An interpretable well logging curve completion method

Abstract

The present invention provides an interpretable well logging curve completion method. Based on sample well logging, a sample data set containing complete well logging curves that have been interpreted and a sample data set containing missing well logging curves to be predicted are constructed, denoted as the source domain and the target domain respectively. The well logging curve sample data in the source domain and the target domain are used as input parameters, and a dual-stream neural network model is used for regression prediction training. The complete well logging curve of the target well is input into the trained dual-stream neural network model to complete the missing well logging curve of the target well. The present invention uses a dual-stream neural network model for regression prediction training. When training, a fully convolutional neural network is introduced to effectively capture and fuse features of different well logging scales, and a random Fourier domain difference loss and a domain unidirectional alignment loss are introduced to improve the calculation efficiency and the interpretability of well logging features.

Description

Technical Field

[0001] The present invention relates to the technical field of geophysical logging, and particularly relates to an interpretable logging curve completion method. Background Technique

[0002] Geophysical logging is a method for measuring geophysical parameters by using the electrochemical properties, conductive properties, acoustic properties, radioactivity and other geophysical properties of rock formations; at present, with the in-depth development of geological survey work in China, geophysical logging methods, as one of the physical measurement means used to describe and analyze underground conditions, have been widely used in geological mapping, hydrogeology, engineering geology and environmental geology exploration, terrestrial gas hydrate geological survey, scientific drilling and other work. Geophysical logging methods can effectively reduce the workload of drilling and coring, improve the exploration speed, and reduce the exploration cost, which is of great significance for oil and gas exploration and development.

[0003] In specific use, geologists and engineers can establish an accurate geological model based on logging data and design exploration and development strategies. However, the acquisition of logging curves is often expensive and time-consuming. In actual measurement, due to various objective reasons, the problem of missing logging data often occurs, and it may also be abandoned to measure some entire logging curves for cost considerations. Therefore, the completion and generation of logging curves is a research with academic and engineering value; however, due to the complex formation conditions and anisotropy, the mapping relationship between different logging curves is extremely complex. Whether it is a traditional physical model or an empirical model, it is difficult to accurately describe the relationship between logging curves and cannot complete the incomplete logging curves.

[0004] Logging data processing and comprehensive interpretation use the in-situ formation acoustic, electrical, nuclear and other geophysical data collected to conduct comprehensive analysis to solve problems such as formation division, formation parameter calculation, oil and gas reservoir evaluation, and other geological and engineering technical problems in exploration and development. Traditional logging interpretation uses artificial means such as experiments, theories and statistics to establish an interpretation model. In the process of data processing, the existing methods require artificial selection of appropriate models and parameters according to different formations where the data is located and different types of data obtained, and then conduct data processing and comprehensive interpretation. The steps are very cumbersome and there are also deviations in the understanding of the same data by different interpreters. Therefore, the final logging interpretation results are easily affected by the subjective factors of the interpreters. Summary of the Invention

[0005] The purpose of the present invention is to provide an interpretable logging curve completion method to solve the problems raised in the above background technique.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An interpretable logging curve completion method, characterized in that it specifically includes the following steps:

[0008] S100. Based on the sample well logs, construct a sample data set containing the fully interpreted well log curves and a sample data set containing the missing well log curves to be predicted, denoted as the source domain and the target domain respectively;

[0009] S200. Use the well log curve sample data in the source domain and the target domain as inputs, and perform regression prediction training using a two-stream neural network model;

[0010] S201. Extract multi-scale features from the well log curve samples through a fully convolutional neural network, and perform feature fusion through the gated multi-scale feature fusion method (GMF);

[0011] S202. Maintain the consistency of the feature representation ability through a fully connected layer, measure the distribution difference of the data based on the maximum mean discrepancy (MMD) criterion, and introduce random Fourier features to construct a domain difference loss;

[0012] S203. Minimize the difference between the reconstructed well log curve samples in the target domain, make the source domain align unidirectionally with the target domain, and achieve feature migration;

[0013] S204. Perform a regression prediction function on the aligned features to reconstruct the missing well log curves;

[0014] S300. Input the complete well log curves of the target domain into the trained two-stream neural network model, and the missing well log curves of the target well log can be completed.

[0015] Furthermore, the regression model training is implemented based on the LogRegX network. The LogRegX network adopts a symmetric two-stream structure of a top stream and a bottom stream, including but not limited to an FCN-GMF module, an RFF-MMD module, a unidirectional alignment module, and a regression module.

[0016] Furthermore, the sample data sets of the well log curve samples in the source domain and the target domain are specifically constructed through the following method:

[0017] S101. For the sample well logs of d complete well log curves given, the following data sets can be constructed:

[0018]

[0019] where: x i is the sample data set of the well log feature vectors corresponding to each complete well log curve, d is the number of complete well log curves, represents the set of real numbers, X is the sample point data set of the well log, Y is the sample point data set of the partially or entirely missing well log curves to be predicted, and n is the number of sample points of the well log;

[0020] S102. Construct the log curve sample data sets for the source domain and the target domain:

[0021]

[0022] Where: D s is the log curve sample data set for the source domain, and D t is the log curve sample data set for the target domain. represents the true value of the complete log curve at the i-th sample point, represents the true value of the log curve to be predicted at the i-th sample point, and n s and n t are the sample numbers of the data sets D s and D t respectively.

[0023] Furthermore, the multi-scale feature extraction and fusion of the log curve samples are specifically implemented by the following method: In the training stage, input the log curve sample data of the source domain and the target domain into the top stream and the bottom stream of the LogRegX network, and based on the fully convolutional network (FCN), convert the number of feature channels and the size of the feature map of the input log curve through multi-layer convolution and downsampling to extract the multi-scale features of the log curve samples; and use the multi-scale feature fusion (GMF) method to fuse the log feature maps of different scales to capture the response characteristics of the log curve.

[0024] Furthermore, the step S202 is implemented by the following method based on the RFF-MMD module: Map the feature vectors of all sample points to the reproducing kernel Hilbert space or other high-dimensional subspaces with similar characteristics, correct them based on the two-stream structure of the RegLogX network sharing the same parameters, and measure the corrected log curves based on the maximum mean discrepancy criterion, and at the same time introduce the random Fourier feature mapping function to approximate the high-dimensional subspace.

[0025] Furthermore, the step S203 is implemented by the following method based on the one-way alignment module: Correct the log curve of the source domain with the log curve of the target domain as the baseline in the high-dimensional subspace to make the probability distribution of the log curve of the source domain approach that of the log curve of the target domain, realize the one-way alignment from the source domain to the target domain, and at the same time introduce the target domain information retention constraint mechanism to reduce the data distribution difference between the source domain and the target domain and make the model interpretable.

[0026] Furthermore, the regression prediction is to map the extracted multi-scale features to the sample label space based on a regression module, create a regression layer including a fully connected layer with two hidden layers and a prediction result of the missing or lost logging curve in the target domain as the last layer, and take the mean square error as the metric criterion during training to measure the difference between the predicted value and the true value of the logging curve in the target domain, and accumulate the mean square error values of each sample to minimize the regression loss.

[0027] As can be seen from the above technical solutions, the present invention uses a two-stream neural network model for regression prediction training. When training, a fully convolutional neural network is introduced to effectively capture and fuse features of different logging scales, and a random Fourier domain difference loss and a domain unidirectional alignment loss are introduced to improve the computational efficiency and the interpretability of logging features. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic flow chart of the steps of the logging curve completion method of the present invention;

[0029] Figure 2 It is a structural block diagram of the LogRegX network;

[0030] Figure 3 It is a basic principle diagram of the gated multi-scale feature fusion method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0031] The following describes in detail a preferred embodiment of the present invention with reference to the accompanying drawings.

[0032] As Figure 1 shown, the interpretable logging curve completion method is implemented based on the LogRegX network, and specifically includes the following steps:

[0033] S100. Based on the sample logging, construct a sample data set including the interpreted complete logging curve and a sample data set including the missing logging curve to be predicted, which are respectively denoted as the source domain and the target domain;

[0034] S200. Use the logging curve sample data in the source domain and the target domain as input parameters, and perform regression prediction training using a two-stream neural network model;

[0035] S300. Input the complete logging curve of the target logging into the trained two-stream neural network model to complete the missing logging curve of the target logging.

[0036] As Figure 2As shown in the figure, the LogRegX network described in this preferred embodiment adopts a symmetric dual-stream structure of a top stream and a bottom stream, and includes an FCN-GMF module, an RFF-MMD module, a unidirectional alignment (UA) module, and a regression module.

[0037] For the above-mentioned step S1, in specific operations, the log curve sample data sets of the source domain and the target domain are specifically constructed by the following method:

[0038] S11. For a given target log with several complete log curves, construct the following data set:

[0039]

[0040] Where: x i is composed of the values of d complete log curves at depth i, represents the set of real numbers, X is the sample point data set of the log, Y is the sample point data set of the partially or entirely missing log curve to be predicted, and n is the number of log sample points;

[0041] Since the log curves record different geophysical properties of the same formation, there is a potential correlation between the log curves. Based on this correlation, the goal of log curve reconstruction and completion is to learn a regression mapping model from the complete log curve to the missing log curve, that is, to learn f: X → Y;

[0042] S12. Construct the source domain data set and the target domain data set:

[0043]

[0044] Where: D s is the source domain data set, D t is the target domain data set, respectively represent the complete log curves of the source domain sample at depth i and the target domain sample at depth j, represents the true value of the log curve to be predicted at the sample point of depth i, n s , n t are respectively the number of sample points in the data sets D s and D t ; the samples in the source domain and the target domain have the same input space can be expressed as The source domain and the target domain are related, but sampled from different distributions, so there is a probability distribution difference.

[0045] Furthermore, the regression prediction training specifically includes the following steps:

[0046] S201. Extract multi-scale features from well logging curve samples through a fully convolutional neural network (FCN) and perform feature fusion through a gated multi-scale feature fusion (GMF) method;

[0047] In this step, the FCN-GMF module is based on the current popular FCN architecture, aiming to generate multi-channel feature maps through multi-layer convolution and downsampling, perform scale transformation, and restore to the size of the original input well logging curve by the deconvolution layer. In specific use, the feature extraction network consists of three convolutional blocks, each corresponding to a convolutional kernel. In this preferred embodiment, the sizes of the three convolutional kernels are 7×1, 5×1, and 3×1 respectively. After each convolution, linear unit correction activation (ReLU) and MaxPooling pooling processing are required.

[0048] Taking the top flow as an example, the processing flow of its three convolutional layers can be expressed as:

[0049]

[0050] In the formula: E s Corresponds to the output feature of the convolutional layer, w i , b i Correspond to the i-th weight matrix and bias of the convolutional layer respectively, and g(x) = max{0, x} is the ReLU activation function.

[0051] After the above processing, the three convolutional blocks sequentially output three abstract feature maps with inconsistent semantics and scales Compare the receptive field sizes of the three: Existing geophysical research shows that the response form of well logging curves is reflected at different scales. For example, for spontaneous potential logging, small-scale curve mutations reflect local sandstone formations, while large-scale curve trends can reflect sedimentary cycle characteristics; inspired by these prior knowledge, in order to more effectively fuse well logging feature maps at different scales to capture the response characteristics of well logging curves, we consider effectively fusing well logging feature maps at different scales to obtain well logging response features. Specifically, a gated multi-scale feature fusion (GMF) method is proposed. This method adopts a gated mechanism and uses the fused feature map as a gated unit to selectively transmit or filter the information contained in the feature map. As Figure 3 shown, the basic process is as follows: First, decode the feature maps and at different scales, and restore them to the input size through linear interpolation upsampling to obtain and Subsequently, and Merge and transform it to between 0 and 1 through the sigmoid function to control the information transmission. This step can be formalized as:

[0052]

[0053] In the formula: is the Kronecker product; and the fusion weight is a real number between 0 and 1, enabling the network to and perform a weighted average operation, defined using the sigmoid activation function:

[0054]

[0055]

[0056] Based on another GMF network, fuse and to obtain the fully fused feature D s , and a fully connected layer is connected after the feature fusion layer for dimensionality reduction to assist in aligning the logging curves of the source domain and the target domain. The feature mapping process of the target domain undergoes the same processing, and the source domain and the target domain feature extraction networks share parameters such as weights to maintain the consistency of feature representation.

[0057] S202. Maintain the consistency of feature representation ability through a fully connected layer, measure the distribution difference of data based on the maximum mean discrepancy (MMD) criterion, and introduce random Fourier features to construct a domain difference loss;

[0058] In this step, due to factors such as sedimentary environment, borehole conditions, and logging equipment, there are drifts in the distribution of logging data on different wells. If the model trained on the source domain is directly applied to the regression prediction of the target domain, the accuracy usually decreases. To address this problem, in this preferred embodiment, the logging curves of the source domain and the target domain are mapped into a high-dimensional subspace, and the data distribution difference between the source domain and the target domain is quantified based on the maximum mean discrepancy (MMD) criterion. The central idea of MMD is to map the feature vectors of all sample points into a reproducing kernel Hilbert space (RKHS), and use the reproducibility of RKHS to solve the non-linear regression problem that cannot be solved in the original space. Based on this idea, the RegLogX network adopts a two-stream structure and shares the same parameters such as weights, biases, and learning rates to ensure that the calibration process is exactly the same, and finally measures the calibrated logging curves based on the MMD criterion. Specifically, the MMD metric loss of the source domain and the target domain is defined as:

[0059]

[0060] In the formula: represents the operation in the RKHS, is the feature mapping function, which is mapped to the RKHS space for operation, and correspond to and the feature maps processed by the FCN - GMF module respectively.

[0061] Furthermore, by using the kernel function to replace the feature mapping function, the above formula is expanded and deduced as:

[0062]

[0063] In the formula: is the kernel function that satisfies translational invariance, h i and h j refer to the feature maps output by the FCN - GMF module respectively, is the kernel matrix, which can be written as:

[0064]

[0065] where: For each element ss of K each element st of K each element ts of K each element tt of K In addition, each element L ij of L is:

[0066]

[0067] In a specific well - logging curve reconstruction task, the number of well - logging sample points on a well can sometimes even reach the order of magnitude of 10 4 This also makes the dimension of the kernel matrix K extremely high. If traditional kernel methods are used, it will lead to huge computational costs and memory consumption; and in this example, the feature mapping function tends to be infinitely dimensional and is an implicit mapping function. Most existing methods reduce the computational amount by setting a small batch size; however, a small batch size is prone to having a biased domain distribution estimate. To solve the above problems, the present invention introduces the random Fourier feature (RFF) mapping. By defining an explicit non - linear mapping function the mapped space is made to approximate the RKHS probabilistically, enabling it to maintain the characteristics of the feature mapping function This function maps the original data to a low - dimensional Euclidean space and estimates the infinite - dimensional kernel representation through finite - dimensional vector operations. The kernel function estimation is achieved through the inner product of two mapping vectors, in the following form:

[0068] k(h i ,h j )≈<φ(h i ),φ(h j )>

[0069] where: is a random mapping feature vector, l is the pre-mapping dimension, usually l << n t +n s . By combining the Fourier transform and Bochner's theorem, and using empirical formulas to replace the complex part, the explicit low-dimensional mapping function φ(·) of the random Fourier features (RFF) can be derived:

[0070]

[0071] In the formula: are all parameters in the mapping process, i = 1, 2,..., l, σ 2 is a predefined variance, and are the normal distribution and the joint distribution respectively, and h refers to the feature vector. For the convenience of writing, define and Then the loss function of the RFF-MMD module can be written as:

[0072]

[0073] In the formula: and are all all-ones vectors. The kernel function can be estimated by setting the value of the parameter l, which can significantly reduce the dimension of the kernel matrix in the mapping process and effectively improve the calculation efficiency; in addition, the explicit mapping function φ(·) can construct a mapping space that approximates the reproducing kernel Hilbert space (RKHS) to the greatest extent, ensuring the accuracy of the maximum mean discrepancy (MMD) metric.

[0074] S203. Minimize the difference between the target domain log curve samples before and after reconstruction, and achieve feature migration by unidirectionally aligning the source domain to the target domain. In the above step S202, RFF generates the representation of the log curve through random non-linear mapping, which causes the features after mapping to lose their original physical meaning. Most existing domain alignment methods adopt bidirectional alignment, mapping the original features to a high-dimensional subspace through non-linear mapping and aligning the features after mapping. However, this often results in the intermediate features having no clear physical meaning. To improve the interpretability of the model, we propose the idea of aligning the source domain to the target domain, that is, unidirectional alignment (UA). Such a unidirectional alignment strategy not only reduces the hypothesis space but also facilitates optimization. Its practical significance is to draw the log curves of the source wells into a probability distribution close to that of the target well log curves (as the baseline curves), so that different wells have similar log responses. Figure 2 Shows the main information flow through FCN-GMF and FNN (fully connected neural network), briefly recorded as follows:

[0075]

[0076] Among them, X t is the target domain log curve matrix, D t is the target domain intermediate feature, is the log curve matrix of the target domain after correction. The source domain log curve matrix X s can obtain the log curve matrix of the source domain after correction through the same mapping process By minimizing the RFF-MMD loss, it can make and have similar data distribution laws. In addition, introduce the Target Information Retaining and Constraining (TRC) mechanism to keep the main geophysical information of the target domain log curves unchanged before and after correction. Formally, the goal of the TRC mechanism is achieved by minimizing the difference between the log curves of the target domain after transformation and the original log curves, that is, the log curve reconstruction error, namely:

[0077]

[0078] By minimizing In the target domain, use FCN-GMF as the encoder and FNN as the decoder, so that Therefore, the function of the UA module can be described as: the target domain hardly changes, while the source domain changes greatly to be consistent with the target domain without losing the original geophysical meaning.

[0079] S204. Perform regression prediction on the aligned features to reconstruct the missing log curves;

[0080] Specifically, the extracted distributed features need to be mapped to the sample label space to complete regression prediction. The core operation is to create a regression layer, which consists of fully connected layers of two hidden layers. The last layer is the prediction result of the missing or lost logging curves in the target domain. During the training process, the Mean Square Error (MSE) is used as the metric criterion to measure the difference between the predicted value and the true value of the logging curves in the target domain, and the MSE values of each sample are accumulated to minimize the regression loss. As described above, the regression loss is defined as:

[0081]

[0082] In the formula: are the true value and the predicted value of the logging curve to be predicted, respectively.

[0083] In summary, by minimizing the linear combination of the regression loss, the one-way alignment loss, and the MMD loss, the regression accuracy of the model for the target domain well data can be ensured, while reducing the data distribution difference between the two wells and making the model interpretable. According to the above description, the network loss function is defined as:

[0084]

[0085] In the formula: γ and λ are both penalty factors, which play a role in maintaining the balance of each item.

[0086] After the training is completed, only using the top flow and inputting the complete logging curves of the samples into it, the missing or lost logging curves in the samples can be regressively predicted.

[0087] The above-described embodiments are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An interpretable well logging curve completion method, characterized in that Specifically, it includes the following steps: S100. Based on sample well logging, construct a sample data set containing the interpreted complete well logging curves and a sample data set containing the missing well logging curves to be predicted, which are respectively denoted as the source domain and the target domain; S200. Use the well logging curve sample data in the source domain and the target domain as input parameters, and perform regression prediction training using a two-stream neural network model; S201. Extract multi-scale features from the well logging curve samples through a fully convolutional neural network and perform feature fusion through a gated multi-scale feature fusion method; S202. Maintain the consistency of the feature representation ability through a fully connected layer, measure the distribution difference of the data based on the maximum mean difference criterion, and introduce random Fourier features to construct a domain difference loss; S203. Minimize the difference between the target domain well logging curve samples before and after reconstruction, and achieve feature migration by unidirectionally aligning the source domain to the target domain; S204. Perform regression prediction on the aligned features to reconstruct the missing well logging curves; S300. Input the complete well logging curve of the target well into the trained two-stream neural network model to complete the missing well logging curve of the target well.

2. The interpretable well logging curve completion method according to claim 1, wherein The regression model training is implemented based on the LogRegX network. The LogRegX network adopts a symmetric two-stream structure of a top stream and a bottom stream, including but not limited to an FCN-GMF module, an RFF-MMD module, a unidirectional alignment module, and a regression module.

3. The interpretable log curve completion method according to claim 1, characterized in that The sample data sets of the well logging curve in the source domain and the target domain are specifically constructed by the following method: S101. For the sample well logging with a given number of complete well logging curves, construct the following data sets: where: x i is the sample data set of the logging feature vectors corresponding to each complete logging curve, d is the number of complete logging curves, represents the set of real numbers, X is the sample point data set of the logging, Y is the sample point data set of the partially or entirely missing logging curve to be predicted, and n is the number of sample points of the logging; S102. Construct the sample data sets of the well logging curves in the source domain and the target domain: Wherein: D s is the set of logging curve sample data in the source domain, D t is the set of logging curve sample data in the target domain, represents the true value of the complete logging curve at the i-th sample point, represents the true value of the logging curve to be predicted at the i-th sample point, n s and n t are respectively the number of sample points in the data sets D s and D t respectively.

4. The interpretable well log curve completion method according to claim 2, wherein The extraction and fusion of the multi-scale features of the well logging curve samples are specifically implemented by the following method: In the training stage, based on the FCN-GMF module, input the well logging curve sample data in the source domain and the target domain into the top stream and the bottom stream of the LogRegX network, and based on the fully convolutional network, convert the number of feature channels and the size of the feature map of the input well logging curve through multi-layer convolution and downsampling to achieve the extraction of the multi-scale features of the well logging curve samples; and use the multi-scale feature fusion method to fuse the well logging feature maps of different scales to capture the response characteristics of the well logging curve.

5. The interpretable log curve completion method according to claim 2, wherein The step S202 is implemented based on the RFF-MMD module by the following method: Map the feature vectors of all sample points to a reproducing kernel Hilbert space or other high-dimensional subspaces with similar characteristics, correct them based on the two-stream structure of the RegLogX network sharing the same parameters, and measure the corrected well logging curves based on the maximum mean difference criterion, and at the same time introduce a random Fourier feature mapping function to approximate the high-dimensional subspace.

6. The interpretable well log curve completion method according to claim 5, wherein The step S203 is implemented based on the unidirectional alignment module by the following method: Correct the source domain well logging curve with the target domain well logging curve as the baseline in the high-dimensional subspace, make the probability distribution of the source domain well logging curve approach that of the target domain well logging curve, achieve unidirectional alignment of the source domain to the target domain, and at the same time introduce a target domain information retention constraint mechanism to reduce the data distribution difference between the source domain and the target domain and make the model interpretable.

7. The interpretable log curve completion method according to claim 2, wherein The regression prediction is based on the regression module to map the extracted multi-scale features to the sample label space, create a regression layer including a fully connected layer with two hidden layers, and the prediction result of the missing or lost logging curve in the target domain in the last layer, and take the mean square error as the metric criterion during the training process to measure the difference between the predicted value and the true value of the logging curve in the target domain, and accumulate the mean square error values of each sample to minimize the regression loss.

Images