Bidirectional recurrent neural network acoustic logging curve reconstruction method

Abstract

The invention provides a bidirectional recurrent neural network acoustic logging curve reconstruction method. The reconstruction method comprises the steps of S1, acquiring data and performing sequence segmentation; s2, carrying out normalization processing on the logging curve; s3, carrying out superposition networking and training on the bidirectional recurrent neural network structure block; s4, model testing and parameter storage; and S5, reading the model parameter file to reload the model, importing the to-be-reconstructed logging data, generating a prediction curve, and storing and exporting a result. A bidirectional recurrent neural network algorithm of artificial intelligence deep learning is adopted, bidirectional depth sequence information contained in a logging curve is effectively captured, the curve reconstruction accuracy is improved, unmeasured, missing and low-quality curves are reconstructed, multi-curve information is fused, and the sensitivity of a reconstructed acoustic curve is improved. The system has the advantages of being low in cost, high in efficiency and high in precision.

Description

Technical Field

[0001] The present invention relates to the field of geophysical logging in the field of oil and gas exploration and development, and in particular to a bidirectional recurrent neural network acoustic logging curve reconstruction method. Background Art

[0002] Due to their strong correlation between well logging and seismic data, acoustic logs are the most frequently used reconstruction targets. By extracting or integrating information reflecting lithologic variations from other well logging curves into the acoustic logs, the sensitivity of the pseudo-acoustic logs to interfaces can be enhanced, thereby improving the accuracy and resolution of seismic inversion results. Furthermore, during actual well logging, factors such as wellbore collapse and instrument failure can cause distortion or loss of logging data, complicating subsequent geological interpretation and logging evaluation. However, due to the economic cost, re-measurement is not feasible. Therefore, various methods can be used to manually reconstruct acoustic logs from existing well logging data. Because these acoustic logs are extracted and integrated from other well logging curves, the pseudo-acoustic logs are more sensitive to lithology and interfaces, further facilitating seismic inversion and refined logging interpretation. Furthermore, deep learning-based model building techniques can establish nonlinear mapping relationships between well logging curves, and the reconstructed logs can also be used for logging quality monitoring and control.

[0003] With the continuous rise of big data, machine learning and deep learning technologies in recent years, this technology has become a current research hotspot in the field of geophysical exploration. Well logging curve reconstruction is a supervised regression task. Compared with traditional curve reconstruction methods, deep learning technology can effectively reduce the tedious feature extraction and summarization process in the early feature engineering. It only requires data cleaning, elimination of abnormal data, design of loss function and training through optimization algorithm to obtain a prediction model. Its model representation ability and accuracy are often higher than traditional methods.

[0004] BP neural networks were first applied to the task of log reconstruction. Wang Yan et al. (2000) used an improved BP algorithm to extract multi-curve morphology and recover the resistivity of thin and poorly distributed layers. Zhang Yongjun et al. (2008) used BP neural networks to construct neutron and density calculation models. Yang Zhili et al. (2008) used BP neural networks to reconstruct acoustic wave curves by integrating natural potential, resistivity, natural gamma ray, and other well logging curves, improving the quality of acoustic wave curves affected by wellbore collapse. However, BP neural networks suffer from shortcomings such as overfitting and a tendency to fall into local minima during optimization. Moreover, BP neural networks only consider the correlation of curves at a single depth and cannot incorporate information about the depth-dependent variations of well logging curves. Therefore, sequence analysis network structures, primarily recurrent neural networks, were introduced to the task of log reconstruction. Zhang Dongxiao et al. (2018) pioneered the use of long short-term memory neural networks (LSTMs) to reconstruct well logging curves. Wang Jun et al. (2020) used gated recurrent units (GRUs) to reconstruct well logging curves, achieving faster training convergence. Data can be constrained using future information. Zhou Xin et al. (2022) used a bidirectional gated recurrent unit neural network to fully utilize the forward and backward data information of well logging data. Summary of the Invention

[0005] In view of the above problems, the present invention is proposed to provide a bidirectional recurrent neural network acoustic logging curve reconstruction method that overcomes the above problems or at least partially solves the above problems.

[0006] According to one aspect of the present invention, a bidirectional recurrent neural network acoustic logging curve reconstruction method is provided, the reconstruction method comprising:

[0007] Step S1: Acquire data and perform sequence segmentation;

[0008] Step S2: normalizing the well logging curve;

[0009] Step S3: Superimpose and train the bidirectional recurrent neural network structure;

[0010] Step S4: model testing and parameter saving;

[0011] Step S5: Read the model parameter file and reload the model, import the well logging data to be reconstructed, generate the prediction curve and save and export the results.

[0012] Optionally, the step S1: acquiring data and performing sequence segmentation specifically includes:

[0013] Data inspection and cleaning: remove abnormal well sections in logging data, ensure that the effective data range and distribution of each logging curve are within a reasonable range, and eliminate system errors caused by different instrument models and environmental factors;

[0014] The overall data is divided into training well data set, verification well data set and test well data set in proportion based on the well number. The data is stored in a single multi-sheet table or three CSV files. The division ratio is determined by the actual number of wells.

[0015] Read each data set and split the data with depth column as index according to sequence design requirements. The data dimension is (data size ,fea size ) is transformed into (new_data size , seq size ,fea size ), requiring that the sequences between wells be discontinuous, the total length of the data will be reduced due to sequence segmentation, and the formula is expressed as:

[0016] new_data size =data size -seq size ×well num

[0017] Where new_data size is the total amount of data after segmentation, data size is the total amount of data before segmentation, seq size is the designed sequence length, well num is the number of wells included in the dataset.

[0018] Optionally, the step S2: normalizing the well logging curve specifically includes:

[0019] Determine the normalization parameters based on the training set;

[0020] The rationality is judged based on the logging curve type and unit, and the parameters are saved in a json file to establish a link with the subsequent training model.

[0021] Optionally, the step S3: superimposing and networking the bidirectional recurrent neural network structure and training specifically includes: the network structure consists of a fully connected neural network and a recurrent neural network model block, the first layer, the third layer, and the fourth layer are fully connected layers, and the second layer is composed of multiple recurrent model blocks stacked together.

[0022] Optionally, the recurrent neural network model block includes: a basic recurrent neural network, a multi-head attention layer, and a forward feedback layer;

[0023] The basic recurrent neural network includes recurrent neural network RNN, long short-term memory neural network LSTM, and gated recurrent neural network GRU;

[0024] The attention mechanism in the multi-head attention layer adopts standardized dot product attention, and uses concat to connect the multi-head results representing different subspaces, and aggregates them into the forward feedback layer through a single linear layer to output the results.

[0025] Optionally, the recurrent neural network model block includes: a dropout layer, which is located after the multi-head attention layer and the forward feedback layer, respectively, and is used to reduce overfitting during training and increase the randomness of the network.

[0026] Optionally, the recurrent neural network model block includes: layer normalization, which is located before the basic recurrent neural network, the multi-head attention layer, and the forward feedback layer, respectively, to accelerate the parameter convergence of the network during the training process.

[0027] Optionally, the step S4: model testing and parameter saving specifically includes:

[0028] The generalization of the network model is verified by the test set. The model parameter file includes the model parameter file and the global training parameter file.

[0029] Optionally, the step S5: reading the model parameter file, reloading the model, importing the well logging data to be reconstructed, generating the prediction curve, and saving and exporting the results specifically includes:

[0030] Read the model parameter file and reload the model;

[0031] The model prediction module pre-processes the logging data to be reconstructed according to the model sequence design requirements to obtain the prediction data;

[0032] The prediction data is exported in the form of a CSV file, including a depth column and a prediction curve column.

[0033] Optionally, the model prediction module can be directly applied multiple times independently based on the existing training model parameter file.

[0034] The present invention provides a bidirectional recurrent neural network acoustic logging curve reconstruction method, which includes: step S1: acquiring data and performing sequence segmentation; step S2: normalizing the logging curve; step S3: superimposing and networking the bidirectional recurrent neural network structure and training it; step S4: model testing and parameter storage; step S5: reading the model parameter file and reloading the model, importing the logging data to be reconstructed, generating a prediction curve, and saving and exporting the results. The bidirectional recurrent neural network algorithm using artificial intelligence deep learning effectively captures the bidirectional depth sequence information contained in the logging curve, improves the curve reconstruction accuracy, reconstructs unmeasured, missing, and low-quality curves, integrates multi-curve information, and enhances the sensitivity of the reconstructed acoustic wave curve. This improves the oilfield logging database and has the advantages of low cost, high efficiency, and high precision.

[0035] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are specifically listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 A flowchart of a bidirectional recurrent neural network acoustic logging curve reconstruction method provided by an embodiment of the present invention;

[0038] Figure 2 A schematic diagram of the neural network model structure provided by an embodiment of the present invention;

[0039] Figure 3 A schematic diagram of the reconstruction results of the acoustic logging curve of the verification well A provided in an embodiment of the present invention;

[0040] Figure 4 Schematic diagram of the reconstruction results of the acoustic logging curve of verification well B provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] Exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0042] The terms "comprises" and "comprising" and any variations thereof in the description, embodiments, claims and drawings of the present invention are intended to cover non-exclusive inclusions, for example, including a series of steps or units.

[0043] The technical solution of the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0044] like Figure 1 As shown, a bidirectional recurrent neural network acoustic logging curve reconstruction method includes the following steps:

[0045] Step (1) Data preparation and sequence segmentation: Check and clean the logging data and remove abnormal well sections. Divide the entire data set into training set, validation set and test set, determine the sequence length requirement, and segment into sequence table data;

[0046] Step (2), normalization of logging curves: Different logging curves have different units and value ranges. Before training the neural network, the data needs to be normalized and the maximum and minimum values ​​of each curve need to be recorded.

[0047] Step (3), model networking and training: Apply the multi-head attention recurrent neural network building block to overlay the network, determine the number of network layers, the number of attention heads, the recurrent neural network hidden layer and other parameters to establish the network model. Use the curves other than the target curve in the training set as input, and the target logging curve as output. During training, use the validation set as the model network parameter selection standard, and use the mean square error as the loss function to train the network until the model converges and meets the accuracy requirements;

[0048] Step (4), model testing and parameter saving: Use the test set to verify the generalization performance of the model and save the model parameter file.

[0049] Step (5), model reloading and prediction: read the model parameter file to reload the model, import the well logging data to be reconstructed, generate the prediction curve and save and export the results.

[0050] The specific implementation method of data inspection and cleaning in step (1) is to remove abnormal well sections in the logging data through identification and mapping observation by the logging software program, ensure that the effective data value range and distribution of each logging curve are within a reasonable range, and eliminate systematic errors caused by different instrument models and environmental factors.

[0051] The dataset division method in step (1) is to divide the overall data into training well dataset, validation well dataset and test well dataset in proportion based on well number, and store them in a single multi-sheet table or three CSV formats, named "train", "vali" and "test" respectively, with a division ratio of 8:1:1 or adjusted according to the actual number of wells;

[0052] In step (1), the data set sequence is divided into two parts: the depth column is used as the index to divide the data, and the data dimension is (data size ,fea size ) is transformed into (new_data size ,seq size ,fea size ), after segmentation, the well sequences are discontinuous, and the total length of the data is reduced due to sequence segmentation, which can be expressed as:

[0053] new_data size =datasize -seq size ×well num

[0054] Where: new_data size is the total amount of data after segmentation, data size is the total amount of data before segmentation, seq size is the designed sequence length, well num is the number of wells included in the dataset.

[0055] In step (2), the normalization parameters are determined by the training set, and their rationality is judged according to the logging curve type and unit. The parameters will be saved in the json file and linked to the subsequent training model. The normalization algorithm uses z-score normalization, and its formula is as follows:

[0056]

[0057] Where, v i is the actual measured value of the logging curve, v max and v min is the maximum and minimum value of the logging curve, V i is the normalized logging curve value;

[0058] In step (3), the network structure consists of a fully connected neural network and a recurrent neural network model block. The first, third, and fourth layers are fully connected layers, and the second layer is composed of multiple recurrent model blocks stacked together. The model structure is as follows: Figure 2 shown.

[0059] Each recurrent network block consists of a basic recurrent neural network, a multi-head attention layer, and a feedforward layer. Basic recurrent neural networks include, but are not limited to, recurrent neural networks (RNNs), long short-term memory (LSTMs), and gated recurrent neural networks (GRUs). The multi-head attention layer uses a scaled dot-product attention mechanism, concatenating multiple results representing different subspaces using a concat method. These results are fed into a single linear layer and fed into the feedforward layer for output.

[0060] Each recurrent network block contains two dropout layers, one after the multi-head attention layer and the other after the forward feedback layer, to reduce overfitting during training and increase the randomness of the network.

[0061] Each recurrent network block contains three layer normalizations (layernorm), which are located before the basic recurrent neural network, the multi-head attention layer, and the forward feedback layer, respectively, to accelerate the convergence of the network during training.

[0062] In step (4), the model parameter file includes a model parameter file (.pth) and a global training parameter file (.json, stored in a dictionary key-value pair format).

[0063] In step (5), the prediction data is exported in the form of a CSV file, including a depth column and a prediction curve column.

[0064] Example 1

[0065] (1) This example is implemented in a certain work area of ​​Shengli Oilfield. A total of 100 wells are used as a data set, which are divided into training wells, verification wells, and test wells in a ratio of 8:1:1. Three curves, namely natural gamma, density, and neutron, are used as input, and a single acoustic logging curve is used as output. In the network model, the number of network layers is set to 4, the number of attention heads is set to 4, and the number of nodes in the recurrent neural network hidden layer is set to 128. The Adam optimizer is used for optimization with a learning rate of 0.0001. The training ends when the loss of the verification set does not decrease for 20 consecutive rounds.

[0066] (2) The method proposed in this invention is used to train the model parameters and perform prediction reconstruction. The final reconstruction result is as follows: Figure 3 The first track in the figure is the depth track, the second track is the mud logging lithology track, the third track is the wellbore CAL and natural gamma ray GR curves, the fourth track is the neutron CNL and density DEN curves, the fifth track is the interpretation conclusion track; and the sixth track is the comparison track between the measured acoustic wave curve and the reconstructed acoustic wave curve.

[0067] (3) Well A in the validation set: the overall predicted length of the well is 1258-3338 m, with a mean absolute error (MAE) of 19.1 and R 2 is 0.83, and the Pearson correlation coefficient is 0.91. Figure 3 Taking the water layer at 2240-2290m ​​in the well section as an example, some small sections such as 2347m are caused by measurement errors. Overall, the acoustic wave reconstruction curve has a good matching effect with the measured curve.

[0068] (4) Test set B well, the overall predicted length of the well is 902-3157m, the mean absolute error MAE is 30.2, R 2 is 0.79, and the Pearson correlation coefficient is 0.89. Figure 4Taking the oil-water layer and oil-water layer at 2930-2970m in the well section as an example, the reservoir section has a good wellbore, and the reconstructed curve basically coincides with the true curve. The upper 2930-2935m mudstone section has a partial diameter expansion, increased gamma, decreased density, and increased neutrons. The reconstructed acoustic wave is larger than the true measurement value, reflecting the increase in apparent porosity caused by the diameter expansion. The original acoustic wave does not show this feature, showing the information gain advantage of the reconstructed curve compared to the original curve. Overall, the acoustic wave reconstructed curve has a good matching effect with the measured curve.

[0069] Beneficial Effects: The bidirectional recurrent neural network algorithm, which uses artificial intelligence deep learning, can effectively capture the bidirectional depth sequence information contained in the logging curve, improve the accuracy of curve reconstruction, reconstruct unmeasured, missing, and low-quality curves, integrate multi-curve information, and enhance the sensitivity of reconstructed acoustic wave curves. This method can also improve the oilfield logging database, with the advantages of low cost, high efficiency, and high precision.

[0070] The multi-head attention mechanism in the Transformer model was introduced to further enhance the model's ability to represent the nonlinear relationship between logging curves. Optimization methods such as layer normalization were used to improve the model's convergence efficiency, reduce training costs, and increase the model's prediction accuracy.

[0071] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A bidirectional recurrent neural network acoustic logging curve reconstruction method, characterized in that: The reconstruction method includes: Step S1: Acquire data and perform sequence segmentation; Step S2: normalizing the well logging curve; Step S3: Superimpose and train the bidirectional recurrent neural network structure; Step S4: model testing and parameter saving; Step S5: Read the model parameter file and reload the model, import the well logging data to be reconstructed, generate the prediction curve and save and export the results.

2. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 1, wherein: The step S1: acquiring data and performing sequence segmentation specifically includes: Data inspection and cleaning: remove abnormal well sections in logging data, ensure that the effective data range and distribution of each logging curve are within a reasonable range, and eliminate system errors caused by different instrument models and environmental factors; The overall data is divided into training well data set, verification well data set and test well data set in proportion based on the well number. The data is stored in a single multi-sheet table or three CSV files. The division ratio is determined by the actual number of wells. Read each data set and split the data with depth column as index according to sequence design requirements. The data dimension is (data size ,fea size ) is transformed into (new_data size , seq size ,fea size ), requiring that the sequences between wells be discontinuous, the total length of the data will be reduced due to sequence segmentation, and the formula is expressed as: new_data size =data size -seq size ×well num Where new_data size is the total amount of data after segmentation, data size is the total amount of data before segmentation, seq size is the designed sequence length, well num is the number of wells included in the dataset.

3. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 1, wherein: The step S2: normalizing the well logging curve specifically includes: Determine the normalization parameters based on the training set; The rationality is judged based on the logging curve type and unit, and the parameters are saved in a json file to establish a link with the subsequent training model.

4. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 1, wherein: The step S3: superimposing and networking the bidirectional recurrent neural network structure and training specifically includes: the network structure is composed of a fully connected neural network and a recurrent neural network model block, the first layer, the third layer, and the fourth layer are fully connected layers, and the second layer is composed of multiple recurrent model blocks stacked together.

5. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 4, wherein: The recurrent neural network model block includes: a basic recurrent neural network, a multi-head attention layer, and a forward feedback layer; The basic recurrent neural network includes recurrent neural network RNN, long short-term memory neural network LSTM, and gated recurrent neural network GRU; The attention mechanism in the multi-head attention layer adopts standardized dot product attention, and uses concat to connect the multi-head results representing different subspaces, and aggregates them into the forward feedback layer through a single linear layer to output the results.

6. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 4, wherein: The recurrent neural network model block includes: a dropout layer, which is located after the multi-head attention layer and the forward feedback layer, respectively, and is used to reduce overfitting during training and increase the randomness of the network.

7. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 4, wherein: The recurrent neural network model block includes: layer normalization, which is located before the basic recurrent neural network, the multi-head attention layer, and the forward feedback layer, respectively, to accelerate the parameter convergence of the network during the training process.

8. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 1, wherein: The step S4: model testing and parameter saving specifically includes: The generalization of the network model is verified by the test set. The model parameter file includes the model parameter file and the global training parameter file.

9. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 1, wherein: The step S5: reading the model parameter file and reloading the model, importing the well logging data to be reconstructed, generating the prediction curve and saving and exporting the results specifically includes: Read the model parameter file and reload the model; The model prediction module pre-processes the logging data to be reconstructed according to the model sequence design requirements to obtain the prediction data; The prediction data is exported in the form of a CSV file, including a depth column and a prediction curve column.

10. The method for reconstructing acoustic logging curves using a bidirectional recurrent neural network according to claim 9, wherein: The model prediction module can be directly applied multiple times on its own based on the existing training model parameter file.

Images