediction results, and restoring the multi-component mineral content prediction results to real mineral content percentages through inverse normalization processing, wherein the multi-component mineral content prediction model is obtained by training a preset deep learning model using multiple samples, the deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion, and each sample includes multiple conventional logging curves of a historical well section and corresponding multi-component mineral content label information.

[0041] ​​Wherein, the method provided by the embodiment of the present application can involve artificial intelligence (Artificial Intelligence, AI) technology, and can be implemented based on artificial intelligence technology, for example, in a deep learning manner, and a corresponding model is obtained by using sample training.

[0042] Wherein, machine learning (Machine Learning, ML) is a multi-field interdisciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other subjects. It is a subject that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence.

[0043] Wherein, deep learning (DL, Deep Learning) specifically refers to machine learning based on deep neural network models and methods. It is developed based on statistical machine learning, artificial neural network algorithms and other models, combined with the development of modern big data and big computing power. The most important technical feature of deep learning is the ability to automatically extract features.

[0044] The above machine learning and deep learning usually include neural network, belief network, reinforcement learning, transfer learning, inductive learning, and rule teaching learning.

[0045] Wherein, the loss function is used in the process of training the neural network. Because it is desired that the output of the neural network is as close as possible to the value that is actually intended to be predicted, the difference between the predicted value of the current network and the target value can be compared, and the weight vector of each layer of the neural network can be updated according to the difference between the two, wherein an initialization process is usually performed before the first update, that is, the parameters of each layer of the neural network are pre-configured. Until the neural network can predict the target value or a value very close to the target value. Therefore, when deep learning, it is necessary to define "how to compare the difference between the predicted value and the target value", which is the loss function.

[0046] As shown in the accompanying Figure 1 , it shows an implementation environment schematic diagram provided by an embodiment of the present application. The implementation environment can include: a training device and a using device.

[0047] The training device and the using device are computer devices; optionally, the computer device is a terminal device, such as a mobile phone, a tablet computer, a PC (Personal Computer, personal computer) and the like; or, the computer device is a server, which can be a server or a server cluster composed of multiple servers, or a cloud computing service center, and the embodiments of the present application do not limit this.

[0048] The training device refers to a computer device with neural network training and learning capability. Optionally, the training device has neural network acquisition capability and is trained and learned according to application requirements. For example, the training device acquires a neural network from other devices through a network, and then trains the neural network through training samples according to application requirements, so that the neural network has the capability to obtain multi-component mineral content prediction results. Optionally, the training device has neural network construction capability, which can construct a neural network according to application requirements, and then train and learn the neural network. For example, the training device constructs a neural network by itself in order to obtain multi-component mineral content prediction results according to multiple conventional logging curves, and then trains and learns the neural network according to application requirements through samples.

[0049] The use device refers to a computer device with neural network use requirements. Optionally, the use device acquires a neural network from other devices through a network according to application requirements. For example, the use device has the requirement to obtain multi-component mineral content prediction results, which can acquire a neural network trained and learned to predict multi-component mineral content from other devices through a network, and use the neural network to predict multi-component mineral content.

[0050] Therefore, the technical solutions of the present application will be described below in combination with several examples.

[0051] Embodiment 1: As shown in the accompanying Figure 2 The embodiment of the present application discloses a multi-component mineral content prediction method, which comprises: Step S110, acquiring input features of a to-be-predicted well section, wherein the input features include multiple conventional logging curves of the to-be-predicted well section; Step S210, inputting the input features to a multi-component mineral content prediction model, outputting corresponding multi-component mineral content prediction results, and restoring to real mineral content percentages through inverse normalization processing, wherein the multi-component mineral content prediction model is obtained by training a preset deep learning model using multiple samples, the deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion, and each sample includes multiple conventional logging curves of a historical well section and corresponding multi-component mineral content label information.

[0052] The input features in the step S110 include multiple conventional logging curves of the well section to be predicted, and it should be noted that the length of the well section to be predicted needs to be consistent with the historical well section length of the sample when the multi-component mineral content prediction model is constructed, and the multiple conventional logging curves include acoustic time difference (AC), neutron porosity (CNL), density (DEN), natural gamma (GR), deep lateral resistivity (RT), shallow lateral resistivity (RXO), and natural alkali content, and further, the multiple conventional logging curves of the well section to be predicted need to be standardized; and the minerals include searlesite, arcanite, and natronarcanite.

[0053] The embodiment of the present application discloses a multi-component mineral content prediction method, a deep learning model adopting a bidirectional multi-scale feature extraction and cross-scale attention fusion strategy is established, and a multi-component mineral content prediction model is obtained by training the deep learning model. The model can not only efficiently capture the global long-term trend and local mutation fluctuation of the logging curve, effectively depict the deep nonlinear features of the logging data, but also automatically learn the mutual restraint relationship between minerals, realize the accurate mapping between the multi-scale logging features and the specific mineral categories, and thus the multi-component mineral content prediction model can realize the collaborative and efficient prediction of the multi-component minerals, and provide an effective data basis for oil and gas exploration and development, reservoir evaluation, and productivity prediction.

[0054] Embodiment 2: as shown in the accompanying Figure 3 The embodiment of the present application is a further optimization of the above-mentioned embodiment, and the construction process of the multi-component mineral content prediction model includes: Step S210, obtaining historical data sets of multiple historical wells, and dividing the historical data sets into a training set, a validation set, and a test set in proportion, wherein the historical data set of each historical well includes multiple conventional logging curves and corresponding component mineral contents; wherein the multiple conventional logging curves include acoustic time difference (AC), neutron porosity (CNL), density (DEN), natural gamma (GR), deep lateral resistivity (RT), shallow lateral resistivity (RXO), and natural alkali content, and the minerals include searlesite, arcanite, and natronarcanite; It should be noted that after obtaining the historical data sets of the multiple historical wells, the obvious abnormal values and invalid measurement sections in the conventional logging curves can be removed first, and then all the historical data sets are divided into the training set, the validation set, and the test set in proportion, and the proportion is set according to the needs, which can be 70%:10%:20%.

[0055] Step S220, performing data preprocessing on the training set, the validation set, and the test set, including: cutting the logging features according to the window size and batch size in each historical data set in the training set to obtain corresponding training samples; Well logging features are truncated from each historical dataset in the validation set according to the window size to obtain the corresponding validation samples; Well logging features are extracted from each historical dataset in the test set according to the window size to obtain the corresponding test samples; The training set, validation set, and test set are standardized using the statistics of the training set, and the label values ​​of mineral content of each component are normalized to the [0,1] interval.

[0056] The window size can be set to Window_size=64 and the batch size can be set to Batch_size=32. Thus, if there are 7 regular logging curves, the dimension of the training samples can be [32,64,7].

[0057] The above standardization process involves using the statistics (mean and variance) of the training set to standardize the training set, validation set, and test set, eliminating numerical differences between different dimensions (such as resistivity and porosity). The statistics of the training set are the well logging curves of the training samples, which are standardized to have a mean of 0 and a variance of 1. The statistics are updated by performing a forward propagation on the final SWA model using the training data before formal inference, updating the mean and variance statistics of the batch normalization layer (BatchNorm).

[0058] Step S230: The pre-defined deep learning model is trained using the training set. During training, a joint loss function is introduced, and a random weight averaging strategy and a progressive learning rate scheduling mechanism are combined to iteratively optimize and train the deep learning model until convergence, resulting in a prediction model for the mineral content of each component. The deep learning model includes a feature embedding layer, a 4Mamba-attention encoder, a mineral-guided Transformer decoder, a prediction head, and a hybrid loss module. The feature embedding layer performs high-dimensional mapping and adds positional encoding. The 4Mamba-attention encoder introduces the bidirectional state-space model Mamba and combines a parallel large / small convolutional kernel strategy to capture the global long-term trend and local abrupt fluctuations of the logging curve, extracting encoded features. The mineral-guided Transformer decoder maps the multi-scale logging features extracted by the encoder to specific mineral categories to obtain decoded features. The prediction head and hybrid loss module obtain the prediction results of the mineral content of each component based on the decoded features.

[0059] As attached Figure 4 As shown, the main framework of the aforementioned deep learning model is an encoder-decoder collaborative network, which is a deep learning architecture that integrates a state-space model and an attention mechanism. Its overall architecture employs a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion, specifically including: The feature embedding layer maps the logging features of the training sample to a high dimension and adds position encoding, and outputs the logging features with a high dimension; for example, the logging features with a dimension of [32, 64, 7] are mapped to a high dimension and added with position encoding, and the logging features with a dimension of [32, 64, 256] are outputted; 4 Mamba-Attention Encoder, as shown in the accompanying Figure 5 Fig. 2, stacked by multiple layers of Mamba blocks, which extracts features from the input logging features, wherein the internal processing mechanism of each layer of Mamba blocks includes: (1) Four-branch parallel processing: the logging features are inputted into four independent state space model branches in parallel, which are forward large-scale Mamba branch, forward small-scale Mamba branch, reverse large-scale Mamba branch and reverse small-scale Mamba branch, respectively; The forward large-scale Mamba branch processes the original sequence by using a Mamba module with a convolution kernel size of 4 (structure as shown in the accompanying Figure 6 Fig. 3), to capture the long-distance geological evolution trend in the logging curve; The forward small-scale Mamba branch uses a Mamba module with a convolution kernel size of 2, which focuses on capturing local lithology mutation details; The reverse large-scale Mamba branch and the reverse small-scale Mamba branch input the input sequence into the large-scale (conv=4) and small-scale (conv=2) Mamba modules, respectively, after reversing (Flip), and finally reverse the output back to the original order; The above design overcomes the limitations of one-way causal modeling of Mamba, and realizes bidirectional perception of the context information of the logging features.

[0060] (2) Cross-scale feature fusion: cross-scale cross-attention is introduced to integrate the output features of each state space model branch, to generate encoding features containing global long-range dependencies and local fine-grained information, with a dimension of [32, 64, 256]; The above cross-scale feature fusion integrates the output features of the four branches, which is to use the global features extracted by the large-scale Mamba branch as the query (Query), and the local features extracted by the small-scale Mamba branch as the key (Key) and value (Value), to calculate the attention weight matrix, dynamically inject local detail information into global structure features, realize cross-scale complementary fusion of features, and finally output the encoding features with a dimension of [32, 64, 256].

[0061] Mineral-guided Transformer Decoder, as shown in the accompanying Figure 4As shown, the mineral query vector corresponding to each component mineral and the multi-layer Transformer decoding layer are included. Through the cross-attention mechanism in the Transformer decoding layer, the mineral query vector is used as the query, and the encoding features are used as the key and the value. The logging response features related to the specific mineral are learned and aggregated autonomously, and the decoding features are output, which can have dimensions of [32, 64, 256].

[0062] The prediction head and the mixed loss module include the prediction head and the mixed loss sub-module, and specifically: The prediction head is as shown in Figure 7 As shown, the decoding features output by the mineral-guided Transformer decoder are dimensionally reduced through the MLP layer (linear transformation -> GELU activation -> linear transformation), and finally pass through the Sigmoid activation function to strictly constrain the output value in the range of [0, 1], which directly corresponds to the percentage content of each component mineral. The final output dimension can be [32, 64, 3]; The mixed loss sub-module implements a training strategy, introduces a joint loss function, and combines a random weight average strategy and a progressive learning rate scheduling mechanism to iteratively optimize and train the deep learning model until convergence. Specifically: The formula for calculating the joint loss function is: The above is a consistency correlation coefficient loss value, which is used to constrain the consistency of the predicted curve and the real logging curve in the overall waveform trend and geological morphology, is an L1 loss value, which is used to constrain the point-to-point numerical absolute error. The two are balanced by an adjustable weight coefficient to consider both the global rationality and the local accuracy of the prediction under limited sample conditions. In the calculation formula, is a balance coefficient, which can be set to 0.5.

[0063] During the optimization process, the SWA (stochastic weight averaging) strategy is introduced to average the model weights at the end of the training, guide the model to converge to the flat minimum value region of the loss plane, and improve the generalization ability. The random weight average (SWA) strategy is to average the model weights of multiple checkpoints (Checkpoints) at the end of the training process to obtain a centralized weight distribution, thereby enhancing the generalization ability of the model on unseen well segment data.

[0064] Step S240, the validation set and the test set are used to verify and test the component mineral content prediction model, respectively, to optimize the model parameters of the component mineral content prediction model, and output the component mineral content prediction model that meets the test evaluation requirements.

[0065] Example 3: To comprehensively verify the effectiveness and generalization ability of the multi-component mineral content prediction model constructed in the present application, three representative test well sections (numbered fy1, mh59 and mh54) are selected for model verification, and mh61 well is additionally selected as an independent blind well test set. In terms of quantitative evaluation, the determination coefficient (R²) and Pearson correlation coefficient (R) are used as core evaluation indicators. In order to intuitively show the actual prediction performance of the model under different geological conditions, the test results of mh54 and blind well mh61 are visualized in the drawings. Figs. 1, 2 and 3 respectively show the test results of the test well and the blind well. Among them, Fig. 1 is a comparison of the true value and the predicted value of well mh54; Fig. 2 is a test well plot of the multi-component mineral content prediction model in well mh54; Fig. 3 is a test well plot of the multi-component mineral content prediction model in blind well mh61. By comparing the predicted curve and the true value curve, the prediction accuracy of the model can be qualitatively evaluated. Figure 8 Figure 9 Figure 10 Figure 8 Figure 9 Figure 10

[0066] The experimental results are shown in Table 1. The multi-component mineral content prediction model exhibits excellent generalization ability and prediction accuracy in different well sections: in terms of overall performance, the average determination coefficient (Mean R²) of the three test well sections is more than 0.73, among which the MH54 well is as high as 0.7984; the average correlation coefficient (Mean R) is stable above 0.86, and the highest reaches 0.8955. At the same time, the error index remains at a low level, proving that the predicted value is highly consistent with the true core test value. In terms of specific mineral components, the model performs particularly well in predicting soddite, with a single mineral R² of 0.9257 in well mh54; for natrocalcite and natro-mangesite, the model also maintains stable prediction ability under complex geological conditions, with R values exceeding 0.85 in most test sets.

[0067] Table 1. Multi-component mineral prediction performance statistics of the model in different well sections of the present application .

[0068] This example shows that the multi-component mineral content prediction model of the present application, by using the 4Mamba-attention encoder and mineral-guided Transformer decoder design, combined with the mixed loss function and SWA training strategy, effectively solves the problems of difficult prediction and low accuracy of multi-component mineral content in complex reservoirs, and realizes high-precision and low-cost continuous evaluation based on conventional logging data.

[0069] Example 4: As shown in Fig. 4, the multi-component mineral content prediction model of the present application is applied to the well mh61, and the predicted curve and the true value curve are compared. It can be seen that the model has a high prediction accuracy, and the predicted curve is basically consistent with the true value curve. Figure 11 ​​​​​​As shown, the embodiment of the present application discloses a multi-component mineral content prediction device, comprising: The data acquisition unit acquires input features of the to-be-predicted well section, wherein the input features include multiple conventional logging curves of the to-be-predicted well section. The prediction unit inputs the input features into a multi-component mineral content prediction model, outputs corresponding multi-component mineral content prediction results, and restores them to real mineral content percentages through inverse normalization processing, wherein the multi-component mineral content prediction model is obtained by training a preset deep learning model using multiple samples, the deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion, and each sample includes multiple conventional logging curves of a historical well section and corresponding multi-component mineral content label information.

[0070] Embodiment 5: as shown in the accompanying Figure 12 As shown, the embodiment of the present application discloses a multi-component mineral content prediction device, comprising: The model construction unit comprises: The data set construction module acquires historical data sets of multiple historical wells, and divides them into a training set, a validation set and a test set in proportion, wherein the historical data set of each historical well includes multiple conventional logging curves and corresponding multi-component mineral content. The preprocessing module performs data preprocessing on the training set, the validation set and the test set respectively. The model training module trains a preset deep learning model using the training set, introduces a joint loss function during training, and combines a random weight average strategy and a progressive learning rate scheduling mechanism to iteratively optimize and train the deep learning model until convergence, thereby obtaining a multi-component mineral content prediction model, wherein the deep learning model includes a feature embedding layer, a 4Mamba-attention encoder, a mineral-guided Transformer decoder, a prediction head and a hybrid loss module; the feature embedding layer performs high-dimensional mapping and adds position encoding; the 4Mamba-attention encoder introduces a bidirectional state space model Mamba, combines a parallel large / small convolution kernel strategy, captures the global long-term trend and local sudden fluctuation of the logging curve, and extracts encoded features; the mineral-guided Transformer decoder maps the multi-scale logging features extracted by the encoder to specific mineral categories to obtain decoded features; and the prediction head and the hybrid loss module obtain multi-component mineral content prediction results according to the decoded features. The model test optimization module verifies and tests the multi-component mineral content prediction model using the validation set and the test set respectively, optimizes the model parameters of the multi-component mineral content prediction model, and outputs a multi-component mineral content prediction model that meets the test evaluation requirements.

[0071] The data acquisition unit acquires input features of the to-be-predicted well section, wherein the input features include multiple conventional logging curves of the to-be-predicted well section.

[0072] The prediction unit inputs the input features into a multi-component mineral content prediction model, outputs corresponding prediction results of the mineral content of each component, and restores the real mineral content percentage through inverse normalization processing. The multi-component mineral content prediction model is obtained by training a preset deep learning model using a plurality of samples. The deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion. Each sample includes a plurality of conventional logging curves of a historical well section and corresponding label information of the mineral content of each component.

[0073] The specific implementation steps and method embodiments of each unit / module in the above embodiments are the same as those of the above embodiments, and will not be described again.

[0074] Embodiment 6: The embodiment of the present application discloses a storage medium, wherein the storage medium stores a computer program readable by a computer, and the computer program is set to execute a multi-component mineral content prediction method when running.

[0075] The above storage medium can include but is not limited to: a U disk, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk, and various storage media that can store computer programs.

[0076] Embodiment 7: The embodiment of the present application discloses an electronic device, which includes a processor and a memory, and the memory stores a computer program. The computer program is loaded and executed by the processor to perform a multi-component mineral content prediction method.

[0077] The above processor can be a central processing unit CPU, a general-purpose processor, a digital signal processor DSP, an ASIC, an FPGA or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. It can realize or execute various exemplary logical blocks, modules and circuits described in combination with the present application. It can also be a combination of computing functions, such as one or more microprocessor combinations, DSP and microprocessor combinations, etc. The memory can include but is not limited to: a U disk, a read-only memory, a mobile hard disk, a magnetic disk or an optical disk, and various storage media that can store computer programs.

[0078] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer-readable program code. The program code may

[0079] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to the present application. In this regard, each flowchart block and / or block in the Figures can represent a module, segment, or portion of code, which comprises one or more executable Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to the present application. In this regard, each flowchart block and / or block in the Figures can represent a module, segment, or portion of code, which comprises one or more executable

[0080] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to the present application. In this regard, each flowchart block and / or block in the Figures can represent a module, segment, or portion of code, which comprises one or more executable Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to the present application. In this regard, each flowchart block and / or block in the Figures can represent a module, segment, or portion of code, which comprises one or more executable

[0081] The above-described embodiments are merely given as examples, and it should be understood that the present application is not limited thereto. Those skilled in the art can make various modifications and substitutions to the specific embodiments without departing from the scope of the present application. Therefore, the technical scope of the present application shall be defined by the technical scope of the claims.

Claims

1. A method of multi-component mineral content prediction, characterized by, The method comprises the following steps: obtaining input features of a to-be-predicted well section, wherein the input features comprise multiple conventional logging curves of the to-be-predicted well section; inputting the input features into a multi-component mineral content prediction model, outputting corresponding multi-component mineral content prediction results, and restoring the multi-component mineral content prediction results to real mineral content percentages through inverse normalization processing, wherein the multi-component mineral content prediction model is obtained by training a preset deep learning model using multiple samples, and the deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion, and each sample comprises multiple conventional logging curves of a historical well section and corresponding multi-component mineral content label information.

2. The multi-component mineral content prediction method of claim 1, wherein, The construction process of the multi-component mineral content prediction model comprises the following steps: obtaining historical data sets of multiple historical wells, and dividing the historical data sets into a training set, a validation set and a test set according to a proportion, wherein the historical data set of each historical well comprises multiple conventional logging curves and corresponding multi-component mineral content; performing data preprocessing on the training set, the validation set and the test set respectively; training the preset deep learning model using the training set, introducing a joint loss function during the training, and combining a random weight average strategy and a progressive learning rate scheduling mechanism to iteratively optimize and train the deep learning model until convergence, thereby obtaining the multi-component mineral content prediction model, wherein the deep learning model comprises a feature embedding layer, a 4Mamba-attention encoder, a mineral-guided Transformer decoder, a prediction head and a hybrid loss module; the feature embedding layer performs high-dimensional mapping and adds position encoding; the 4Mamba-attention encoder introduces a bidirectional state space model Mamba, combines a parallel large / small convolution kernel strategy, captures global long-term trends and local sudden fluctuations of the logging curves, and extracts encoded features; the mineral-guided Transformer decoder maps the multi-scale logging features extracted by the encoder to specific mineral categories to obtain decoded features; the prediction head and the hybrid loss module obtain multi-component mineral content prediction results according to the decoded features; verifying and testing the multi-component mineral content prediction model using the validation set and the test set respectively, optimizing model parameters of the multi-component mineral content prediction model, and outputting a multi-component mineral content prediction model that meets the test evaluation requirements.

3. The multi-component mineral content prediction method of claim 2, wherein, performing data preprocessing on the training set, the validation set and the test set respectively, comprising the following steps: performing logging feature cutting on each historical data set in the training set according to a window size and a batch size to obtain corresponding training samples; performing logging feature cutting on each historical data set in the validation set according to the window size to obtain corresponding validation samples; performing logging feature cutting on each historical data set in the test set according to the window size to obtain corresponding test samples; performing standardization processing on the training set, the validation set and the test set using statistical quantities of the training set, and normalizing label values of the multi-component mineral content to the interval [0, 1].

4. The method of multi-component mineral content prediction of claim 2, wherein, The 4Mamba-attention encoder is stacked by multiple Mamba blocks, and performs feature extraction on input logging features, wherein the internal processing mechanism of each Mamba block comprises the following steps: The four-branch parallel processing inputs well logging features into four independent state-space model branches for feature extraction. These four independent state-space model branches include: The forward large-scale Mamba branch uses a Mamba module with a convolution kernel size of 4 to process the original sequence and capture long-distance geological evolution trends in well logging curves; The forward small-scale Mamba branch utilizes a Mamba module with a convolution kernel size of 2 to focus on capturing details of local lithological abrupt changes; The reverse large-scale Mamba branch and the reverse small-scale Mamba branch reverse the input sequence and then input it into the large-scale and small-scale Mamba modules respectively, and reverse the output back to the original order. Cross-scale feature fusion introduces cross-scale attention and integrates the output features of each state space model branch to generate encoded features that contain global long-range dependencies and local fine-grained information.

5. The method of multi-component mineral content prediction of claim 2, wherein, The mineral-guided Transformer decoder includes learnable mineral query vectors that correspond one-to-one with multi-component minerals and multi-layer Transformer decoding layers. Through the cross-attention mechanism in the Transformer decoding layers, it uses mineral query vectors as queries and encoded features as keys and values ​​to autonomously learn and aggregate logging response features related to specific minerals, and outputs decoded features.

6. The multi-component mineral content prediction method of claim 1 or 2, wherein, The formula for calculating the joint loss function is: wherein, is a joint loss value; is a consistency correlation coefficient loss value; is an L1 loss value; is a balancing coefficient.

7. A multi-component mineral content prediction device applying the method according to any one of claims 1 to 6, characterized in that, include: The data acquisition unit acquires the input features of the well section to be predicted, including various conventional logging curves of the well section to be predicted. The prediction unit takes input features into the multi-component mineral content prediction model and outputs the corresponding prediction results of each component mineral content. After inverse normalization, the results are restored to the true mineral content percentage. The multi-component mineral content prediction model is trained on a pre-set deep learning model using multiple samples. The deep learning model adopts a strategy of bidirectional multi-scale feature extraction and cross-scale attention fusion. Each sample includes various conventional logging curves of historical well sections and corresponding mineral content label information for each component.

8. The multi-component mineral content prediction apparatus of claim 7, wherein, Model building units include: The dataset construction module acquires historical datasets from multiple historical wells and divides them into training, validation, and test sets according to a set ratio. The historical datasets of each historical well include various conventional logging curves and the corresponding mineral contents of each component. The preprocessing module performs data preprocessing on the training set, validation set, and test set respectively. The model training module trains the preset deep learning model by using the training set, introduces a joint loss function during the training, and combines a random weight average strategy and a gradual learning rate scheduling mechanism to iteratively optimize and train the deep learning model until convergence, thereby obtaining the component mineral content prediction model, wherein the deep learning model comprises a feature embedding layer, a 4Mamba-attention encoder, a mineral guided Transformer decoder, a prediction head, and a hybrid loss module; the feature embedding layer performs high-dimensional mapping and adds position encoding; the 4Mamba-attention encoder introduces a bidirectional state space model Mamba, combines a parallel large / small convolution kernel strategy, captures global long-term trends and local sudden fluctuations of the well logging curve, and extracts encoded features; the mineral guided Transformer decoder maps the multi-scale logging features extracted by the encoder to a specific mineral category to obtain decoded features; and the prediction head and the hybrid loss module obtain the component mineral content prediction results according to the decoded features. The model test optimization module verifies and tests the component mineral content prediction model by using the verification set and the test set, respectively, optimizes the model parameters of the component mineral content prediction model, and outputs the component mineral content prediction model that meets the test evaluation requirements.

9. An electronic device, comprising: The computer program is loaded and executed by the processor to realize the steps in the method of any one of claims 1 to 6.

10. A storage medium, characterized by The storage medium stores a computer program readable by a computer, and the computer program is configured to execute the steps in the method of any one of claims 1 to 6 when running.

Images