Gamma logging curve correction and prediction method based on federal cross-domain collaboration

Abstract

The invention relates to the technical field of oil exploration and development, and discloses a gamma logging curve correction and prediction method based on federal cross-domain collaboration. Aiming at the problems that traditional prediction depends on single-oil-field data, data scarcity and privacy leakage, historical logging-while-drilling data of multiple oil fields are collected and preprocessed, and a gamma logging curve is corrected in combination with the distance between a drill bit and a sensor; a time sequence prediction model based on a self-attention mechanism is constructed, a federated learning framework is established, and a parameter strategy is initialized; and respectively training a global model and a local model, and comparing and determining a final model through index evaluation such as a relative error. According to the method, on the premise that data privacy is protected, the gamma logging curve prediction precision is improved through the cross-oil-field data, and the real-time geological modeling requirement is met.

Description

Technical Field

[0001] This invention belongs to the field of petroleum exploration and development technology, and specifically relates to a method for gamma logging curve correction and prediction based on federal cross-domain collaboration. Background Technology

[0002] Gamma-ray logging, as a core technology for identifying formation lithology and characterizing reservoir properties, is widely used in key scenarios such as shale oil and gas reservoir identification, sedimentary cycle delineation, and reservoir property evaluation. However, conventional logging techniques are limited by the significant time lag between data acquisition systems and drilling operations, making it difficult to meet the high-precision requirements of real-time geological modeling. In recent years, artificial intelligence (AI) technology has been profoundly reshaping the theoretical framework and practical models of the petroleum industry. At the same time, the development of AI technology increasingly relies on high-quality, large-scale datasets, and model performance is directly limited by the diversity and sample size of the data. Although AI brings new opportunities to the petroleum industry, this field still faces the dual core challenges of data acquisition and application. First, in the field of intelligent drilling engineering, due to the axial distance between the measurement unit and the drill bit, geological sensors generally lag behind the drill string assembly by 5-18 meters in logging-while-drilling technology. This dynamic delay severely affects the real-time accuracy of geological modeling. Second, constrained by data privacy protection regulations and industry technical barriers, the scope of data sharing between oilfields is limited, leading to a significant "data silo" phenomenon, and cross-regional data integration faces multiple obstacles. While transfer learning can improve a model's generalization ability in scenarios with limited data, its reliance on the availability of original data makes it difficult to meet strict privacy protection requirements. Federated learning, as a distributed machine learning method, can train high-precision prediction models jointly from multiple oilfields without sharing the original data. Therefore, developing a method that, while protecting data privacy, corrects logging-while-drilling data and uses cross-oilfield data for gamma-ray logging curve prediction has significant practical application value. Summary of the Invention

[0003] To overcome the shortcomings of the existing technology, the present invention aims to provide a gamma-ray logging curve correction and prediction method based on federated cross-domain collaboration. This method alleviates the data scarcity problem while protecting data privacy by achieving depth-synchronous calibration of measurement-while-drilling data and utilizing a deep learning model architecture to achieve high-precision intelligent prediction of gamma-ray logging curves. Its main steps include:

[0004] S1. Collect historical logging-while-drilling data from different oilfields and the distance between the corresponding logging-while-drilling tool drill bit and sensor, and perform data preprocessing on the historical logging-while-drilling data;

[0005] S2. Based on the preprocessed data in step S1, perform gamma logging curve correction according to the distance between the drill bit and the sensor of the logging-while-drilling tool, and construct training sets and test sets for different oilfields.

[0006] S3. Construct a time-series prediction model for gamma logging curves, which can be used as a tool for implementing federated learning algorithms;

[0007] S4. Based on the model built in step S3, build a federated learning framework, including a global model and a client model, and initialize the model parameter filtering and model parameter aggregation strategies of the federated learning framework.

[0008] S5. Based on the model constructed in step S3, establish local models in each oilfield as comparative experimental models;

[0009] S6. The global model is trained based on the federated learning framework built in step S4, and the local model built in step S3 is trained in each oil field. The trained global model and local model are used to predict gamma logging values ​​for the test set of each oil field, and the model evaluation is calculated to obtain the model evaluation index calculation results.

[0010] S7. Based on the calculation results of the global model and local model evaluation indicators in step S6, analyze and compare them to determine the final prediction model.

[0011] In step S1, historical logging-while-drilling data from different oilfields are collected, including historical gamma logging curve data and corresponding real-time drilling data. The characteristics of historical gamma logging curve data include gamma measurement depth and gamma logging value. The characteristics of real-time drilling data include measurement depth, drilling pressure, standpipe pressure, torque, mechanical drilling speed, rotational speed, mud rheological parameters, and hook load.

[0012] The data preprocessing in step S1 includes: performing preliminary cleaning on the collected data to remove outlier data exceeding three times the standard deviation; performing linear interpolation on missing data; and performing cubic Lagrange interpolation on data with negative characteristic gamma logging values.

[0013] In step S2, the pre-processed data is used to perform gamma logging curve correction based on the distance between the drill bit and the sensor in the logging-while-drilling tool, and training and test sets are constructed for different oilfields. The process includes:

[0014] S2.1 Extract historical gamma logging data, corresponding real-time drilling data, and the distance between the drill bit and sensor in the logging-while-drilling tool;

[0015] S2.2 For the gamma measurement depth sampling interval, based on the fixed depth difference between the drill bit and the sensor, all gamma measurement depth values ​​are summed with the fixed depth difference between the drill bit and the sensor to obtain a new depth value. The gamma logging value corresponding to the new depth value is consistent with the original gamma measurement depth, and finally, new historical gamma logging curve data are obtained. The specific formula is as follows:

[0016]

[0017] Among them, BD i The depth value is measured using gamma, and ΔX is the fixed depth difference between the drill bit and the sensor. The new depth value is given, where n is the number of samples.

[0018] S2.3 The new historical gamma logging curve data obtained from S2.2 is merged with the real-time drilling data according to depth to finally obtain a new drilling dataset;

[0019] S2.4 The data collected from each oilfield are corrected for gamma logging curves according to S2.1 to S2.3, and the final dataset is used to construct training and test sets for each oilfield.

[0020] In step S3, the process of constructing a time-series prediction model for gamma logging curves includes:

[0021] S3.1. A deep learning model based on an encoder-decoder is constructed using a self-attention mechanism as the core framework. The model consists of an input sequence encoding layer, a position encoding layer, a softmax layer, a forward propagation layer, a multi-head attention mechanism layer, a masked multi-head attention mechanism layer, and a linear layer.

[0022] S3.2 Set the model dimension to 256, the model layer to 2, the model neuron random dropout rate to 0.1, and the model learning rate to 0.0001;

[0023] S3.3 The model optimizer uses the Adam optimizer for parameter updates; the loss function uses the mean squared error, as shown in the following formula:

[0024]

[0025] Where yi is the actual gamma logging value. The model predicts the gamma logging value, and n is the number of samples.

[0026] In step S4, a federated learning framework is built, including a global model and a client model, and the model parameter filtering and aggregation strategies of the federated learning framework are initialized, further including:

[0027] S4.1. Establish client models for each oil field based on the number of oil fields, and establish a server-side model containing a global model.

[0028] S4.2 Global Model Parameter Distribution: The global model distributes model parameters to each client, including feature embedding parameters, attention mechanism parameters, and feedforward network parameters, to ensure that each client has a unified initial state;

[0029] S4.3 Client Model Training and Parameter Upload: Each client model uses the training set of each oilfield for model training, and transmits its own model parameters to the server in real time for each training batch.

[0030] S4.4 Parameter Filtering: After receiving the parameters from each client, the server first filters the parameters, using a normalization strategy that does not rely on batch statistics to filter out parameters that skip batch normalization.

[0031] S4.5 Model Parameter Partitioning: Based on the model structure, the model parameters are divided into a local private layer and a shared layer; the parameters in the local private layer are reserved for each client and are unique to that client; the parameters in the shared layer are shared by all clients.

[0032] S4.6 Personalized Parameter Aggregation Strategy: Based on the local private layer and shared layer parameters defined in S4.5, aggregation is performed only on shared layer parameters. During aggregation, the data volume of each client is considered as a weight. Local private layer parameters are updated independently by each client and do not participate in global aggregation. The formula is as follows:

[0033]

[0034] in, Let represent the weight of the k-th client. W represents the shared layer parameters of the k-th client. s n represents the aggregated global shared layer parameters. k This represents the amount of data from the k-th client, where K is the total number of clients.

[0035] S4.7 Model Update: Update the parameters W after aggregation. s Deploy it to the global model and distribute it to each client;

[0036] In step S6, a global model is trained based on a federated learning framework, and a local model is trained in each oilfield. The trained global model and local model are used to predict gamma logging values ​​on the test set of each oilfield, and model evaluation calculations are performed to obtain the model evaluation index calculation results. This includes the following sub-steps:

[0037] S6.1 The global model is trained based on the federated learning framework. The training ends when the global model loss converges, and the number of times the global model is trained is recorded. Local models are trained separately using the training sets of each oil field. The number of training times of the local models is the same as the number of training times of the federated learning framework models. This is used to conduct a comparative experiment with the global model.

[0038] S6.2. Use the trained global model and local model to predict gamma logging values ​​for each oilfield test set;

[0039] S6.3 Obtain the predicted and actual values ​​of gamma logging values ​​for each oilfield using the global model and local model. Use three complementary evaluation indicators—relative error, accuracy, and mean square error—to evaluate and calculate the predicted and actual gamma logging values.

[0040] In step S7, the calculation results of the evaluation indicators of the global model and the local model are analyzed and compared. The predictive performance of the model is evaluated from multiple perspectives, including error distribution, trend fitting and relative accuracy, and the final model is obtained.

[0041] The advantages of this invention compared to the prior art are mainly reflected in:

[0042] (1) Improve prediction accuracy through cross-oilfield data collaboration: The federated learning framework enables collaborative modeling of multi-oilfield data, breaking through the limitations of training with single-oilfield data, and significantly improving the generalization ability and prediction accuracy of the model. In particular, it shows stronger adaptability in the case of scarce data and large geological differences.

[0043] (2) Protect data privacy and avoid data leakage: adopt federated learning to avoid sharing raw data, ensure the security of oilfield data privacy, and effectively prevent cross-oilfield data leakage;

[0044] (3) Applicable to large-scale multi-oilfield application scenarios: The solution provided by this invention has high scalability and can adapt to large-scale multi-oilfield data collaboration. It is particularly suitable for collaborative prediction needs of different regions and different geological units in the petroleum industry and has important application value.

[0045] In summary, this invention provides an efficient, privacy-preserving, and accurate intelligent prediction method for gamma-ray logging curves by employing a multi-oilfield data collaboration model based on federated learning and a deep learning architecture with a self-attention mechanism. This method effectively alleviates the problem of data scarcity, overcomes the limitations of traditional models, significantly improves prediction accuracy through cross-oilfield data collaboration, and ensures the security of data privacy across oilfields. Through an innovative federated learning framework and parameter aggregation strategy, it offers a new solution for intelligent prediction of gamma-ray logging curves, possessing strong adaptability and broad application potential, particularly suitable for cross-regional, multi-oilfield data sharing and collaboration. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the embodiments will be briefly introduced below.

[0047] Figure 1 This is an overall flowchart of the method of the present invention;

[0048] Figure 2 A schematic diagram of gamma logging curve correction;

[0049] Figure 3 This is a schematic diagram of the structure of a self-attention-based neural network;

[0050] Figure 4 A schematic diagram of the federated learning framework;

[0051] Figure 5 This diagram illustrates the comparison of gamma logging prediction results between the global model and the local model. Detailed Implementation

[0052] Example 1:

[0053] See Figure 1 The present invention provides a method for gamma logging curve correction and prediction based on federal cross-domain collaboration, comprising the following steps:

[0054] S1. Collect historical logging-while-drilling data from different oilfields and the distance between the corresponding logging-while-drilling tool drill bit and sensor, and perform data preprocessing on the historical logging-while-drilling data;

[0055] S2. Based on the preprocessed data in step S1, perform gamma logging curve correction according to the distance between the drill bit and the sensor of the logging-while-drilling tool, and construct training sets and test sets for different oilfields.

[0056] S3. Construct a time-series prediction model for gamma logging curves, which can be used as a tool for implementing federated learning algorithms;

[0057] S4. Based on the model built in step S3, build a federated learning framework, including a global model and a client model, and initialize the model parameter filtering and model parameter aggregation strategies of the federated learning framework.

[0058] S5. Based on the model constructed in step S3, establish local models in each oilfield as comparative experimental models;

[0059] S6. The global model is trained based on the federated learning framework built in step S4, and the local model built in step S3 is trained in each oil field. The trained global model and local model are used to predict gamma logging values ​​for the test set of each oil field, and the model evaluation is calculated to obtain the model evaluation index calculation results.

[0060] S7. Based on the calculation results of the global model and local model evaluation indicators in step S6, analyze and compare them to determine the final prediction model.

[0061] In this embodiment, in step S1, the relevant historical gamma logging curve data, corresponding real-time drilling data, and the distance between the drill bit and sensor of the corresponding logging-while-drilling tool are collected in real time through sensors and monitoring systems at oil well sites in three regions: Bohai Oilfield, Xinjiang Oilfield, and Beihai Oilfield. The historical gamma logging curve data features include gamma measurement depth and gamma logging value. The real-time drilling data features include measurement depth, drill pressure, standpipe pressure, torque, mechanical drilling speed, rotational speed, mud rheological parameters, and hook load. The corresponding real-time drilling data and the distance between the drill bit and sensor of the corresponding logging-while-drilling tool are 7.84 meters. The collected data is then preprocessed as follows:

[0062] S1.1. Perform preliminary cleaning on the collected data to remove outliers exceeding three standard deviations; Table 1 shows a sample of processed real-time drilling data from a certain oilfield, and Table 2 shows a sample of historical gamma-ray logging curves.

[0063] Table 1. Example of Real-Time Drilling Data Samples

[0064]

[0065] Table 2. Examples of Historical Gamma Logging Curve Sample Data

[0066]

[0067] S1.2 After the initial data cleaning, linear interpolation was used to fill in the missing values. For data with negative characteristic gamma logging values, cubic Lagrange interpolation was performed.

[0068] In this embodiment, in step S2, the pre-processed data is used to perform gamma logging curve correction based on the distance between the drill bit and the sensor in the logging-while-drilling tool. Figure 2 The process of creating a diagram illustrating the principle of gamma curve correction, and constructing training and test sets for different oil fields, includes:

[0069] S2.1 Extract historical gamma logging data, corresponding real-time drilling data, and the distance between the drill bit and sensor in the logging-while-drilling tool;

[0070] S2.2 For the gamma measurement depth sampling interval, based on the fixed depth difference between the drill bit and the sensor, all gamma measurement depth values ​​are summed with the fixed depth difference between the drill bit and the sensor to obtain a new depth value. The gamma logging value corresponding to the new depth value is consistent with the original gamma measurement depth, and finally, new historical gamma logging curve data are obtained. The specific formula is as follows:

[0071]

[0072] Among them, BD i The depth value is measured by gamma, and ΔX is the fixed depth difference between the drill bit and the sensor. The new depth value is given, where n is the number of samples.

[0073] Table 3 shows some sample data from a certain oil field after gamma curve correction:

[0074] Table 3. Examples of partial sample data after gamma curve correction.

[0075]

[0076] In this embodiment, step S3 involves constructing a time-series oil production prediction model based on a self-attention mechanism, aiming to capture the dynamic patterns of oil production parameters and other characteristics changing over time. Figure 3 The model structure diagram shows that model construction specifically includes the following sub-steps:

[0077] S3.1 This model uses a self-attention mechanism as its core framework to construct a deep learning model of encoder-decoder based on the self-attention mechanism. It consists of an input sequence encoding layer, a position encoding layer, a softmax layer, a forward propagation layer, a multi-head attention mechanism layer, a masked multi-head attention mechanism layer, and a linear layer.

[0078] S3.2 Set the model dimension to 256, the model layer to 2, the model neuron random dropout rate to 0.1, and the model learning rate to 0.0001;

[0079] S3.3 The model optimizer uses the Adam optimizer for parameter updates; the loss function uses the mean squared error, as shown in the following formula:

[0080]

[0081] Where yi represents the actual oil production value. Here, n represents the predicted oil production value from the model, and n is the number of samples. During the model training and optimization process, the optimizer uses the Adam optimizer to update the parameters.

[0082] In this embodiment, in step S4, the federated learning framework is built, see... Figure 4 It includes a global model and a client model, and defines the model parameter filtering and aggregation strategies for the federated learning framework, specifically including the following:

[0083] S4.1. Based on the number of oil fields, establish three client-side models for three oil field regions, and establish a server-side model containing a global model.

[0084] S4.2 Global Model Parameter Distribution: The global model distributes model parameters to each client, including feature embedding parameters, attention mechanism parameters, and feedforward network parameters, to ensure that each client has a unified initial state;

[0085] S4.3 Client Model Training and Parameter Upload: Each client model uses the training set of each oilfield for model training, and transmits its own model parameters to the server in real time for each training batch.

[0086] S4.4 Parameter Filtering: After receiving the parameters from each client, the server first filters the parameters, using a normalization strategy that does not rely on batch statistics to filter out parameters that skip batch normalization.

[0087] S4.5 Model Parameter Partitioning: Based on the model structure, the model parameters are divided into a local private layer and a shared layer; the parameters in the local private layer are reserved for each client and are unique to that client; the parameters in the shared layer are shared by all clients.

[0088] S4.6 Personalized Parameter Aggregation Strategy: Based on the local private layer and shared layer parameters defined in S4.5, aggregation is performed only on shared layer parameters. During aggregation, the data volume of each client is considered as a weight. Local private layer parameters are updated independently by each client and do not participate in global aggregation. The formula is as follows:

[0089]

[0090] in, Let represent the weight of the k-th client. W represents the shared layer parameters of the k-th client. s n represents the aggregated global shared layer parameters. k This represents the amount of data from the k-th client, where K is the total number of clients, and K = 3.

[0091] S4.7 After parameter aggregation is complete, the aggregated parameter W... s Deploy it into the global model and distribute it to each client.

[0092] In this embodiment, step S6 involves training a global model based on a federated learning framework, training local models in each oilfield, using the trained global and local models to predict gamma logging values ​​for the test sets of each oilfield, and performing model evaluation calculations to obtain the model evaluation index calculation results. This includes the following steps:

[0093] S6.1. The global model is trained based on the federated learning framework until the global model loss converges, at which point the training is considered complete. The number of training iterations of the global model is recorded. The model uses the batch gradient descent method, where the batch size for each prediction is 64. The training is set to 100 rounds, and the learning rate is adjusted by a decay factor of 0.5 every 10 training cycles.

[0094] S6.2 uses three oilfield training sets to train local models separately, namely local model 1, local model 2 and local model 3. The number of training cycles is the same as that of the federated learning framework model, which is 100 training cycles. The batch size for each prediction is 64. The training is set to 100 rounds and the learning rate is adjusted by a decay factor of 0.5 every 10 training cycles. This is used for comparison experiments with the global model.

[0095] S6.3 Test the trained global model and local model in the test sets of three regions respectively. Set the global model to predict the gamma logging value of 1 meter down using historical 3-meter data in the three oilfield regions respectively, and obtain the predicted gamma logging value.

[0096] S6.4 Obtain the gamma logging values ​​and actual values ​​of the global model and local model in each oilfield. Use three complementary evaluation indicators—relative error, accuracy, and mean square error—to evaluate and calculate the predicted and actual gamma logging values. The model evaluation parameter formulas are as follows:

[0097]

[0098] Where yi is the true value. For predicted values, is the mean of the actual values, and n is the total number of samples;

[0099] In this embodiment, step S7 involves analyzing and comparing the calculation results of the global model and the four local models to evaluate the model's predictive performance from multiple perspectives, including error distribution, trend fitting, and relative accuracy, ultimately obtaining the optimal model. The model evaluation results for the three oilfield prediction results are shown in Table 4.

[0100] Table 4. Model evaluation results for each oilfield

[0101]

[0102] Figure 5 This diagram compares the gamma logging prediction results of the global model and the local model on test sets in four regions. The results show that the federated learning-based time-series prediction model for gamma logging curves predicts 10-meter gamma logging results at 30-meter intervals more accurately than the results obtained using the local model. As shown in the table, the predictive performance of the model in this invention surpasses that of the local model across all indicators.

Claims

1. A method for gamma logging curve correction and prediction based on federal cross-domain collaboration, characterized in that, Includes the following steps: S1. Collect historical logging-while-drilling data from different oilfields and the distance between the corresponding logging-while-drilling tool drill bit and sensor, and perform data preprocessing on the historical logging-while-drilling data; S2. Based on the preprocessed data in step S1, perform gamma logging curve correction according to the distance between the drill bit and the sensor of the logging-while-drilling tool, and construct training sets and test sets for different oilfields. S3. Construct a time-series prediction model for gamma logging curves, which can be used as a tool for implementing federated learning algorithms; S4. Based on the model built in step S3, build a federated learning framework, including a global model and a client model, and initialize the model parameter filtering and model parameter aggregation strategies of the federated learning framework. S5. Based on the model constructed in step S3, establish local models in each oilfield as comparative experimental models; S6. The global model is trained based on the federated learning framework built in step S4, and the local model built in step S3 is trained in each oil field. The trained global model and local model are used to predict gamma logging values ​​for the test set of each oil field, and the model evaluation is calculated to obtain the model evaluation index calculation results. S7. Based on the calculation results of the global model and local model evaluation indicators in step S6, analyze and compare them to determine the final prediction model.

2. The method according to claim 1, characterized in that, In step S1, historical logging-while-drilling data from different oilfields are collected, including historical gamma logging curve data and corresponding real-time drilling data. The characteristics of historical gamma logging curve data include gamma measurement depth and gamma logging value. The characteristics of real-time drilling data include measurement depth, drilling pressure, standpipe pressure, torque, mechanical drilling speed, rotational speed, mud rheological parameters, and hook load.

3. The method according to claim 1, characterized in that, In step S1, data preprocessing includes: performing preliminary cleaning on the collected data to remove outlier data exceeding three times the standard deviation; performing linear interpolation on missing data; and performing cubic Lagrange interpolation on data with negative characteristic gamma logging values.

4. The method according to claim 1, characterized in that, In step S2, the pre-processed data is used to perform gamma logging curve correction based on the distance between the drill bit and the sensor in the logging-while-drilling tool, and training and test sets are constructed for different oilfields. The process includes: S2.1 Extract historical gamma logging data, corresponding real-time drilling data, and the distance between the drill bit and sensor in the logging-while-drilling tool; S2.2 For the gamma measurement depth sampling interval, based on the fixed depth difference between the drill bit and the sensor, all gamma measurement depth values ​​are summed with the fixed depth difference between the drill bit and the sensor to obtain a new depth value. The gamma logging value corresponding to the new depth value is consistent with the original gamma measurement depth, and finally, new historical gamma logging curve data are obtained. The specific formula is as follows: Among them, BD i The depth value is measured using gamma, and ΔX is the fixed depth difference between the drill bit and the sensor. The new depth value is given, where n is the number of samples. S2.3 The new historical gamma logging curve data obtained from S2.2 is merged with the real-time drilling data according to depth to finally obtain a new drilling dataset; S2.4 The data collected from each oilfield are corrected for gamma logging curves according to S2.1 to S2.3, and the final dataset is used to construct training and test sets for each oilfield.

5. The method according to claim 1, characterized in that, In step S3, the process of constructing a time-series prediction model for gamma logging curves includes: S3.

1. A deep learning model based on an encoder-decoder is constructed using a self-attention mechanism as the core framework. The model consists of an input sequence encoding layer, a position encoding layer, a softmax layer, a forward propagation layer, a multi-head attention mechanism layer, a masked multi-head attention mechanism layer, and a linear layer. S3.2 Set the model dimension to 256, the model layer to 2, the model neuron random dropout rate to 0.1, and the model learning rate to 0.0001; S3.3 The model optimizer uses the Adam optimizer for parameter updates; the loss function uses the mean squared error, as shown in the following formula: Among them, y i This represents the true gamma logging value. is the gamma logging value predicted by the model, and n is the number of samples.

6. The method according to claim 1, characterized in that, In step S4, a federated learning framework is built, including a global model and a client model, and the model parameter filtering and aggregation strategies of the federated learning framework are initialized, further including: S4.

1. Establish client models for each oil field based on the number of oil fields, and establish a server-side model containing a global model. S4.2 Global Model Parameter Distribution: The global model distributes model parameters to each client, including feature embedding parameters, attention mechanism parameters, and feedforward network parameters, to ensure that each client has a unified initial state; S4.3 Client Model Training and Parameter Upload: Each client model uses the training set of each oilfield for model training, and transmits its own model parameters to the server in real time for each training batch. S4.4 Parameter Filtering: After receiving the parameters from each client, the server first filters the parameters, using a normalization strategy that does not rely on batch statistics to filter out parameters that skip batch normalization. S4.5 Model Parameter Partitioning: Based on the model structure, the model parameters are divided into a local private layer and a shared layer; the parameters in the local private layer are reserved for each client and are unique to that client; the parameters in the shared layer are shared by all clients. S4.6 Personalized Parameter Aggregation Strategy: Based on the local private layer and shared layer parameters defined in S4.5, aggregation is performed only on shared layer parameters. During aggregation, the data volume of each client is considered as a weight. Local private layer parameters are updated independently by each client and do not participate in global aggregation. The formula is as follows: in, Let represent the weight of the k-th client. W represents the shared layer parameters of the k-th client. s n represents the aggregated global shared layer parameters. k This represents the amount of data from the k-th client, where K is the total number of clients. S4.7 Model Update: Update the parameters W after aggregation. s Deploy it into the global model and distribute it to each client.

7. In step S6, the global model is trained based on the federated learning framework, and the local model is trained in each oilfield. The trained global model and the local model are used to predict gamma logging values ​​on the test set of each oilfield, and model evaluation calculations are performed to obtain the model evaluation index calculation results. This includes the following sub-steps: S6.1 The global model is trained based on the federated learning framework. The training ends when the global model loss converges, and the number of times the global model is trained is recorded. Local models are trained separately using the training sets of each oil field. The number of training times of the local models is the same as the number of training times of the federated learning framework models. This is used to conduct a comparative experiment with the global model. S6.

2. Use the trained global model and local model to predict gamma logging values ​​for each oilfield test set; S6.3 Obtain the predicted and actual gamma logging values ​​of the global model and local model in each oilfield, and use three complementary evaluation indicators—relative error, accuracy, and mean square error—to evaluate and calculate the predicted and actual gamma logging values.

8. The method according to claim 1, characterized in that, In step S7, the calculation results of the evaluation indicators of the global model and the local model are analyzed and compared. The predictive performance of the model is evaluated from multiple perspectives, including error distribution, trend fitting and relative accuracy, and the final model is obtained.

Images