A small sample reservoir classification method based on meta-transfer learning

Abstract

This invention proposes a small-sample reservoir classification method based on meta-transfer learning, primarily involving artificial intelligence, machine learning, and the petroleum exploration field. The main steps include: designing two case studies based on actual oilfield conditions, utilizing samples from neighboring and more distant blocks respectively to address the problem of insufficient target block samples; designing a masked attention mechanism to avoid interference from non-reservoir elements while learning the relationship between reservoirs and non-reservoir elements, improving the model's feature extraction capability; designing a meta-transfer learning strategy to achieve rapid parameter optimization during transfer learning, improving the model's convergence speed; and designing a value-aware module within meta-transfer learning to enable the model to focus on challenging tasks, learning valuable transfer knowledge and avoiding negative transfer caused by geological differences between blocks. This invention addresses the issue of limited target block data, which hinders model training. By leveraging meta-transfer learning, the reservoir classification performance under two practical case studies is evaluated, and the meta-transfer learning method is optimized, effectively solving the small-sample problem in reservoir classification.

Description

Technical Field

[0001] This invention relates to the fields of data mining, deep learning, and petroleum exploration, specifically to a small-sample reservoir classification method based on meta-transfer learning. Background Technology

[0002] Reservoir classification primarily aims to determine formation fluid properties, which is crucial for improving reservoir evaluation and optimizing production strategies. Traditional methods classify reservoirs by fitting logging data such as porosity and permeability to establish empirical formulas, mainly including data inversion and numerical simulation methods. However, these methods often make assumptions about mathematical parameters during reservoir classification, greatly simplifying the actual formation conditions. Furthermore, traditional methods typically require significant manpower and time, and their identification effectiveness is often limited by human experience. Therefore, data-driven intelligent reservoir identification methods have attracted considerable attention.

[0003] The widespread deployment of oil wells has generated a wealth of Earth observation data, enabling further understanding of geological structures. Traditional machine learning methods have been widely applied in the petroleum industry. Well logging data, due to its continuity and high resolution, can effectively reflect subsurface geological information, and some researchers use data-driven methods for reservoir and lithology identification, such as random forest algorithms, XGBoost algorithms, and Naive Bayes methods. However, these traditional machine learning methods struggle to characterize the complex nonlinear relationships of geological features.

[0004] Deep learning methods have been widely used in reservoir classification and related research due to their effectiveness in extracting geological features from samples. Researchers have combined multilayer perceptrons and oversampling to improve lithology and fluid identification. However, this method often gets stuck in local minima and has weak generalization ability. Some scholars have also proposed a fuzzy process neural network model to classify well logging data to identify water-flooded layers. Recurrent neural network-based methods are widely used for well logging curve classification because they can capture the depth and temporal characteristics of well logging data. Some have proposed using bidirectional long short-term memory networks to learn well logging sequence information for reservoir classification; however, all of the above machine learning methods use expert-labeled data for model-supervised learning, which still leads to error accumulation. Furthermore, the target block has relatively few wells that have been validated through actual oilfield development, meaning there are limited wells with real labels. This is a typical small sample problem in petroleum artificial intelligence.

[0005] Some researchers have proposed transfer learning to address the problem of insufficient samples in petroleum research. This involves defining a domain with enough samples to support model training as the source domain for pre-training the model, and then fine-tuning the pre-trained model using target domain data to achieve knowledge transfer. However, faults divide oilfields into multiple blocks with significant geological differences. Transferring all knowledge learned in the source block to the target block is not always effective; knowledge differing from the target domain inevitably leads to negative transfer. Therefore, there is an urgent need to design an effective method to overcome the small sample problem in reservoir classification. Summary of the Invention

[0006] To address the shortcomings and deficiencies of existing technologies, the present invention aims to propose a small-sample reservoir classification method based on meta-transfer learning.

[0007] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution:

[0008] S1. Based on the actual situation of the oilfield, two cases are designed to solve the problem of insufficient target block samples by using samples from neighboring blocks and distant blocks respectively.

[0009] S2. Design a masked attention mechanism to avoid interference from non-reservoir layers, while learning the relationship between reservoir and non-reservoir layers to improve the model's feature extraction capability.

[0010] S3. Design a meta-transfer learning strategy to achieve rapid parameter optimization in transfer learning and improve the convergence speed of the model.

[0011] S4. In meta-transfer learning, design a value perception module to enable the model to focus on difficult tasks in order to learn valuable transfer knowledge and avoid negative transfer caused by geological differences between blocks.

[0012] In step S1 above, based on the actual situation of the oilfield, two cases are designed to address the problem of insufficient samples in the target block by using samples from neighboring blocks and more distant blocks, respectively. Specifically, these include:

[0013] S11. Due to insufficient reservoir samples in the target block, we considered transferring learning from classification knowledge in another block. Since the geological characteristics of neighboring blocks are relatively similar, we selected samples from neighboring blocks for model pre-training and then transferred the model to the target block; this is Case 1.

[0014] S12. When there are not enough samples in the neighboring blocks to train the model, it is necessary to select a block that is slightly farther away from the target block as the source block for knowledge transfer. This is Case 2.

[0015] In step S2 above, a masked attention mechanism is designed to avoid interference from non-reservoir layers, while simultaneously learning the relationship between reservoir and non-reservoir layers to improve the model's feature extraction capability. Specifically, this includes:

[0016] S21. Perform feature mapping, as follows:

[0017] A reservoir sample contains n logging curves, i.e., n logging features used for model training. A reservoir sample X contains multiple depth points, corresponding to a reservoir category Y, such as oil layer, water layer, etc. The length of the reservoir sample is set to l. Inputting the reservoir sample X into the model for feature mapping can be represented as:

[0018]

[0019] in The dimension of the reservoir sample is denoted by , and → represents the feature mapping. In each round of model training, multiple reservoir samples are fed into the model.

[0020] S22. Avoid interference from non-reservoir elements, specifically as follows:

[0021] Similar to traditional self-attention, masked attention constructs three vectors for each depth point: query Q, key K, and value V. The Q vector of x1 represents the relationship between x1 and other depth points. Assuming x1 is a non-reservoir depth point, the Q vector of x1 is set to zero, thereby automatically masking non-reservoir features.

[0022] S23. Study the relationship between reservoirs and non-reservoirs, as detailed below:

[0023] Next, three vectors, Q, K, and V, are created for depth point x2, where x2 is the reservoir depth point. The Q vector of x2 needs to query the K vectors of other depth points. When querying depth point x1, the K vector of x1 is set to a minimum value, thus taking into account the relationship between the non-reservoir depth point feature x1 and the reservoir depth point x2.

[0024] S24. Calculate the mask attention score to represent the relationship between depth points and implement classification, as detailed below:

[0025] After calculating the mask attention score, the features are residually connected and layer normalized before being fed into the feedforward module. The last feedforward layer maps the hidden layer features, and the number of neurons represents the number of reservoir categories. Reservoir classification is achieved by outputting the predicted probability p(x) of the reservoir category.

[0026] In step S3 above, a meta-transfer learning strategy is designed to achieve rapid parameter optimization and improve the convergence speed of the model through meta-learning in transfer learning. Specifically, this includes:

[0027] S31, Pre-training phase;

[0028] During the pre-training phase, reservoir samples from the source blocks are used to train the model. The last linear layer is the classifier, and the rest are feature extractors. The trained feature extractors are then used for value-aware meta-transfer training to improve the robustness and convergence speed of the reservoir classification method.

[0029] S32, Meta-training phase;

[0030] In meta-training, the model randomly selects samples from the source block reservoir to form a meta-task with a distribution p(T). A meta-task consists of two parts: a support set and a query set. The support set is used to update the base learner, and the query set is used to update the meta-learner. The updated meta-learner is then a trained reservoir classification model.

[0031] S32, Meta-testing phase;

[0032] In the meta-test, the task of sampling from reservoir samples in the target block is also divided into a support set and a query set. The optimized parameters from the meta-training are loaded into the meta-test for initialization, and the base learner is fine-tuned using the support set. Then, the meta-learner is tested using the query set, and the test results are evaluated using accuracy and F1 score to achieve reservoir classification.

[0033] In step S4 above, a value-aware module is designed in meta-transfer learning to enable the model to focus on difficult tasks in order to learn valuable transfer knowledge and avoid negative transfer caused by geological differences between blocks. Specifically, this includes:

[0034] Meta-learning treats each meta-task equally during training; however, in actual reservoir classification, the difficulty of learning meta-tasks varies. Meta-tasks with lower accuracy or F1 scores are defined as difficult tasks, and they should receive more attention from the model. Since the values ​​of accuracy and F1 score are between 0 and 1, a log function is used to scale the loss function.

[0035] Taking accuracy as an example, -log(accuracy) is a very small value when accuracy is high, which will reduce the loss value of the corresponding task. Conversely, -log(accuracy) will amplify the loss when accuracy is low. Therefore, the accuracy Acc of the current task should be set... i and F1 score F1 i Introducing traditional cross-entropy loss In the value-aware module, the loss of each base learner is dynamically adjusted. By assigning greater gradient information to difficult tasks, the model differentiates its focus to different tasks, learning valuable transfer knowledge and accelerating model convergence, thereby improving reservoir classification performance. The value-aware loss VL is represented as:

[0036]

[0037] Where μ and v are weight parameters, and λ is a scaling factor used to adjust the model's focus on difficult tasks.

[0038] The beneficial effects of this invention are as follows: A small-sample reservoir classification method based on meta-transfer learning is proposed. Considering the insufficient reservoir data in the target block, which cannot support model training to achieve reservoir classification, the meta-transfer learning method is used to achieve value-based reservoir classification through a mask attention mechanism and a value perception module. This method has significant application value in reservoir evaluation. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of the reservoir classification process based on meta-transfer learning according to the present invention.

[0041] Figure 2 This is a schematic diagram of the reservoir classification model based on mask attention according to the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Due to insufficient data in the target block, it is difficult to train a model for reservoir classification research. This invention addresses the small sample problem in reservoir classification by proposing a small sample reservoir classification method based on meta-transfer learning. The main steps include: designing two cases based on the actual conditions of the oilfield, respectively using samples from neighboring and more distant blocks to address the insufficient sample size in the target block; designing a masked attention mechanism to avoid interference from non-reservoir areas while learning the relationship between reservoirs and non-reservoirs, improving the model's feature extraction capability; designing a meta-transfer learning strategy to achieve rapid parameter optimization during transfer learning, improving the model's convergence speed; and designing a value-aware module in meta-transfer learning to enable the model to focus on difficult tasks, learning valuable transfer knowledge and avoiding negative transfer caused by geological differences between blocks. This invention optimizes the meta-transfer learning method to effectively solve the small sample problem in reservoir classification, addressing the limited amount of data in the target block.

[0044] A few-sample reservoir classification method based on meta-transfer learning, the specific process is as follows: Figure 1 As shown, the implementation steps are as follows:

[0045] S1. Based on the actual situation of the oilfield, two cases are designed to solve the problem of insufficient target block samples by using samples from neighboring blocks and distant blocks respectively.

[0046] S2. Design a masked attention mechanism to avoid interference from non-reservoir layers, while learning the relationship between reservoir and non-reservoir layers to improve the model's feature extraction capability.

[0047] S3. Design a meta-transfer learning strategy to achieve rapid parameter optimization in transfer learning and improve the convergence speed of the model.

[0048] S4. In meta-transfer learning, design a value perception module to enable the model to focus on difficult tasks in order to learn valuable transfer knowledge and avoid negative transfer caused by geological differences between blocks.

[0049] In step S1 above, based on the actual situation of the oilfield, two cases are designed to address the problem of insufficient samples in the target block by using samples from neighboring blocks and more distant blocks, respectively. Specifically, these include:

[0050] S11. Due to insufficient reservoir samples in the target block, we considered transferring learning from classification knowledge in another block. Since the geological characteristics of neighboring blocks are relatively similar, we selected samples from neighboring blocks for model pre-training and then transferred the model to the target block; this is Case 1.

[0051] S12. When there are not enough samples in the neighboring blocks to train the model, it is necessary to select a block that is slightly farther away from the target block as the source block for knowledge transfer. This is Case 2.

[0052] In step S2 above, a masked attention mechanism is designed to avoid interference from non-reservoir layers, while simultaneously learning the relationship between reservoir and non-reservoir layers to improve the model's feature extraction capability. The model structure is as follows: Figure 2 As shown, it specifically includes:

[0053] S21. Perform feature mapping, as follows:

[0054] A reservoir sample contains n logging curves, i.e., n logging features used for model training. A reservoir sample X contains multiple depth points, corresponding to a reservoir category Y, such as oil layer, water layer, etc. The length of the reservoir sample is set to l. Inputting the reservoir sample X into the model for feature mapping can be represented as:

[0055]

[0056] in The dimension of the reservoir sample is denoted by , and → represents the feature mapping. In each round of model training, multiple reservoir samples are fed into the model.

[0057] S22. Avoid interference from non-reservoir elements, specifically as follows:

[0058] Similar to traditional self-attention, masked attention constructs three vectors for each depth point: query Q, key K, and value V. The Q vector of x1 represents the relationship between x1 and other depth points. Assuming x1 is a non-reservoir depth point, the Q vector of x1 is set to zero, thereby automatically masking non-reservoir features.

[0059] S23. Study the relationship between reservoirs and non-reservoirs, as detailed below:

[0060] Next, three vectors, Q, K, and V, are created for depth point x2, where x2 is the reservoir depth point. The Q vector of x2 needs to query the K vectors of other depth points. When querying depth point x1, the K vector of x1 is set to a minimum value, thus taking into account the relationship between the non-reservoir depth point feature x1 and the reservoir depth point x2.

[0061] S24. Calculate the mask attention score to represent the relationship between depth points and implement classification, as detailed below:

[0062] After calculating the mask attention score, the features are residually connected and layer normalized before being fed into the feedforward module. The last feedforward layer maps the hidden layer features, and the number of neurons represents the number of reservoir categories. Reservoir classification is achieved by outputting the predicted probability p(x) of the reservoir category.

[0063] In step S3 above, a meta-transfer learning strategy is designed to achieve rapid parameter optimization and improve the convergence speed of the model through meta-learning in transfer learning. Specifically, this includes:

[0064] S31, Pre-training phase;

[0065] During the pre-training phase, reservoir samples from the source blocks are used to train the model. The last linear layer is the classifier, and the rest are feature extractors. The trained feature extractors are then used for value-aware meta-transfer training to improve the robustness and convergence speed of the reservoir classification method.

[0066] S32, Meta-training phase;

[0067] In meta-training, the model randomly selects samples from the source block reservoir to form a meta-task with a distribution p(T). A meta-task consists of two parts: a support set and a query set. The support set is used to update the base learner, and the query set is used to update the meta-learner. The updated meta-learner is then a trained reservoir classification model.

[0068] S32, Meta-testing phase;

[0069] In the meta-test, the task of sampling from reservoir samples in the target block is also divided into a support set and a query set. The optimized parameters from the meta-training are loaded into the meta-test for initialization, and the base learner is fine-tuned using the support set. Then, the meta-learner is tested using the query set, and the test results are evaluated using accuracy and F1 score to achieve reservoir classification.

[0070] In step S4 above, a value-aware module is designed in meta-transfer learning to enable the model to focus on difficult tasks in order to learn valuable transfer knowledge and avoid negative transfer caused by geological differences between blocks. Specifically, this includes:

[0071] Meta-learning treats each meta-task equally during training; however, in actual reservoir classification, the difficulty of learning meta-tasks varies. Meta-tasks with lower accuracy or F1 scores are defined as difficult tasks, and they should receive more attention from the model. Since the values ​​of accuracy and F1 score are between 0 and 1, a log function is used to scale the loss function.

[0072] Taking accuracy as an example, -log(accuracy) is a very small value when accuracy is high, which will reduce the loss value of the corresponding task. Conversely, -log(accuracy) will amplify the loss when accuracy is low. Therefore, the accuracy Acc of the current task should be set... i and F1 score F1 i Introducing traditional cross-entropy loss In the value-aware module, the loss of each base learner is dynamically adjusted. By assigning greater gradient information to difficult tasks, the model differentiates its focus to different tasks, learning valuable transfer knowledge and accelerating model convergence, thereby improving reservoir classification performance. The value-aware loss VL is represented as:

[0073]

[0074] Where μ and v are weight parameters, and λ is a scaling factor used to adjust the model's focus on difficult tasks.

[0075] This invention addresses the issue of insufficient reservoir data in the target block, which prevents model training for reservoir classification. Therefore, it utilizes meta-transfer learning methods, employing a masked attention mechanism and a value-aware module to achieve value-aware reservoir classification, which has significant application value in reservoir assessment and well logging interpretation.

[0076] 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 principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A few-sample reservoir classification method based on meta-transfer learning, characterized in that, The specific steps are as follows: S1. Based on the actual situation of the oilfield, two cases are designed to solve the problem of insufficient target block samples by using samples from neighboring blocks and distant blocks respectively. S2. Design a masked attention mechanism to avoid interference from non-reservoir layers, while learning the relationship between reservoir and non-reservoir layers to improve the model's feature extraction capability. S3. Design a meta-transfer learning strategy to achieve rapid parameter optimization in transfer learning and improve the convergence speed of the model. S4. A value-aware module is designed in meta-transfer learning to enable the model to focus on difficult tasks, thereby learning valuable transfer knowledge and avoiding negative transfer caused by geological differences between blocks. The feature is that a value-aware module is designed in meta-transfer learning to enable the model to focus on difficult tasks, thereby learning valuable transfer knowledge and avoiding negative transfer caused by geological differences between blocks, including: Meta-learning treats each meta-task equally during training; however, in actual reservoir classification, the difficulty of learning meta-tasks varies, with meta-tasks having lower accuracy or F1 scores defined as difficult tasks. Since accuracy and F1 scores range from 0 to 1, they are used in a more nuanced way. The function scales the loss function; When accuracy is high, it is a very small value, which will reduce the loss value of the corresponding task; conversely, Lower accuracy will amplify the loss; therefore, the accuracy of the current task should be... and F1 score Introducing traditional cross-entropy loss In the value-aware module, the loss of each base learner is dynamically adjusted. By allocating greater gradient information to difficult tasks, the model differentiates its focus to different tasks, learning valuable transfer knowledge and accelerating model convergence, thereby improving reservoir classification performance. (Value-aware loss) Represented as: in , For weight parameters, This is a scaling factor used to adjust the model's focus on difficult tasks.

2. The small-sample reservoir classification method based on meta-transfer learning according to claim 1, characterized in that, In step S1, based on the actual situation of the oilfield, two cases are designed to address the problem of insufficient samples in the target block by utilizing samples from neighboring blocks and more distant blocks, respectively. These mainly include: S11. Since there are not enough reservoir samples in the target block, we consider transferring the knowledge of classification in another block to the target block. The geological characteristics of neighboring blocks are relatively similar. We select samples from neighboring blocks for model pre-training and then transfer the knowledge to the target block. This is Case 1. S12. When there are not enough samples in the neighboring blocks to train the model, it is necessary to select a block that is slightly farther away from the target block as the source block for knowledge transfer. This is Case 2.

3. The small-sample reservoir classification method based on meta-transfer learning according to claim 1, characterized in that, In step S2, a masked attention mechanism is designed to avoid interference from non-reservoir layers, while simultaneously learning the relationship between reservoir and non-reservoir layers to improve the model's feature extraction capability, including: S21. Perform feature mapping, as follows: A reservoir sample contains The logging curve, i.e. One logging feature is used for model training, and one reservoir sample is used. It contains multiple depth points, corresponding to a reservoir category. The length of the reservoir sample is set to reservoir samples The input is fed into the model for feature mapping, represented as: in The dimension of the reservoir sample. This represents feature mapping, where multiple reservoir samples are fed into the model during each round of model training; S22. Avoid interference from non-reservoir elements, specifically as follows: Masked attention constructs a query for each depth point ,key ,value Three vectors, of Vectors are for querying Relationship with other depth points, assuming For non-reservoir depth points, of The vector is set to zero, thereby automatically masking non-reservoir features in this way; S23. Study the relationship between reservoirs and non-reservoirs, as detailed below: Next is the depth point. create , , Three vectors, For reservoir depth points, of The vector needs to query other depth points. Vector, when querying depth points hour, vector Set to a minimum value to take into account characteristics at non-reservoir depth points. and reservoir depth points The relationship between them; S24. Calculate the mask attention score to represent the relationship between depth points and implement classification, as detailed below: After calculating the mask attention score, the features are residually connected and layer normalized before being fed into the feedforward module. The last feedforward layer maps the hidden features, and the number of neurons represents the number of reservoir classes. The predicted probability of each reservoir class is then output. Achieve reservoir classification.

Images