Well logging curve completion method and device

Abstract

The invention discloses a well logging curve completion method and device, which can be used in the technical field of well logging information processing, and the method comprises the steps: inputting a to-be-completed well logging curve into a well logging curve completion network, determining a curve set of the well logging curve and a missing code corresponding to the curve set according to the input to-be-completed well logging curve, the missing codes are converted into a weight feature sequence; calculating the dynamic weight of each channel according to the weight feature sequence, determining the dynamic feature of the target dimension of each channel according to the curve set and the dynamic weight of each channel, and combining the dynamic features of the target dimension of each channel to obtain the dynamic combination feature of the curve set; the dynamic combination features of the curve set are corrected, missing curve estimation is carried out according to the corrected dynamic combination features, and a well logging curve after completion is output. According to the method, missing curves can be accurately and rapidly complemented, and the consistency and completeness of logging information are improved.

Description

Technical Field

[0001] This invention relates to the field of well logging data processing technology, and in particular to a method and apparatus for well logging curve completion. Background Technology

[0002] This section is intended to provide background or context for embodiments of the present invention. The description herein is not intended to imply that it is prior art simply because it is included in this section.

[0003] Well logging data is fundamental for detailed reservoir characterization and reservoir modeling. Ideally, complete well logging data should provide a series of conventional logging curves, including wellbore diameter, spontaneous potential, spontaneous gamma, bulk density, sonic transit time, compensated neutron, and microresistivity, to comprehensively characterize the physical properties of the subsurface formation. However, due to cost-saving considerations, operational limitations, differences in instrument combinations, and instrument malfunctions during measurement, not all wells (especially production and injection wells) possess complete logging curves in actual production. While a small number of missing logging curves may not significantly impact formation evaluation in manual well logging interpretation, the increasing digitalization and intelligentization of oil exploration and development has led to higher demands on the consistency and completeness of well logging data from various intelligent well logging interpretation algorithms based on machine learning and deep learning. Therefore, there is an urgent need to use appropriate methods to complete the missing logging curves in each well, improving the consistency and completeness of well logging data to meet the new requirements of the current digitalization and intelligentization process in oil exploration and development. Summary of the Invention

[0004] This invention provides a method for completing well logging curves, which accurately and quickly completes missing curves, improving the consistency and completeness of well logging data. The method includes:

[0005] The well logging curve to be completed is input into the well logging curve completion network, and the completed well logging curve is output through the well logging curve completion network. The well logging curve completion network is obtained by training a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed.

[0006] The well logging curve completion network completes the well logging curve to be completed in the following manner:

[0007] Based on the input logging curve to be completed, determine the set of logging curves and the missing codes corresponding to the set of curves, and convert the missing codes into a weighted feature sequence. The missing codes are the codes corresponding to the missing parts in the set of curves.

[0008] The dynamic weights of each channel are calculated based on the weighted feature sequence. The dynamic features of the target dimension of each channel are determined based on the curve set and the dynamic weights of each channel. The dynamic features of the target dimension of each channel are combined to obtain the dynamic combination features of the curve set.

[0009] The dynamic combination characteristics of the curve set are corrected, the missing curves are estimated based on the corrected dynamic combination characteristics, and the completed logging curves are output based on the missing curve estimation results.

[0010] This invention also provides a well logging curve completion device for accurately and quickly completing missing curves, thereby improving the consistency and completeness of well logging data. The device includes:

[0011] The well logging curve completion module is used to input the well logging curve to be completed into the well logging curve completion network, and output the completed well logging curve through the well logging curve completion network. The well logging curve completion network is obtained by training a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed.

[0012] The well logging curve completion module includes:

[0013] The curve conversion unit is used to determine the curve set of the well logging curve and the missing code corresponding to the curve set based on the input well logging curve to be completed, and convert the missing code into a weighted feature sequence. The missing code is the code corresponding to the missing part in the curve set.

[0014] The feature combination unit is used to calculate the dynamic weight of each channel based on the weight feature sequence, determine the dynamic features of the target dimension of each channel based on the curve set and the dynamic weight of each channel, and combine the dynamic features of the target dimension of each channel to obtain the dynamic combination features of the curve set.

[0015] The missing curve estimation unit is used to correct the dynamic combination characteristics of the curve set, estimate the missing curves based on the corrected dynamic combination characteristics, and output the completed logging curves based on the missing curve estimation results.

[0016] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described well logging curve completion method.

[0017] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described well logging curve completion method.

[0018] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described well logging curve completion method.

[0019] In this embodiment of the invention, the well logging curve to be completed is input into a well logging curve completion network, and the completed well logging curve is output through the well logging curve completion network. The well logging curve completion network is trained on a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed. The well logging curve completion network completes the well logging curve to be completed in the following way: based on the input well logging curve to be completed, the curve set of the well logging curve and the missing code corresponding to the curve set are determined, and the missing code is converted into a weight feature sequence, wherein the missing code is the code corresponding to the missing part in the curve set; the dynamic weight of each channel is calculated based on the weight feature sequence, and the dynamic features of the target dimension of each channel are determined based on the curve set and the dynamic weight of each channel; the dynamic features of the target dimension of each channel are combined to obtain the dynamic combination features of the curve set; the dynamic combination features of the curve set are corrected, the missing curve is estimated based on the corrected dynamic combination features, and the completed well logging curve is output based on the missing curve estimation result. In this way, a logging curve completion network is established based on a loss function that includes supervised learning and contrastive learning. By calculating dynamic weights and dynamically combining features to estimate missing curves, only one network model needs to be learned to effectively cope with various possible data conditions in the logging data of the actual work area. It can predict and complete the specified missing curves in the entire work area, and has obvious advantages in terms of accuracy and convenience. It can accurately and quickly complete the missing curves, and improve the consistency and completeness of logging data. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0021] Figure 1 This is a flowchart of the logging curve completion method provided in the embodiments of the present invention;

[0022] Figure 2 This is a schematic diagram illustrating the missing logging curves of each well in the work area provided in this embodiment of the invention;

[0023] Figure 3 This is a schematic diagram of the logging curve completion network structure provided in an embodiment of the present invention;

[0024] Figure 4 This is a schematic diagram illustrating the learning process of the logging curve completion network provided in this embodiment of the invention.

[0025] Figure 5 This is a schematic diagram of the logging curve completion device provided in an embodiment of the present invention;

[0026] Figure 6 This is a structural block diagram of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0028] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0029] In the description of this specification, the terms "comprising," "including," "having," and "containing" are open-ended terms, meaning that they include but are not limited to. The terms "an embodiment," "a specific embodiment," "some embodiments," and "for example," etc., refer to specific features, structures, or characteristics described in connection with that embodiment or example that are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. The order of steps involved in the various embodiments is used to illustrate the implementation of this application, and the order of steps is not limited and can be adjusted appropriately as needed.

[0030] Existing methods for predicting missing logging curves mainly fall into two categories: 1) Inter-well interpolation prediction methods: These methods are based on the continuity of formation development and the similarity between logging curve responses of adjacent wells. By spatially interpolating the logging curves of adjacent wells, the missing curve of the current well can be completed. However, this type of method requires accurate inter-well formation correlation results and is only applicable to densely populated well areas with stable formation development, which severely limits its application in complex geological conditions and when there is limited well data in the early stages of exploration. 2) Curve regression prediction methods: These methods mainly utilize the correlation between logging curves. Based on well data with complete logging data, samples are created to establish regression mapping relationships between missing curves and other curves, and regression prediction is used to complete the missing logging curves. Due to its low data requirements and ease of implementation, this type of method has been extensively studied and widely applied. Early research mainly relied on statistical regression or machine learning regression algorithms such as support vector machines, random forests, and gradient descent lift-up trees. With the widespread application of deep learning technology, a large number of missing curve completion algorithms based on end-to-end networks such as CNN, LSTM, GRU, and ConvLSTM have been proposed. Compared with traditional point-to-point regression algorithms, these networks can perform missing curve regression prediction in a curve-to-curve manner, which not only solves the problem caused by differences in curve resolution, but also makes full use of the correlation between different curves in terms of shape to make more accurate predictions.

[0031] However, existing methods all use fixed curve combinations as input for regression, which has limitations in completing missing curves in actual work areas. Figure 2 This paper presents statistics on missing logging curves for 39 wells in a specific work area. Figure 2 The results show that curve missingness in actual operations manifests in the following ways: 1) The degree of curve missingness varies across different wells. Only natural gamma and deep lateral logging are deployed across all wells, while other logging curves may be missing in all wells; 2) The degree of curve missingness varies across different depth segments within a specific well. This is because these wells are measured in multiple well runs, and the logging combinations used in different runs are not the same. This distribution of missing curves limits the use of fixed curve combinations as input for regression modeling in actual work area missing curve completion processing. Take the missing neutron curve as an example:

[0032] 1) If regression modeling is performed using all eight logging curves (natural gamma, density, sonic transit time, microelectrode, microgradient, shallow lateral, medium lateral, deep lateral) as input curves, on the one hand, since only a small number of wells have complete curves in certain well sections, this results in regression modeling learning based on only a small amount of data, facing the problem of insufficient samples; on the other hand, it also means that the resulting regression model cannot be deployed and applied in a large number of wells, such as W02 to W15, which lack microelectrode and microgradient.

[0033] 2) If regression modeling is performed using the combination of {natural gamma and deep lateral} curves common to all wells as input, although the data from the aforementioned 39 wells can all be used as sample data and the resulting model can be applied to all wells, the input curve information is relatively scarce. It is often difficult to accurately predict the compensated neutron curve based solely on natural gamma and deep lateral curves, wasting a large amount of density and acoustic transit time curve information that has already been measured in the wells and is most closely related to compensated neutrons.

[0034] 3) A compromise approach is to discard logging curves available only to a small number of wells, using only curves available to most wells as input for regression modeling. For example, inputs could include {natural gamma, density, sonic transit time, shallow lateral, deep lateral}. However, because the availability of missing curves varies across different wells and depths, the resulting regression model can only be applied to certain sections of some wells. For instance, it cannot be deployed for wells like W12 and W13, which lack density curves.

[0035] 4) To achieve compensated neutron curve prediction for all depth segments of the aforementioned 39 wells using a fixed curve input combination, it is necessary to set a large number of different input curve combinations based on the curve measurement conditions of each well at each depth segment to cover all possible logging data situations. Then, sample collection and regression modeling would be performed based on different input curve combinations, establishing a large number of different models to support application deployment. This is undoubtedly a tedious and time-consuming process.

[0036] Based on this, this invention addresses the problem that the missing curves in different wells and well sections are highly variable in actual work areas, making it difficult to use a unified curve combination input for regression modeling. It provides a well logging curve completion method, such as... Figure 1 As shown, it includes:

[0037] Step 101: Input the logging curve to be completed into the logging curve completion network, and output the completed logging curve through the logging curve completion network. The logging curve completion network is trained on a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical logging curves to be completed and corresponding historical completed logging curves. The logging curve completion network completes the logging curve to be completed in the following way:

[0038] Step 1011: Based on the input logging curve to be completed, determine the set of logging curves and the missing codes corresponding to the set of curves, and convert the missing codes into a weighted feature sequence. The missing codes are the codes corresponding to the missing parts in the set of curves.

[0039] Step 1012: Calculate the dynamic weight of each channel based on the weight feature sequence, determine the dynamic features of the target dimension of each channel based on the curve set and the dynamic weight of each channel, and combine the dynamic features of the target dimension of each channel to obtain the dynamic combination features of the curve set.

[0040] Step 1013: Correct the dynamic combination characteristics of the curve set, estimate the missing curves based on the corrected dynamic combination characteristics, and output the completed logging curves based on the missing curve estimation results.

[0041] The well logging curve completion method proposed in this invention establishes a well logging curve completion network based on a loss function that includes supervised learning and contrastive learning. By calculating dynamic weights and dynamically combining features to estimate missing curves, only one network model needs to be learned to effectively cope with various possible data conditions in the actual well logging data of the work area. It can predict and complete the specified missing curves in the entire work area, and has obvious advantages in terms of accuracy and convenience. It can accurately and quickly complete missing curves, and improve the consistency and completeness of well logging data.

[0042] In specific implementation, the logging curve completion network structure that supports dynamic input in this invention is as follows: Figure 3 As shown, this network takes a set of K curves of length L as input, uses an end-to-end CNN to perform curve-to-curve regression prediction, and outputs a target curve Y of length L. * The estimation results are presented. However, unlike existing methods, this invention allows for missing data in the input curve set X, and uses a missing code M of the same size to identify the missing data at each data point. Both the input curve set X and the missing code M are given in the form of a K-channel one-dimensional sequence, where the k-th channel x... k Let m represent the k-th dimension input curve, and m represent the k-th dimension input curve. k Then the identifier x k Is the previous sampling point missing?: If the k-th dimension input curve is missing at the l-th sampling point, then let the corresponding x... k,l =0 and m k,l =0, otherwise x k,l Take the corresponding response value and m k,l =1. In this way, if there are no missing curves in the input curve set X, then all corresponding M values ​​are 1. If some curves in X are missing in certain depth segments, then only the corresponding elements in X and M need to be set to 0, which can represent various cases in a unified form.

[0043] In one embodiment, obtaining historical logging curve samples to be completed includes:

[0044] For the acquired historical logging curve samples, a sample curve of a preset length is extracted using a sliding window method;

[0045] Randomly select a portion of the curves from the sample curves, and set the corresponding values ​​of the selected curves to 0 to obtain a sample of historical logging curves to be completed.

[0046] In one embodiment, converting the missing code into a weighted feature sequence includes:

[0047] By setting pointwise convolutional layers and PReLU activation layers, the missing encoding is transformed into a weighted feature sequence.

[0048] In one embodiment, calculating the dynamic weight of each channel based on the weight feature sequence includes:

[0049] Using multiple dynamic combination units, the dynamic weights required for linear sufficiency of each channel in the curve set are calculated based on the weight feature sequence. The dynamic combination units include pointwise convolutional layers and PReLU activation layers.

[0050] Based on the input curve set X and the missing code M, the network will first use, for example... Figure 3 The dynamic combined feature extraction module shown dynamically determines the curve combination weights based on the missing curve represented by M to transform X into a dynamic combined feature Z. Specifically, it first uses a pointwise convolutional layer and a PReLU activation layer to transform the missing code M into a 64-channel weighted feature sequence V = {v1, v2, ..., v...}. 64},in:

[0051]

[0052] In formula (1) and These are the learnable network parameters in the aforementioned pointwise convolutional layer and PReLU activation layer, respectively. The dynamic feature combination extraction module will use 64 dynamic combination units to continue calculating the dynamic weights required for linear sufficiency of the K channels in X based on V. A dynamic combination unit also consists of a pointwise convolutional layer and a PReLU activation layer, and the dynamic weights obtained by the j-th dynamic combination unit can be expressed as... in:

[0053]

[0054] In equation (2) and ... k This is to ensure that when the k-th dimension input curve is missing at the l-th sampling point, the corresponding dynamic combination weights are used. Values ​​that are always zero or missing are not included in the calculation of dynamic features. The dynamic feature combination extraction module will use W... (j) The K-channel input X is linearly recombined into the j-th dimension dynamic feature:

[0055]

[0056] This process does not involve learning network parameters. Concatenating the results of the 64 dynamic combination units yields the corresponding 64-channel dynamic combination feature Z = {z1, z2, ..., z...}. 64 The core of the dynamic combination feature extraction module lies in learning from the samples and adaptively determining the weights used to linearly reconstruct the input curves in X based on the missing data points reflected by M, so that the resulting 64-channel dynamic combination feature Z is as unaffected as possible by the missing data points in X.

[0057] In one embodiment, determining the dynamic combination characteristics of the curve set and modifying the dynamic combination characteristics of the curve set includes:

[0058] Identify the morphological response features related to the logging curve to be completed in the dynamic combination features;

[0059] Based on the morphological response characteristics, the dynamic combination characteristics are modified through residual connections.

[0060] On this basis Figure 3 The network shown will continue to use the missing curve reconstruction prediction module to perform regression prediction and reconstruction of the missing curve based on Z. Specifically, the implementation will first cascade four residual connection modules (each residual connection module consists of a PReLU activation layer, a one-dimensional convolutional layer with a kernel size of 9, and a normalization layer). These modules will fully utilize the curve morphology feature extraction capability of the convolutional neural network to mine the morphological response features in Z related to the missing curve to be predicted, and correct Z through residual connections to obtain a more accurate 64-channel feature U = {u1, u2, ..., u...}. 64 Then, this module will use shared weights, employing the same set of PReLU activation layers and pointwise convolutional layers, to fuse Z and U into the corresponding missing curves for preliminary estimation of Y. P The final estimate of Y from the missing curve * ,Right now:

[0061]

[0062] in and Y P and Y *The learnable parameters in the shared PReLU activation layers and pointwise convolutional layers are calculated. The missing curve is ultimately estimated as Y. * yes Figure 3 The final output of the network, and the initial estimate of Y P The dynamically combined feature Z is used as an auxiliary output of the network to assist in its learning. Here, a shared PReLU activation layer and a pointwise convolutional layer Z are fused together to ensure that the four residual connection modules used only fine-tune the details in Z to U for more accurate missing curve prediction, without significantly changing the semantics of the dynamically combined feature Z.

[0063] In one embodiment, the well logging curve completion network is trained as follows:

[0064] Obtain historical logging curve samples to be completed, determine the historical completed logging curve samples corresponding to the historical logging curves to be completed, and construct training and testing sets based on the historical logging curves to be completed and the historical completed logging curve samples corresponding to the historical logging curves to be completed.

[0065] An initial curve completion network is constructed using a contrastive learning approach. Supervised loss terms and contrastive loss terms are defined, and a loss function is constructed based on the supervised loss terms and contrastive loss terms.

[0066] Based on the loss function, with minimizing the loss function value as the training objective, the initial curve completion network is trained and tested using the training set and the test set.

[0067] When the loss function value reaches the target value, the logging curve completion network is obtained.

[0068] For the actual work area data to be processed, this embodiment of the invention will first collect samples of the logging curve Y to be completed. Specifically, samples will be collected in a sliding window manner, with 512 logging points as one well segment. Any well segment with complete Y values, regardless of whether other curves are complete, can be considered as a sample. The specific curve input combination is taken as the union of the input curves in the collected samples. For a specific sample, if there are missing values ​​in the input X, the corresponding elements of X and M for the missing well segment will be set to 0, while the corresponding elements of M for the unmissing part will remain unchanged at 1. Based on these samples, this invention will proceed according to... Figure 4 The curve completion network is learned and constructed in the manner shown.

[0069] A sample collected from the work area can be denoted as {X1, M1: Y}, where X1 and M1 are the input curve set and their corresponding missing codes, respectively, and Y is the label of the curve to be completed. {X1, M1} can be a complete input set without any missing values, or it can be an input set with partial missing values. By simply setting the values ​​corresponding to the missing measurement points in X1 and M1 to zero, a unified method can be used to represent various situations. Based on this, we can randomly set some curves (or local sections of some curves) in {X1, M1} to zero to obtain another corresponding sample {X2, M2: Y}. If the input {X1, M1} has no missing values, this process actually simulates possible curve missing situations; if {X1, M1} itself has missing values, this process actually simulates situations with more missing values.

[0070] Based on {X1,M1:Y} and {X2,M2:Y}, the aforementioned curve completion network can be trained using a contrastive learning approach. Let the network model be... Where Θ represents the network parameters that need to be determined through learning, then substituting {X1,M1} and {X2,M2} into... It can be calculated that:

[0071]

[0072] Since the ultimate goal of this network is to predict and complete the missing curve Y, then the first requirement is... The final estimate on {X1,M1} and {X2,M2} and To ensure consistency with its label Y as much as possible, a supervised loss term can be defined:

[0073]

[0074] However, simply backpropagating the error to the final output of the network might prevent the dynamic feature extraction module from learning effectively. Considering that the purpose of this module is to adaptively and dynamically combine X point-by-point into Z based on M, minimizing the impact of missing values ​​in the input on Z, then for {X1,M1} and {X2,M2} that differ only in their missing values, the difference between their corresponding Z1 and Z2 should be as small as possible. Therefore, a contrastive loss term can be defined:

[0075]

[0076] Simply requiring Z1 and Z2 to be as consistent as possible may lead to erroneous local minima; for example, the magnitudes of Z1 and Z2 should be as small as possible, achieving optimality when both approach 0. To avoid this, we employ the following measures. First, as described in the aforementioned network structure, we require that Z1 and Z2 can be used to obtain a preliminary estimate of Y through a PReLU activation layer and a pointwise convolutional layer, as shown in Equation (4). This can be achieved by requiring a preliminary estimate... and It should be as consistent as possible with its label Y, that is:

[0077]

[0078] At the same time, to prevent Z1 and Z2 from satisfying Since the magnitudes tend to be smaller, we first adopt the shared parameters according to the forms in equations (4) and (5). and Preliminary and final estimates were calculated. Additionally, [the calculations were performed]. Added a magnitude regularization term, requiring It must not be greater than 1, that is:

[0079]

[0080] By introducing the loss term in equations (9) and (10), the network's learning requires the dynamic combination feature extraction module to eliminate the influence of missing data, adaptively expand X1 and X2 into Z1 and Z2 according to M1 and M2, and minimize the difference between them. Simultaneously, Z1 and Z2 are required to be of the same order of magnitude as Y and contain sufficient information for preliminary prediction of Y. In summary, the loss function for curve network learning in this invention is:

[0081]

[0082] The network can be optimized by minimizing the loss function in equation (11) based on the sample data collected from the actual work area to be processed using the aforementioned method. The parameters Θ are learned and fitted to establish the required missing curve completion network. For the logging data of the well to be completed, a new curve set input X and corresponding missing codes M are constructed according to the order of the curves in the set of input curves. If a missing curve exists, the corresponding values ​​of X and M are set to 0. X and M are then substituted into the established network model. The process only requires reasoning, and then taking the predicted Y. * This can be used as the final prediction result for the curve to be completed.

[0083] For example, based on Figure 2The advantages of this method over existing methods are illustrated using the missing neutron curve completion data of the work area shown in the figure. Specific applications will be compared with existing pointwise regression mapping methods and methods based on RNNs and CNNs. First, in the sample collection phase, we reserved well W35, with the most complete curve measurement, as the test well. Then, according to several curve combinations shown in Appendix 1, we collected learning samples from the remaining 38 wells and counted the number of samples collected in this work area, as shown in Appendix 1. It can be seen that if we use combination 4, with {natural gamma, deep lateral} which is present in all wells as input, a total of 238,843 samples can be collected. However, if we use combination 1, with all other curves as input, only 33,826 samples can meet the conditions. For existing methods, how the input curve is set directly affects the number of samples that the network can use for learning. However, for the method in this invention, since it can dynamically accommodate all possible curve combinations, all 238,843 samples from the remaining 38 wells can be used for network learning. This will enable the construction of networks based on the method of this invention to obtain more sufficient sample support.

[0084] Table 1. Example of missing neutron curve completion input settings

[0085]

[0086] In specific applications, the method of this invention is compared with pointwise prediction DNNs and similar end-to-end (curve-to-curve) prediction CNNs, LSTMs, and GRUs. Corresponding models are built according to the input combinations shown in Appendix 1, and the NMSE relative error of the prediction results on the W35 well is calculated. It is important to note that for the four algorithms being compared, since a single model cannot be compatible with different input curve combinations, a separate network model needs to be learned for each of the four combinations. That is, each row of these algorithms in Appendix 2 is derived from a different model. However, for the method of this invention, only one model needs to be trained, and this model can use all 238843. During prediction, only the channels of the corresponding curve in the input curve combination X and the corresponding missing code M need to be set to 0 for prediction. This allows for compatibility with different input curve combinations, fully demonstrating the convenience of the proposed method.

[0087] Table 2. Compensated Neutron Curve Completion Prediction NMSE for W35 Inoue Algorithms

[0088]

[0089] In specific applications, the method of this invention is compared with point-by-point prediction DNNs and similar end-to-end (curve-to-curve) prediction methods such as CNNs, LSTMs, and GRUs. Corresponding models are built according to the input combinations shown in Appendix 1, and the NMSE relative error of the prediction results on the W35 well is calculated. It is important to note that for the four algorithms being compared, since a single model cannot be compatible with different input curve combinations, a separate network model needs to be learned for each of the four combinations. That is, each row of these algorithms in Appendix 2 is derived from a different model. However, for the method of this invention, only one model needs to be trained, and this model can be learned using all 238,843 samples. During prediction, only the channels of the corresponding curves in the input curve combination X and the corresponding missing code M need to be set to 0, thus enabling compatibility with different input curve combinations and fully demonstrating the convenience of the proposed method.

[0090] Comparing the NMSE prediction results of W35 Inoue under different algorithms and combinations in Appendix Table 2, it can be seen that end-to-end (curve-to-curve) prediction of missing curves is always superior to point-by-point prediction. However, for existing end-to-end CNN, LSTM, and GRU, they all achieved the best application results on combination 3. This is mainly because for combination 4, only natural gamma and deep lateral can effectively predict the compensated neutron response; while for groups 1 and 2, although more curves are included as input, the higher requirement for the completeness of the input curves during the sample collection stage leads to fewer collected samples, and the network cannot learn more fully, resulting in poor generalization ability. Since the method of this invention can use all 238,843 samples for learning, it is not affected by the above-mentioned sample quantity problem. As can be seen from the results in Appendix Table 2, as the number of input curves increases, the relative error of the prediction NMSE of the method of this invention decreases significantly, achieving the best results on group 1. This shows that the method of this invention can make fuller use of the input curve information. Furthermore, due to more sufficient sample support, the method of this invention achieves the minimum prediction error in all input combinations. This further demonstrates the advantage of the method of this invention in prediction accuracy compared to existing methods.

[0091] This invention also provides a logging curve completion device, as described in the following embodiments. Since the principle by which this device solves the problem is similar to that of the logging curve completion method, the implementation of this device can be referred to the implementation of the logging curve completion method, and repeated details will not be elaborated further.

[0092] Figure 5 This is a schematic diagram of the logging curve completion device provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the device includes:

[0093] The well logging curve completion module 501 is used to input the well logging curve to be completed into the well logging curve completion network, and output the completed well logging curve through the well logging curve completion network. The well logging curve completion network is trained on a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed. The well logging curve completion module includes:

[0094] The curve conversion unit 5011 is used to determine the curve set of the well logging curve and the missing code corresponding to the curve set based on the input well logging curve to be completed, and convert the missing code into a weighted feature sequence. The missing code is the code corresponding to the missing part in the curve set.

[0095] The feature combination unit 5012 is used to calculate the dynamic weight of each channel according to the weight feature sequence, determine the dynamic features of the target dimension of each channel according to the curve set and the dynamic weight of each channel, and combine the dynamic features of the target dimension of each channel to obtain the dynamic combination features of the curve set.

[0096] The missing curve estimation unit 5013 is used to correct the dynamic combination characteristics of the curve set, estimate the missing curves based on the corrected dynamic combination characteristics, and output the completed logging curves based on the missing curve estimation results.

[0097] In one embodiment, a network training module is further included, specifically for:

[0098] Obtain historical logging curve samples to be completed, determine the historical completed logging curve samples corresponding to the historical logging curves to be completed, and construct training and testing sets based on the historical logging curves to be completed and the historical completed logging curve samples corresponding to the historical logging curves to be completed.

[0099] An initial curve completion network is constructed using a contrastive learning approach. Supervised loss terms and contrastive loss terms are defined, and a loss function is constructed based on the supervised loss terms and contrastive loss terms.

[0100] Based on the loss function, with minimizing the loss function value as the training objective, the initial curve completion network is trained and tested using the training set and the test set.

[0101] When the loss function value reaches the target value, the logging curve completion network is obtained.

[0102] In one embodiment, the network training module is specifically used for:

[0103] For the acquired historical logging curve samples, a sample curve of a preset length is extracted using a sliding window method;

[0104] Randomly select a portion of the curves from the sample curves, and set the corresponding values ​​of the selected curves to 0 to obtain a sample of historical logging curves to be completed.

[0105] In one embodiment, the curve conversion unit 5011 is specifically used for:

[0106] By setting pointwise convolutional layers and PReLU activation layers, the missing encoding is transformed into a weighted feature sequence.

[0107] In one embodiment, the feature combining unit 5012 is specifically used for:

[0108] Using multiple dynamic combination units, the dynamic weights required for linear sufficiency of each channel in the curve set are calculated based on the weight feature sequence. The dynamic combination units include pointwise convolutional layers and PReLU activation layers.

[0109] In one embodiment, the missing value estimation unit 5013 is specifically used for:

[0110] Identify the morphological response features related to the logging curve to be completed in the dynamic combination features;

[0111] Based on the morphological response characteristics, the dynamic combination characteristics are modified through residual connections.

[0112] Based on the aforementioned inventive concept, such as Figure 6 As shown, the present invention also proposes a computer device 600, including a memory 610, a processor 620, and a computer program 630 stored in the memory 610 and executable on the processor 620. When the processor 620 executes the computer program 630, it implements the aforementioned well logging curve completion method.

[0113] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described well logging curve completion method.

[0114] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described well logging curve completion method.

[0115] In summary, in this embodiment of the invention, the well logging curve to be completed is input into a well logging curve completion network, and the completed well logging curve is output through the well logging curve completion network. The well logging curve completion network is trained on a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed. The well logging curve completion network completes the well logging curve to be completed in the following way: based on the input well logging curve to be completed, the curve set of the well logging curve and the missing codes corresponding to the curve set are determined; the missing codes are converted into weight feature sequences, where the missing codes are the codes corresponding to the missing parts in the curve set; the dynamic weights of each channel are calculated based on the weight feature sequences; the dynamic features of the target dimension of each channel are determined based on the curve set and the dynamic weights of each channel; the dynamic features of the target dimension of each channel are combined to obtain the dynamic combination features of the curve set; the dynamic combination features of the curve set are corrected; the missing curve is estimated based on the corrected dynamic combination features; and the completed well logging curve is output based on the missing curve estimation results. In this way, a logging curve completion network is established based on a loss function that includes supervised learning and contrastive learning. By calculating dynamic weights and dynamically combining features to estimate missing curves, only one network model needs to be learned to effectively cope with various possible data conditions in the logging data of the actual work area. It can predict and complete the specified missing curves in the entire work area, and has obvious advantages in terms of accuracy and convenience. It can accurately and quickly complete the missing curves, and improve the consistency and completeness of logging data.

[0116] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0117] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments 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 within the scope of protection of the present invention.

Claims

1. A method for completing well logging curves, characterized in that, include: The well logging curve to be completed is input into the well logging curve completion network, and the completed well logging curve is output through the well logging curve completion network. The well logging curve completion network is obtained by training a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed. The well logging curve completion network completes the well logging curve to be completed in the following manner: Based on the input logging curve to be completed, determine the set of logging curves and the missing codes corresponding to the set of curves, and convert the missing codes into a weighted feature sequence. The missing codes are the codes corresponding to the missing parts in the set of curves. The dynamic weights of each channel are calculated based on the weighted feature sequence. The dynamic features of the target dimension of each channel are determined based on the curve set and the dynamic weights of each channel. The dynamic features of the target dimension of each channel are combined to obtain the dynamic combination features of the curve set. The dynamic combination characteristics of the curve set are corrected, the missing curves are estimated based on the corrected dynamic combination characteristics, and the completed logging curves are output based on the missing curve estimation results.

2. The method as described in claim 1, characterized in that, The well logging curve completion network was trained in the following way: Obtain historical logging curve samples to be completed, determine the historical completed logging curve samples corresponding to the historical logging curves to be completed, and construct training and testing sets based on the historical logging curves to be completed and the historical completed logging curve samples corresponding to the historical logging curves to be completed. An initial curve completion network is constructed using a contrastive learning approach. Supervised loss terms and contrastive loss terms are defined, and a loss function is constructed based on the supervised loss terms and contrastive loss terms. Based on the loss function, with minimizing the loss function value as the training objective, the initial curve completion network is trained and tested using the training set and the test set. When the loss function value reaches the target value, the logging curve completion network is obtained.

3. The method as described in claim 2, characterized in that, Obtain historical logging curve samples that need to be supplemented, including: For the acquired historical logging curve samples, a sample curve of a preset length is extracted using a sliding window method; Randomly select a portion of the curves from the sample curves, and set the corresponding values ​​of the selected curves to 0 to obtain a sample of historical logging curves to be completed.

4. The method as described in claim 1, characterized in that, The missing codes are transformed into a weighted feature sequence, including: By setting pointwise convolutional layers and PReLU activation layers, the missing encoding is transformed into a weighted feature sequence.

5. The method as described in claim 1, characterized in that, The dynamic weights of each channel are calculated based on the weighted feature sequence, including: Using multiple dynamic combination units, the dynamic weights required for linear sufficiency of each channel in the curve set are calculated based on the weight feature sequence. The dynamic combination units include pointwise convolutional layers and PReLU activation layers.

6. The method as described in claim 1, characterized in that, The dynamic combination characteristics of the curve set are modified, including: Identify the morphological response features related to the logging curve to be completed in the dynamic combination features; Based on the morphological response characteristics, the dynamic combination characteristics are modified through residual connections.

7. A logging curve completion device, characterized in that, include: The well logging curve completion module is used to input the well logging curve to be completed into the well logging curve completion network, and output the completed well logging curve through the well logging curve completion network. The well logging curve completion network is obtained by training a machine learning model based on a loss function that includes supervised learning and contrastive learning, using historical well logging curves to be completed and historical completed well logging curves corresponding to the historical well logging curves to be completed. The well logging curve completion module includes: The curve conversion unit is used to determine the curve set of the well logging curve and the missing code corresponding to the curve set based on the input well logging curve to be completed, and convert the missing code into a weighted feature sequence. The missing code is the code corresponding to the missing part in the curve set. The feature combination unit is used to calculate the dynamic weight of each channel based on the weight feature sequence, determine the dynamic features of the target dimension of each channel based on the curve set and the dynamic weight of each channel, and combine the dynamic features of the target dimension of each channel to obtain the dynamic combination features of the curve set. The missing curve estimation unit is used to correct the dynamic combination characteristics of the curve set, estimate the missing curves based on the corrected dynamic combination characteristics, and output the completed logging curves based on the missing curve estimation results.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

Images