Lithofacies classification system combined with optimization algorithm

Abstract

The invention provides a lithofacies classification system combined with an optimization algorithm, and the system comprises a data obtaining module which is used for obtaining logging data from an oil and gas field exploration project; the data filling module is used for filling missing values in the logging data; the balance sampling module is used for balancing the proportion of different lithofacies categories in the training data by using an improved oversampling technology; the hybrid model identification module is used for constructing a hybrid deep learning model, inputting the logging data processed by the balanced sampling module into the hybrid deep learning model to identify logging data features, and outputting a lithofacies classification result of each data point; and the report generation module is used for generating a lithofacies classification report according to the lithofacies classification result output by the mixed deep learning model. According to the system, an efficient and accurate lithofacies classification system is provided for geologists and petroleum engineers, and the efficiency and benefits of oil and gas exploration and development can be improved.

Description

Technical Field

[0001] The invention relates to the technical field of geological exploration, and in particular to a lithofacies classification system combined with an optimization algorithm. Background Art

[0002] In the process of oil and gas exploration and development, lithofacies classification is crucial, which helps to understand the reservoir structure, predict the distribution of oil and gas reservoirs, and formulate efficient exploration and production strategies. However, the current lithofacies classification technology has significant deficiencies and is difficult to meet the actual production needs. The main problems are data missing and integrity problems. Well logging data are usually missing, noisy and inconsistent in quality due to well site operation errors, equipment failures and other reasons, which affect the integrity of the data. Traditional lithofacies classification methods are difficult to effectively deal with the problem of missing data, resulting in the difficulty of the classification model to obtain stable and reliable results; data imbalance problem. Data of different lithofacies categories are usually unbalanced. Rare lithofacies samples are easily ignored by the classification model, resulting in the classification results being biased towards categories with a high proportion of data, thereby reducing the classification accuracy of the model; complex geological data characteristics. Geological data are diverse and complex, and there are a large number of nonlinear relationships. Traditional classification algorithms have limited capabilities in feature extraction and data analysis, and it is difficult to accurately handle the complex relationships in logging data, reducing the efficiency and accuracy of lithofacies classification.

[0003] In recent years, with the rapid development of machine learning and deep learning technologies, these technologies have shown significant potential in geological data processing. In particular, convolutional neural networks (CNN) and long short-term memory networks (LSTM), which have powerful feature extraction and time series analysis capabilities, respectively, can effectively process complex well logging data. However, the direct application of deep learning technology to lithofacies classification still faces some challenges, including data missing and data imbalance. Existing deep learning models cannot directly handle the problem of missing data. The lack of well logging data will lead to unstable model training and affect the final classification results. At the same time, different lithofacies categories in well logging data are usually unbalanced. Traditional models may tend to predict the majority category during training, resulting in reduced prediction accuracy. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a lithofacies classification system combined with an optimization algorithm, which ensures the integrity of logging data by introducing the hot card filling technology of the optimization algorithm; utilizes improved oversampling technology to enhance the ability of the classification model to handle data imbalance; and realizes lithofacies classification with the help of the feature extraction and sequence analysis capabilities of the hybrid deep learning model.

[0005] To achieve the above-mentioned object, the present invention provides a lithofacies classification system combined with an optimization algorithm, the system comprising:

[0006] Data acquisition module: used to obtain logging data from oil and gas field exploration projects. The logging data of each well forms a multi-dimensional feature vector according to the sampling depth;

[0007] Data filling module: used to balance the proportion of different facies categories in training data using improved oversampling technology;

[0008] Balanced sampling module: used to balance the data ratio between different lithofacies categories in the training logging data;

[0009] Hybrid model identification module: used to build a hybrid deep learning model, input the well logging data processed by the balanced sampling module into the hybrid deep learning model to identify the well logging data features, and output the lithofacies classification results of each data point;

[0010] Report generation module: used to generate a lithofacies classification report based on the lithofacies classification results output by the hybrid deep learning model.

[0011] Furthermore, the missing values ​​in the logging data are filled by using the hot card filling technology, specifically:

[0012] S11, identifying missing values ​​in the well logging data set, for each well logging data with missing values, looking for one or more similar complete records in the data set, dividing all complete records into different groups according to the characteristics of the well logging data, and the records in each group are similar in key variables as potential donors;

[0013] S12, using a statistical method to determine the missing record that is most similar to the complete record, randomly selecting one or more records from a donor pool as donors, and using their values ​​to fill in the missing values ​​of the well logging data, wherein the donor pool is a collection of donor samples that can be used to fill in the missing values ​​of the well logging data;

[0014] S13, assigning the selected donor's value to the corresponding variable of the missing record of the well logging data to complete the filling of the missing value of the well logging data, and repeating steps S11 to S13 for all the missing values ​​of the well logging data in the data set until all the missing values ​​of the well logging data are filled.

[0015] Furthermore, an improved oversampling technique is used to balance the proportion of different lithofacies in the training data, specifically:

[0016] S21, marking the minority class samples on the boundary of the well logging data processed by the data filling module, calculating the Euclidean distances from the minority class samples to all other samples, and selecting K samples from the minority class samples with the shortest Euclidean distances to all other samples;

[0017] S22, dividing the K samples into safe class samples, boundary class samples and noise class samples, calculating the Euclidean distance between the boundary class samples and all other minority class samples, and selecting the samples with the shortest spatial distance from the boundary class samples to all other minority class samples;

[0018] S23, selecting a sample with the shortest time span from the samples with the shortest spatial distance in step S23, and synthesizing the sample with the shortest spatial distance and the sample with the shortest time span into a new sample point;

[0019] S24, adding the new sample point to the minority class samples, calculating the sample category ratio around the minority class samples, finding the noise class samples and deleting them.

[0020] Furthermore, the hybrid model recognition module includes a spatial feature extraction unit, a time series feature extraction unit and a data feature extraction unit. The spatial feature extraction unit is used to extract the spatial features of the logging data, the time series feature extraction unit is used to process the time series features of the logging data, and the data feature extraction unit is used to process the long-distance dependencies in the logging data.

[0021] Furthermore, the spatial feature extraction unit uses CNN to identify spatial patterns and structures in the logging data through multi-layer convolution and pooling operations, specifically:

[0022] S31, taking the logging data as the input of CNN, and calculating the similarity weight between the input logging data and the prediction target through CNN;

[0023] S32, using the Softmax function to normalize the similarity weight to obtain a normalized weight;

[0024] S33, weighting the input logging data to obtain key feature information of the logging data, and using a local filter and a sliding filter to capture the spatial features between the logging data.

[0025] Furthermore, the time series feature extraction unit uses the memory capacity of LSTM to capture the time series relationship and trend changes in the logging data, specifically:

[0026] S41, using the output result of CNN as the input information of LSTM to obtain comprehensive information of well logging data;

[0027] S42, building a connection between current key information and historical information of logging data through an LSTM recurrent neural network layer, calculating a time similarity weight, and normalizing the time similarity weight using a Softmax function to obtain a normalized time similarity weight;

[0028] S43, weighting the normalized time similarity weight and the comprehensive information of the logging data output by LSTM to obtain the key time features of the logging data.

[0029] Furthermore, the data feature extraction unit uses a transformer to capture long-distance dependencies between sequences and extracts complex features from the logging data, specifically:

[0030] S51, the output result of LSTM is linearly transformed and divided into multiple attention heads, each of which performs independent self-attention calculation;

[0031] S52, using multiple attention heads to process different subspace information in parallel, calculating the similarity between each subspace and other subspaces to obtain a weight coefficient, and performing normalization operation through the softmax function to obtain the attention weight of each position to other positions;

[0032] S53, multiply the attention weight with the corresponding value vector and sum them, concatenate the outputs of each attention head, and perform a final linear transformation to obtain the self-attention representation of each position to capture the long-distance dependencies in the logging data sequence.

[0033] Furthermore, the hybrid deep learning model performs lithofacies classification, specifically:

[0034] S61, inputting the padded and balanced logging data into the hybrid deep learning model;

[0035] S62, extracting spatial features, time series features and long-distance dependencies of complex features in the logging data and performing comprehensive analysis, and outputting the lithofacies category probability and final lithofacies classification result corresponding to each logging data point.

[0036] Compared with the prior art, the present invention has the following beneficial effects:

[0037] (1) Solve the problem of missing data: By adopting the hot card filling technology, the missing values ​​in the logging data are compensated, the integrity and consistency of the data are maintained, and reliable basic data is provided for the training of the hybrid deep learning model, which helps to improve the stability and accuracy of lithofacies classification.

[0038] (2) Dealing with data imbalance: Using improved oversampling technology to balance training samples and improve the data ratio between different lithofacies categories, the hybrid deep learning model can learn different lithofacies characteristics more evenly and improve the recognition ability of rare categories.

[0039] (3) Improve classification accuracy: By combining the feature extraction capability of CNN, the time series analysis capability of LSTM, and the long-distance dependency processing capability of the transformer, a hybrid deep learning model is established to improve the accuracy and reliability of lithofacies classification.

[0040] (4) Adaptability to practical applications: The system provides an efficient and accurate technical solution for lithofacies classification in oil and gas exploration and development, meets the needs of actual production, helps optimize exploration and development strategies, and improves the development efficiency and benefits of oil and gas resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only preferred embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 It is a schematic diagram of the overall structure of a lithofacies classification system combined with an optimization algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The enumerated embodiments are only used to explain the present invention and are not used to limit the scope of the present invention.

[0044] Reference Figure 1 This embodiment provides a lithofacies classification system combined with an optimization algorithm, the system comprising:

[0045] Data acquisition module: used to obtain logging data from oil and gas field exploration projects. The logging data of each well forms a multi-dimensional feature vector according to the sampling depth. Obtain logging data from oil and gas field exploration projects, including various logging curves such as acoustic wave, density, gamma, resistivity, etc.

[0046] Data filling module: used to fill the missing values ​​in the logging data by hot deck imputation technology. Hot deck imputation technology is used to process the missing values ​​in the logging data. The missing items are filled by randomly selecting appropriate values ​​from a complete similar data set to ensure the integrity and consistency of the data.

[0047] Furthermore, the missing values ​​in the logging data are filled by using the hot card filling technology, specifically:

[0048] S11. Identify missing values ​​in the well logging data set. For each well logging data with missing values, find one or more complete records similar to it in the data set. Divide all complete records into different groups according to the characteristics of the well logging data. The records in each group are similar in key variables and serve as potential donors.

[0049] S12. Use statistical methods to determine the missing records that are most similar to the complete records, randomly select one or more records from the donor pool as donors, and use their values ​​to fill the missing values ​​of well logging. The donor pool is a collection of donor samples that can be used to fill the missing values ​​of well logging data.

[0050] S13, assigning the selected donor's value to the corresponding variable of the missing record of the well logging data to complete the filling of the missing value of the well logging data, and repeating steps S11 to S13 for all the missing values ​​of the well logging data in the data set until all the missing values ​​of the well logging data are filled.

[0051] Hot card filling is a statistical missing data processing technology. Its basic principle is to use complete records in the data set to fill missing values. This method believes that in the data set, although some records are missing in one or some variables, they are similar to other records in other variables. Therefore, the corresponding values ​​of these similar records can be used to fill the missing values ​​of the well logging data. This method can better maintain the distribution characteristics of the well logging data and reduce the problem of inaccurate lithofacies classification caused by improper processing of missing values ​​in the well logging data. Hot card filling finds the most similar complete samples in the well logging data set to fill the missing values, and uses the values ​​of these similar samples to fill the missing values. This method can use the existing intrinsic relationships in the well logging data set to provide a more reasonable missing value estimate. The data processed by hot card filling can enable the subsequent hybrid deep learning model to better capture the patterns and trends in the well logging data, thereby improving the performance and prediction accuracy of the hybrid deep learning model.

[0052] Balanced sampling module: used to balance the proportion of different lithofacies in the training data using improved oversampling techniques. After preprocessing, unbalanced data processing techniques are used to balance the data proportions of different lithofacies, generate synthetic new samples on lithofacies with scarce data samples, expand the data set, and make the training data more balanced.

[0053] An improved oversampling technique is used to balance the proportions of different lithofacies in the training data, specifically:

[0054] S21. Mark the minority class samples on the boundary of the well logging data processed by the data filling module, calculate the Euclidean distance from the minority class samples to all other samples, and select K samples with the shortest Euclidean distance from the minority class samples to all other samples.

[0055] S22, divide the K samples into safe class samples, boundary class samples and noise class samples, calculate the Euclidean distance from the boundary class samples to all other minority class samples, and select the boundary class samples with the shortest spatial distance to all other minority class samples. Boundary class samples are usually located on the decision boundary between minority class samples and majority class samples, and are easily misclassified by the classifier.

[0056] S23, select the sample with the shortest time span from the samples with the closest spatial distance in step S23, and synthesize the sample with the closest spatial distance and the sample with the shortest time span into a new sample point. These synthesized new sample points expand the range of minority class samples and enhance the learning ability of the hybrid deep learning model for minority class features.

[0057] S24, adding the new sample point to the minority class samples, calculating the sample category ratio around the minority class samples, finding the noise class samples and deleting them.

[0058] Unbalanced data can cause the hybrid deep learning model to favor majority class samples. In logging data, minority classes may represent some important geological phenomena or reservoir characteristics. In actual oil and gas exploration, being able to accurately identify all lithology categories is crucial for evaluating reservoir quality and formulating exploration strategies. Through unbalanced data processing technology, it can be ensured that these important information is not ignored by the hybrid deep learning model, thereby improving the accuracy of lithofacies classification. By increasing the number of minority class samples, it is possible to avoid the hybrid deep learning model from being biased against majority class samples during training, thereby improving the hybrid deep learning model's ability to recognize minority class samples. By synthesizing new samples, the diversity of the data set is increased, which helps the hybrid deep learning model learn a more comprehensive data representation in the lithology identification of complex reservoirs, thereby improving the accuracy of lithofacies identification.

[0059] Hybrid model identification module: used to build a hybrid deep learning model, input the logging data processed by the balanced sampling module into the hybrid deep learning model to identify the logging data characteristics, and output the lithofacies classification results of each data point.

[0060] Furthermore, the hybrid model recognition module includes a spatial feature extraction unit, a time series feature extraction unit and a data feature extraction unit. The spatial feature extraction unit is used to extract the spatial features of the logging data, the time series feature extraction unit is used to process the time series features of the logging data, and the data feature extraction unit is used to process the long-distance dependencies in the logging data.

[0061] The hybrid deep learning model improves the generalization ability of the model for unknown well logging data by combining different network structures. Well logging data may contain complex spatial and temporal structures. The hybrid deep learning model can better handle these complex structures, improve the depth and breadth of data processing, and make it more stable and reliable in practical applications. At the same time, the hybrid deep learning model has strong robustness and generalization ability, which can reduce the risk of overfitting, improve the performance of the model on different well logging data sets, and improve the accuracy of lithofacies classification.

[0062] The spatial feature extraction unit uses CNN to identify spatial patterns and structures in well logging data through multi-layer convolution and pooling operations, specifically:

[0063] S31. Use the logging data as the input of CNN, and use CNN to calculate the similarity weight between the input logging data and the predicted target.

[0064] S32. Use the Softmax function to normalize the similarity weight to obtain a normalized weight.

[0065] S33, weighting the input logging data to obtain key feature information of the logging data, and using a local filter and a sliding filter to capture the spatial features between the logging data.

[0066] According to the first law of geography, the spatial correlation of logging representation and reservoir lithofacies classification targets has a great influence on recognition accuracy. CNN can capture the spatial characteristics of different logging sequences at the same depth, thereby mining more logging information. CNN can perform translation-invariant classification of input information according to its hierarchical structure. The nonlinear fitting ability and weight sharing characteristics of the network reduce the amount of computation in the entire feature extraction process, which helps the constructed hybrid deep learning model to extract more logging spatial information from the nonlinear relationship between representation and lithofacies, thereby improving the accuracy of the hybrid deep learning model, reducing the number of parameters and reducing computational costs.

[0067] The time series feature extraction unit uses the memory capacity of LSTM to capture the time series relationship and trend changes in the logging data, specifically:

[0068] S41. Use the output of CNN as the input information of LSTM to obtain comprehensive information of well logging data.

[0069] S42. The connection between the current key information and the historical information of the logging data is constructed through the LSTM recurrent neural network layer, the time similarity weight is calculated, and the time similarity weight is normalized using the Softmax function to obtain the normalized time similarity weight.

[0070] S43, weighting the normalized time similarity weight and the comprehensive information of the logging data output by LSTM to obtain the key time features of the logging data.

[0071] LSTM automatically extracts the key features of logging time within a certain depth range based on the correlation between the current key logging information and the historical time series information of the CNN output results during the network training process. LSTM uses its unique memory gate to store the historical logging depth information, reduce information loss, improve information transmission efficiency, and provide the converter with comprehensive logging information within a certain depth range. Extracting the key features of logging time based on LSTM improves the time series feature extraction capability of the hybrid deep learning model, and solves the network loss and other performance degradation problems caused by long training time.

[0072] The data feature extraction unit uses a transformer to capture long-distance dependencies between sequences and extracts complex features from logging data, specifically:

[0073] S51. The output result of LSTM is linearly transformed and divided into multiple attention heads, and each attention head performs independent self-attention calculation.

[0074] S52. Use multiple attention heads to process different subspace information in parallel, calculate the similarity between each subspace and other subspaces to obtain a weight coefficient, and perform normalization operation through the softmax function to obtain the attention weight of each position to other positions.

[0075] S53, multiply the attention weight with the corresponding value vector and sum them, concatenate the outputs of each attention head, and perform a final linear transformation to obtain the self-attention representation of each position to capture the long-distance dependencies in the logging data sequence.

[0076] In the transformer, the sequence data is processed by the self-attention mechanism to capture the time dependency in the logging data, which helps to understand the changes in lithofacies. The self-attention mechanism calculates the similarity between each position and other positions to obtain a weight coefficient, which is used to weight the correlation between the positions, that is, self-attention calculates the similarity scores between different vectors, which reflect the correlation between the vector at a specific position and other vectors. For each position in the input sequence, three vectors are obtained through linear transformation: query vector, key vector, and value vector. The dot product of the query vector and all key vectors is calculated, and then normalized by the softmax function to obtain the attention weight of each position to other positions. Finally, these attention weights are multiplied and summed with the corresponding value vector to obtain the self-attention representation of each position. The output of LSTM is linearly transformed and divided into multiple attention heads. Each attention head performs independent self-attention calculation. Finally, the outputs of each self-attention head are spliced ​​and the final output is obtained after the final linear transformation. By processing different subspace information in parallel through multiple attention heads, the long-distance dependencies in the sequence can be better captured. This mechanism enables the transformer to perform well in processing long sequence tasks and has strong parallelization capabilities, which improves the accuracy and durability of lithology prediction and reduces the possibility of overfitting.

[0077] The hybrid deep learning model performs lithofacies classification, specifically:

[0078] S61. Input the padded and balanced logging data into the hybrid deep learning model.

[0079] S62, extracting spatial features, time series features and long-distance dependencies of complex features in the logging data and performing comprehensive analysis, and outputting the lithofacies category probability and final lithofacies classification result corresponding to each logging data point.

[0080] The balanced logging data is input into the hybrid deep learning model for training. The cross-validation method is used to evaluate the model performance and adjust the model parameters to ensure that the model can accurately distinguish different lithofacies. The model performance is evaluated using indicators such as classification accuracy, confusion matrix and AUC value. The network structure and learning rate are adjusted according to the results to optimize the model's feature extraction and classification capabilities. The processed logging data is input into the trained model for lithofacies classification prediction. The model will output the lithofacies category probability and the final prediction result corresponding to each piece of data.

[0081] Report generation module: used to generate a lithofacies classification report from the lithofacies classification results output by the hybrid deep learning model. The lithofacies classification report includes the lithofacies classification results, that is, the lithofacies classification results of each logging section are listed, including the predicted lithofacies type and the corresponding confidence or probability value; model performance evaluation, providing performance indicators of the model on the training set and test set, such as accuracy, recall rate, F1 score, etc.; feature importance analysis, analyzing which features have the greatest impact on the model's classification decision and how these features are associated with lithofacies types; conclusions and suggestions, based on the classification results of the model, geological interpretations and exploration suggestions are proposed.

[0082] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A lithofacies classification system combined with an optimization algorithm, characterized in that: The system comprises: Data acquisition module: used to obtain logging data from oil and gas field exploration projects. The logging data of each well forms a multi-dimensional feature vector according to the sampling depth; Data filling module: used to fill missing values ​​in the well logging data by using hot card filling technology; Balanced sampling module: used to balance the proportion of different facies categories in training data using improved oversampling techniques; Hybrid model identification module: used to build a hybrid deep learning model, input the well logging data processed by the balanced sampling module into the hybrid deep learning model to identify the well logging data features, and output the lithofacies classification results of each data point; Report generation module: used to generate a lithofacies classification report based on the lithofacies classification results output by the hybrid deep learning model.

2. The petrographic classification system combined with the optimization algorithm according to claim 1, characterized in that: The missing values ​​in the logging data are filled by using the hot card filling technology, specifically: S11, identifying missing values ​​in the well logging data set, for each well logging data with missing values, looking for one or more similar complete records in the data set, dividing all complete records into different groups according to the characteristics of the well logging data, and the records in each group are similar in key variables as potential donors; S12, using a statistical method to determine the missing record that is most similar to the complete record, randomly selecting one or more records from a donor pool as donors, and using their values ​​to fill in the missing values ​​of the well logging data, wherein the donor pool is a collection of donor samples that can be used to fill in the missing values ​​of the well logging data; S13, assigning the selected donor's value to the corresponding variable of the missing record of the well logging data to complete the filling of the missing value of the well logging data, and repeating steps S11 to S13 for all the missing values ​​of the well logging data in the data set until all the missing values ​​of the well logging data are filled.

3. The petrographic classification system combined with the optimization algorithm according to claim 1, characterized in that: An improved oversampling technique is used to balance the proportions of different lithofacies in the training data, specifically: S21, marking the minority class samples on the boundary of the well logging data processed by the data filling module, calculating the Euclidean distances from the minority class samples to all other samples, and selecting K samples from the minority class samples with the shortest Euclidean distances to all other samples; S22, dividing the K samples into safe class samples, boundary class samples and noise class samples, calculating the Euclidean distance between the boundary class samples and all other minority class samples, and selecting the samples with the shortest spatial distance from the boundary class samples to all other minority class samples; S23, selecting a sample with the shortest time span from the samples with the shortest spatial distance in step S23, and synthesizing the sample with the shortest spatial distance and the sample with the shortest time span into a new sample point; S24, adding the new sample point to the minority class samples, calculating the sample category ratio around the minority class samples, finding the noise class samples and deleting them.

4. The petrographic classification system combined with an optimization algorithm according to claim 1, characterized in that: The hybrid model recognition module includes a spatial feature extraction unit, a time series feature extraction unit and a data feature extraction unit. The spatial feature extraction unit is used to extract the spatial features of the logging data, the time series feature extraction unit is used to process the time series features of the logging data, and the data feature extraction unit is used to process the long-distance dependencies in the logging data.

5. The petrographic classification system combined with the optimization algorithm according to claim 4, characterized in that: The spatial feature extraction unit uses CNN to identify spatial patterns and structures in well logging data through multi-layer convolution and pooling operations, specifically: S31, taking the logging data as the input of CNN, and calculating the similarity weight between the input logging data and the prediction target through CNN; S32, using the Softmax function to normalize the similarity weight to obtain a normalized weight; S33, weighting the input logging data to obtain key feature information of the logging data, and using a local filter and a sliding filter to capture the spatial features between the logging data.

6. The petrographic classification system combined with the optimization algorithm according to claim 5, characterized in that: The time series feature extraction unit uses the memory capacity of LSTM to capture the time series relationship and trend changes in the logging data, specifically: S41, using the output result of CNN as the input information of LSTM to obtain comprehensive information of well logging data; S42, building a connection between current key information and historical information of logging data through an LSTM recurrent neural network layer, calculating a time similarity weight, and normalizing the time similarity weight using a Softmax function to obtain a normalized time similarity weight; S43. Weighting the normalized time similarity weight and the comprehensive information of the logging data output by LSTM to obtain the key time features of the logging data.

7. The petrographic classification system combined with the optimization algorithm according to claim 6, characterized in that: The data feature extraction unit uses a transformer to capture long-distance dependencies between sequences and extracts complex features from logging data, specifically: S51, the output result of LSTM is linearly transformed and divided into multiple attention heads, each of which performs independent self-attention calculation; S52, using multiple attention heads to process different subspace information in parallel, calculating the similarity between each subspace and other subspaces to obtain a weight coefficient, and performing normalization operation through the softmax function to obtain the attention weight of each position to other positions; S53, multiply the attention weight with the corresponding value vector and sum them, concatenate the outputs of each attention head, and perform a final linear transformation to obtain the self-attention representation of each position to capture the long-distance dependencies in the logging data sequence.

8. The petrographic classification system combined with an optimization algorithm according to claim 1, characterized in that: The hybrid deep learning model performs lithofacies classification, specifically: S61, inputting the padded and balanced logging data into the hybrid deep learning model; S62, extracting spatial features, time series features and long-distance dependencies of complex features in the logging data and performing comprehensive analysis, and outputting the lithofacies category probability and final lithofacies classification result corresponding to each logging data point.

Images