Data generation method and device based on geophysical large model, equipment and storage medium

Abstract

The application discloses the technical field of geophysical logging, and discloses a data generation method and device based on a geophysical large model, equipment and a storage medium, the method comprises the following steps: obtaining standard logging data of a sample well point, and using a mask to cover part of the logging data in the standard logging data to obtain a mask sequence; in the process of generating the part of the logging data covered by the mask by training a geophysical logging large model, the features of different logging series in the mask sequence are interacted by using a multi-head self-attention layer to establish cross-series association; obtaining original logging data of a target well point, and generating target logging data missing the original logging data by using the trained geophysical logging large model. The application can generate relevant logging data by the model, so that the missing logging data caused by various interference factors in the data acquisition process is made up, and the accuracy of geological research is further improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of geophysical logging, and in particular to a data generation method and device based on a geophysical large model, equipment and a storage medium. BACKGROUND

[0002] Geological information can be obtained by measuring the physical properties of the stratum, so as to explore and develop underground resources (such as oil fields). However, conventional logging data is easily disturbed by various non-geological factors during the acquisition process, such as measurement environment, instrument difference, etc., resulting in missing data (such as missing some data, setting data of a logging series, or data affected by noise to be unusable) in the logging data, affecting the accuracy of geological research.

[0003] To sum up, how to effectively solve the problem of missing logging data and the like is a technical problem that needs to be solved by the technical personnel in the field at present. SUMMARY

[0004] The purpose of the present application is to provide a data generation method and device based on a geophysical large model, equipment and a storage medium, which generates target logging data missing in original logging data by using the measured original logging data as a medium, so as to make up for the missing logging data caused by various interference factors in the data acquisition process, and further improve the accuracy of geological research.

[0005] To solve the above technical problems, the present application provides the following technical solutions.

[0006] A data generation method based on a geophysical large model, comprising:

[0007] Obtaining standard logging data of a sample well point, and using a mask to mask part of the logging data in the standard logging data to obtain a mask sequence;

[0008] Using the mask sequence and taking generation of the part of the logging data as a target, training a geophysical logging large model with a multi-head self-attention layer;

[0009] Obtaining original logging data of a target well point, and using the trained geophysical logging large model to generate target logging data missing in the original logging data;

[0010] In the process of training the geophysical logging large model, the multi-head self-attention layer is used to interact the features of different logging series in the mask sequence, to establish cross-series association, so as to generate the part of the logging data.

[0011] Preferably, the mask is used to mask part of the logging data in the standard logging data to obtain a mask sequence, comprising:

[0012] Proportionally mask data in a single logging series in the standard logging data to obtain an internal mask sequence;

[0013] Proportionally mask data in different logging series in the standard logging data to obtain an interactive mask sequence;

[0014] Superimpose the internal mask sequence and the interactive mask sequence to obtain a comprehensive mask sequence;

[0015] Determine the internal mask sequence, the interactive mask sequence, or the comprehensive mask sequence as the mask sequence.

[0016] Preferably, superimposing the internal mask sequence and the interactive mask sequence comprises:

[0017] If the same data point is masked in both the internal mask sequence and the interactive mask sequence, discard the mask corresponding to the internal mask sequence.

[0018] Preferably, the geophysical logging large model with multi-head self-attention layer comprises:

[0019] Create a converter and add a multi-head self-attention layer to the converter to capture the global dependency between different depth points in the logging curve using the multi-head self-attention layer;

[0020] Add a feedforward network for nonlinear transformation of features for each depth point;

[0021] Use residual structure and normalization to accelerate training convergence and alleviate gradient vanishing;

[0022] Embed the longitudinal position information of the logging curve through depth position encoding to enable the geophysical logging large model to perceive the relative positions of different depth points in the formation;

[0023] Perform Fourier transform on the logging curve to extract periodic features and fuse them with original time domain features;

[0024] Concatenate or weightedly fuse the features coded in time domain and frequency domain to form multi-scale feature representation and input to the converter for global modeling.

[0025] Preferably, the trained geophysical logging large model is used to generate target logging data missing in the original logging data, comprising:

[0026] Serializing the original logging data to obtain a numerical sequence;

[0027] Input the numerical sequence into the trained geophysical logging large model for data generation processing to obtain the target logging data.

[0028] Preferably, the target logging data missing the original logging data is generated by using the trained geophysical logging large model, comprising:

[0029] The original logging data of the missing target structured logging data series is input into the trained geophysical logging large model to generate the target logging data corresponding to the target structured logging data series;

[0030] The target structured logging data series is at least one of a lithology logging series, a resistivity logging series and a porosity logging series, and the target logging data is at least one logging data in the target structured logging data series.

[0031] Preferably, after obtaining the original logging data of the target well point and generating the target logging data missing the original logging data by using the trained geophysical logging large model, the method further comprises:

[0032] The original logging data is supplemented by using the target logging data to obtain the to-be-interpreted logging data;

[0033] Fine-tuning the trained geophysical logging large model;

[0034] Interpreting the to-be-interpreted logging data by using the fine-tuned geophysical logging large model.

[0035] A data generation device based on a geophysical large model, comprising:

[0036] A mask sequence acquisition module is configured to acquire standard logging data of a sample well point, and to obtain a mask sequence by using a mask to mask part of the logging data in the standard logging data;

[0037] A model training module is configured to train a geophysical logging large model having a multi-head self-attention layer by using the mask sequence and aiming to generate the part of the logging data; wherein, in the process of training the geophysical logging large model, the multi-head self-attention layer is used to interact the features of different logging series in the mask sequence, to establish cross-series association, so as to generate the part of the logging data;

[0038] A data generation module is configured to acquire original logging data of a target well point, and to generate target logging data missing the original logging data by using a trained geophysical logging large model.

[0039] An electronic device, comprising:

[0040] A memory is configured to store a computer program;

[0041] A processor for implementing the steps of the above-mentioned geophysical large model-based data generation method when executing the computer program.

[0042] A readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the above-mentioned geophysical large model-based data generation method.

[0043] By applying the method provided in the embodiments of the present application, standard logging data of sample well points is obtained, and part of the logging data in the standard logging data is masked by using a mask to obtain a mask sequence; the geophysical logging large model with a multi-head self-attention layer is trained by using the mask sequence and taking generation of the part of the logging data as a target; original logging data of target well points is obtained, and target logging data missing in the original logging data is generated by using the trained geophysical logging large model; and in the process of training the geophysical logging large model, the multi-head self-attention layer is used to interact features of different logging series in the mask sequence to establish cross-series association so as to generate the part of the logging data.

[0044] In the present application, it is considered that due to various factors, the measured original logging data may have a problem of data missing in the measurement process. Based on the fact that there is some association relationship between different logging data, it is proposed to generate missing target logging data based on existing original logging data. Specifically, standard logging data of sample well points can be provided first, and part of the logging data in the standard logging data is masked by using a mask, that is, logging data with part of the data missing is obtained as a mask sequence. Then, the multi-head self-attention layer can interact different features and establish association of different types of data, and in the present application, the geophysical logging large model with the multi-head self-attention layer is trained by using the mask sequence. In the process of training the geophysical logging large model to generate the part of the logging data masked by the mask, the multi-head self-attention layer is used to interact features of different logging series in the mask sequence to establish cross-series association. That is, through training, the geophysical logging large model can learn the association between different logging series based on the multi-head self-attention layer therein, so that the geophysical logging large model can generate missing logging parameters. After the model training is completed, the trained geophysical logging large model can be used to generate target logging data missing in the original logging data collected by target well points. In this way, the related logging data can be generated by the model, so as to make up for the missing logging data caused by various interference factors in the data collection process, and further improve the accuracy of geological research.

[0045] Correspondingly, the application further provides an apparatus and a device for generating data based on a geophysical large model and a storage medium, which have the above technical effects and will not be described here. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description only represent some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort.

[0047] Figure 1 An implementation flowchart of a method for generating data based on a geophysical large model in an embodiment of the present application;

[0048] Figure 2 An SP curve data consistency analysis diagram of a target well point before processing well logging data in an embodiment of the present application;

[0049] Figure 3 An SP curve data consistency analysis diagram of a target well point after processing well logging data in an embodiment of the present application;

[0050] Figure 4 An implementation flowchart of a model training in an embodiment of the present application;

[0051] Figure 5 A structural diagram of an apparatus for generating data based on a geophysical large model in an embodiment of the present application;

[0052] Figure 6 A structural diagram of an electronic device in an embodiment of the present application;

[0053] Figure 7 A specific structural diagram of an electronic device in an embodiment of the present application. DETAILED DESCRIPTION

[0054] In order to make the person skilled in the art better understand the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort are within the scope of protection of the present application.

[0055] Please refer to Figure 1 , Figure 1 A flowchart of a method for generating data based on a geophysical large model in an embodiment of the present application, which comprises the following steps.

[0056] S101. Obtain standard logging data for sample well points, and use a mask to cover part of the logging data in the standard logging data to obtain a mask sequence.

[0057] The sample well points can be representative well points in the area to be logged. The standard logging data can be logging data that has been verified and is confirmed to be complete.

[0058] Specifically, the standard logging data can be preprocessed into a series of structured logging data, such as lithological logging series (natural gamma (GR), spontaneous potential (SP), caliper (CAL)), resistivity logging series (deep lateral resistivity (RILD), medium lateral resistivity (RILM), shallow lateral resistivity (RILS)), and porosity logging series (neutron (CN), density (DEN), acoustic (AC)), by following steps such as consistency, normalization, and outlier handling, and then converted into logging data in the form of a numerical sequence of equal length.

[0059] To gain a holistic view of the data distribution across all well logging curves, a violin plot was created for well logging data consistency analysis. Taking the SP curve as an example... Figure 2 This paper demonstrates the data distribution characteristics of different logging curves before data consistency processing. It reveals significant and highly uneven distributions of SP logging curves from different wells. To eliminate inter-well discrepancies and standardize dimensions, data processing was performed on each sample well. Firstly, based on the integrity of core data and logging interpretation, W5 was selected as the standard well, and the data was processed using the formula... The data distribution range of other wells is incorporated into the range of standard wells to achieve the goal of logging data consistency. The processed results are as follows: Figure 3 As shown, the data distribution range of all wells is basically the same, eliminating the inter-well discrepancies in the original logging data. Among them, For the processed data; The raw logging data for the target well to be processed; This represents the minimum value for the wells to be processed. The maximum value of the target well point to be processed; This is the maximum value for a standard well; This is the minimum value for a standard well.

[0060] The data was then normalized and outliers were removed. The normalization method used was Z-score normalization. Specifically, [the following steps were taken]. Processing is carried out, among which, Normalized data, The data before normalization. , These are the mean and standard deviation of the original dataset, respectively.

[0061] The data consistency of the standard well is aligned, and the dimension and distribution range of the multi-well data are unified by combining Z-score normalization. By dynamically mapping the well data to the standard well distribution, systematic deviation is eliminated, and the migration error of the model between different wells is reduced.

[0062] In the present application, in order to make the model can be based on the correlation between the existing logging data and the missing logging data, so as to generate the missing logging data based on the existing logging data. The mask sequence with part of the logging data missing can be obtained by masking the standard logging data in advance.

[0063] That is, the difference between the mask sequence and the standard logging data is that part of the logging data in the mask sequence is masked by the mask.

[0064] In a specific embodiment of the present application, the mask is used to mask part of the logging data in the standard logging data to obtain a mask sequence, comprising:

[0065] The data of a single logging series in the standard logging data is masked in proportion to obtain an internal mask sequence;

[0066] The data of different logging series in the standard logging data is masked in proportion to obtain an interactive mask sequence;

[0067] Superimpose the internal mask sequence and the interactive mask sequence to obtain a comprehensive mask sequence;

[0068] The internal mask sequence, the interactive mask sequence or the comprehensive mask sequence is determined as the mask sequence.

[0069] Wherein, superimposing the internal mask sequence and the interactive mask sequence comprises: if the same data point is masked in the internal mask sequence and the interactive mask sequence, discarding the mask corresponding to the internal mask sequence.

[0070] Considering that in the actual logging process, the missing data can have many situations, such as it can be only a number of logging data of a logging system, or it can be a number of logging data of multiple logging series, therefore, in actual application, different masking methods can be used to simulate the data loss corresponding to data acquisition.

[0071] Specifically, the data of a single logging series in the standard logging data can be masked in proportion to obtain an internal mask sequence simulating the data loss corresponding to a single logging series.

[0072] The data of different logging series in the standard logging data can also be proportionally masked to obtain an interactive mask sequence simulating missing data of multiple logging series.

[0073] The internal mask sequence and the interactive mask sequence can also be superimposed to obtain a comprehensive mask sequence simulating missing data in a complex situation.

[0074] The mask sequence can be the internal mask sequence, the interactive mask sequence or the comprehensive mask sequence.

[0075] For example, the following illustrates that a complex input data is generated through a multi-level mask strategy to simulate local noise and cross-series data missing in a real logging scene, so as to train the robust modeling capability of the model for multi-source logging data.

[0076] Specifically, the internal mask is generated to simulate single-series local noise and data missing. The data can be proportionally masked in a single logging series (such as resistivity, porosity). 80% of the data is retained to maintain the integrity of most of the original data and provide context information; 10% is replaced with [MASK] to force the model to learn to reconstruct the masked key data points through the context; and 10% is randomly replaced to simulate instrument random errors (such as drift or transient interference) and enhance the robustness of the model to noise. Further, the single-series local noise and data segment missing problem can be solved, such as instrument error or local data loss.

[0077] The interactive mask is generated to simulate cross-series synchronous missing. Parallel masking can be performed between multiple logging series (such as resistivity and acoustic wave) in proportion. 30% of the data between series: randomly select part of the series combination; 10% of the data is parallelly masked: the data at the same depth position in the selected series is synchronously masked (such as resistivity and porosity), which simulates multi-instrument synchronous failure or data acquisition blind area. Further, the model can be forced to mine potential correlations between cross-series (such as the cooperative relationship between resistivity anomaly and lithology change), and the adaptability to cross-series data missing scene is improved.

[0078] The mask sequence fusion constructs a complex input data. The mask data generated by the internal mask (single-series noise) and the interactive mask (cross-series synchronous missing) can be superimposed to form the final input sequence. The key rule is that if a data point is covered by both the internal mask and the interactive mask, the priority follows the interactive mask (because it simulates a more complex cross-series missing scene). In the final input sequence, the mask area needs to retain the original mask symbol (such as [MASK]) or random noise, and the non-mask area maintains the original value. This module generates a complex input that can reflect both single-series local anomalies and cross-series synchronous missing, and comprehensively trains the comprehensive reconstruction capability of the model.

[0079] The composite mask strategy (internal mask simulates local noise, and interactive mask simulates cross-series synchronous missing) is adopted. In a real logging scene, instrument errors and multi-instrument synchronous failures (such as high temperature causing distortion of resistivity and acoustic data at the same time) frequently occur. By forcing the model to reconstruct the masked data from the context, the reconstruction accuracy of the model is greatly improved when 30% of the cross-series data is missing, and the sensitivity to random noise is reduced.

[0080] By modeling the mask sequence, learning general rules from unlabeled data, and combining dynamic alignment to reduce regional distribution differences, the fine-tuning period of new oilfield data migration is shortened, and similar performance to full supervision can be achieved with only part of the labeled data.

[0081] S102, training a geophysical logging large model with a multi-head self-attention layer by using the mask sequence and targeting the generation of the partial logging data.

[0082] During the training of the geophysical logging large model, the multi-head self-attention layer is used to interact the features of different logging series in the mask sequence, establish cross-series association, and generate the partial logging data.

[0083] Because the multi-head self-attention layer interacts and associates different features, in this embodiment, the geophysical logging large model with the multi-head self-attention layer can be trained.

[0084] In a specific embodiment of the present application, a geophysical logging large model with a multi-head self-attention layer is built, including:

[0085] A transformer is created, and a multi-head self-attention layer is added to the transformer to capture the global dependency between different depth points in the logging curve using the multi-head self-attention layer;

[0086] A feedforward network is added for nonlinear transformation of the features of each depth point;

[0087] Residual structure and normalization are used to accelerate training convergence and alleviate gradient vanishing;

[0088] The longitudinal position information of the logging curve is embedded through depth position encoding, so that the geophysical logging large model can perceive the relative positions of different depth points in the stratum;

[0089] The logging curve is subjected to Fourier transform to extract periodic features, and the original time domain features are fused;

[0090] The time domain and frequency domain encoded features are spliced or weighted fused to form multi-scale feature representation, which is input to the transformer for global modeling.

[0091] For ease of description, the above steps will be described in combination.

[0092] The geophysical logging large model adopts a basic network architecture based on Transformer, and excavates the semantic dependency relationship between and among multi-source series data through the self-attention layer in the Transformer architecture. The core task of this part is to design and implement a multi-modal Transformer architecture suitable for the characteristics of geophysical logging data, realize global dependency modeling and deep feature extraction of multi-source logging data by introducing time-frequency domain coding, optimizing self-attention mechanism and network level configuration.

[0093] Specifically, the network infrastructure design of the geophysical logging large model improves the standard Transformer. That is, the multi-head attention mechanism (Multihead Self Attention) is introduced on the architecture of the standard Transformer to capture the global dependency relationship between different depth points in the logging curve. For example, the nonlinear correlation between the resistivity of a certain layer and the porosity of the layers above and below it. The feedforward network (FeedForward Network, FFN) is added to perform nonlinear conversion on the features of each depth point, enhancing the model's expression ability. Residual connection (Residual Connection) and layer normalization (Layer Normalization) can accelerate training convergence and alleviate gradient vanishing. The input data can be serialized, and the multi-source logging data can be converted into equal-length numerical sequences to ensure uniform input format.

[0094] For time-frequency domain coding, integration can be performed, mainly divided into time domain coding, frequency domain coding, and fusion strategy. First, the depth position coding is used to embed the longitudinal position information of the logging curve, so that the model can perceive the relative position of different depth points in the formation (such as the lithology difference between shallow and deep layers). Then, the Fourier transform is performed on the logging curve to extract periodic features (such as frequency domain fluctuations of sand-shale rhythmic layer sequences), and the original time domain features are fused to enhance the model's ability to analyze periodic changes in the formation. Finally, the time and frequency domain coded features are spliced or weighted fused to form multi-scale feature representation, which is input to the subsequent Transformer layer for global modeling.

[0095] The self-attention weight dynamically calculates the correlation strength between different depth points (for example, there is a strong correlation between the acoustic time difference of a certain depth point and the resistivity of the adjacent layer), revealing the deep coupling rules of multi-source logging data. For long sequence logging data (such as kilometer-level depth sequence), local window attention or axial attention is used to reduce computational complexity and adapt to memory limitations. In the multi-head self-attention layer, the features of different logging series are allowed to interact, directly modeling cross-series correlation (such as the collaborative mode of resistivity anomalies and lithology changes).

[0096] A Transformer architecture based on time-frequency domain coding is introduced. Stratigraphic sediments have periodic characteristics (such as rhythmic layers of sandstone and mudstone). By extracting periodic fluctuations (such as the frequency domain components of natural gamma curves) through frequency domain coding and combining them with time-domain depth coding, the model's accuracy in identifying cyclic strata is greatly improved, and its ability to detect abrupt changes in physical properties in thin layers (<0.5 meters) is enhanced.

[0097] During the training of the geophysical logging big model to generate the part of the logging data covered by the mask, the multi-head self-attention layer is used to interact with the features of different logging series in the mask sequence to establish cross-series correlation.

[0098] like Figure 4 As shown, the number N of encoders and decoders can be set according to actual conditions. The geophysical well logging large model training uses masked sequence modeling as the training task. The specific training process is as follows: The input data types of the large model include continuous logging curve features, such as density and porosity; formation attribute features, such as oil content, physical properties, sand content, and lithology; spatial location features, such as formation horizons and spatial coordinates (e.g., the horizontal and vertical coordinates of the well location). Model label types include: classification labels: taking water-flooded layer identification as an example, including binary classification (0 / 1) or multi-class classification (e.g., non-water-flooded layer, slightly water-flooded, heavily water-flooded); regression values: taking physical property parameters as an example, including key formation physical parameters such as porosity and permeability. The above data is processed through the above process and used as input data for training. For the loss function setting, the regression task uses mean squared error (MSE) loss to calculate the difference between predicted and true values ​​(e.g., porosity and resistivity numerical reconstruction). The classification task uses cross-entropy loss (e.g., lithology classification, water-flooded layer identification). The training objective is to reconstruct the masked data in the input based on the unmasked multiseries data, and to recover the original data masked by [MASK] or other numerical values.

[0099] After training, the large geophysical logging model is stored in a model structure parameter format. For model structure storage, the Transformer network's layer configuration (such as the number of encoder layers, attention heads, hidden layer dimensions, etc.) is saved in a configuration file (such as JSON or YAML) to ensure the network architecture is reproducible. For model parameter storage, the trained weight parameters (such as self-attention weights, feedforward network parameters) are stored in binary format (such as PyTorch's .pt file) to support fast loading and deployment.

[0100] S103. Obtain the original logging data of the target well point, and use the trained geophysical logging model to generate the target logging data that is missing from the original logging data.

[0101] After the geophysical logging large model is trained by using the mask sequence, when original logging data of a target scene point is obtained, the trained geophysical logging large model can be used to generate target logging data missing in the original logging data.

[0102] In an embodiment of the present application, generating the target logging data missing in the original logging data by using the trained geophysical logging large model comprises: performing serialization processing on the original logging data to obtain a numerical sequence; and inputting the numerical sequence into the trained geophysical logging large model for data generation processing to obtain the target logging data.

[0103] That is, in order to ensure the uniformity of the input format of the data, after the original logging data is obtained, the original logging data can be subjected to serialization processing to obtain a numerical sequence. The serialization processing process can refer to the processing process of the standard logging data described above. Then, the numerical sequence is input into the trained geophysical logging large model for data generation processing to obtain the target logging data.

[0104] In an embodiment of the present application, generating the target logging data missing in the original logging data by using the trained geophysical logging large model comprises: inputting original logging data missing in a target structured logging data series into the trained geophysical logging large model to generate target logging data corresponding to the target structured logging data series; wherein the target structured logging data series is at least one of a lithology logging series, a resistivity logging series and a porosity logging series, and the target logging data is at least one of the target structured logging data series.

[0105] That is, in actual application, the logging data can be divided into a lithology logging series (natural gamma (GR), natural potential (SP), caliper (CAL)), a resistivity logging series (deep lateral resistivity (RILD), medium lateral resistivity (RILM), shallow lateral resistivity (RILS)), a porosity logging series (neutron (CN), density (DEN), acoustic wave (AC)) and other structured logging data series. When the logging data is generated, at least one of the target structured logging data series missing is generated.

[0106] In an embodiment of the present application, after the original logging data of a target well point is obtained and the trained geophysical logging large model is used to generate target logging data missing in the original logging data, the method further comprises: supplementing the original logging data with the target logging data to obtain to-be-interpreted logging data; fine-tuning the trained geophysical logging large model; and using the fine-tuned geophysical logging large model to interpret the to-be-interpreted logging data.

[0107] For ease of description, the above steps will be described in combination.

[0108] After the target logging data is generated, the original logging data can be supplemented based on the target logging data, so as to obtain the to-be-interpreted logging data without missing logging data.

[0109] Since the trained geophysical logging large model can mine the correlation between different logging data, the interpretation information of the to-be-interpreted logging data is obtained by fine-tuning the geophysical logging large model and then using the fine-tuned geophysical logging large model. That is, the geophysical logging large model outputs the specific meaning of the to-be-interpreted logging data, that is, gives the physical characteristics of the target well point (such as whether watered out).

[0110] By performing step S102, the pre-trained base model can extract general features (such as porosity resistivity correlation patterns). For specific tasks, only 10% of the parameters need to be fine-tuned, which can improve the inference speed while maintaining the accuracy, and adapt to the real-time interpretation requirements underground.

[0111] By applying the method provided in the present application, the standard logging data of the sample well point is obtained, and part of the logging data in the standard logging data is masked using a mask to obtain a mask sequence; the geophysical logging large model with a multi-head self-attention layer is trained using the mask sequence and taking generation of the part of the logging data as a target; the original logging data of the target well point is obtained, and the target logging data missing in the original logging data is generated using the trained geophysical logging large model; wherein in the process of training the geophysical logging large model, the features of different logging series in the mask sequence are interacted using the multi-head self-attention layer, and cross-series correlation is established, so as to generate the part of the logging data.

[0112] In the present application, it is considered that, in the process of measurement, the measured original logging data may have data missing due to various factors. Based on the correlation between different logging data, it is proposed to generate the missing target logging data based on the existing original logging data. Specifically, standard logging data of sample well points can be provided first, and part of the logging data in the standard logging data can be covered by a mask, i.e. a mask sequence with part of the data missing can be obtained. Then, since the multi-head self-attention layer can interact different features and establish a correlation between different types of data, in the present application, the mask sequence is used to train the geophysical logging large model with a multi-head self-attention layer. In the process of training the geophysical logging large model to generate the part of the logging data covered by the mask, the multi-head self-attention layer is used to interact the features of different logging series in the mask sequence and establish cross-series correlation. That is, through training, the geophysical logging large model can learn the correlation between different logging series based on the multi-head self-attention layer therein, so that the geophysical logging large model can generate the missing logging data. After the model training is completed, the trained geophysical logging large model can be used to generate the missing target logging data of the original logging data collected at the target well point. In this way, the missing logging data caused by various interference factors in the data collection process can be made up by generating relevant logging data through the model, and the accuracy of geological research can be further improved.

[0113] Corresponding to the above method embodiment, the present application embodiment also provides a data generation device based on a geophysical large model. The data generation device based on a geophysical large model described below can be mutually corresponding and referred to with the data generation method based on a geophysical large model described above.

[0114] Referring to Figure 5 As shown in the figure, the device comprises the following modules:

[0115] The mask sequence acquisition module 101 is configured to acquire standard logging data of sample well points, and cover part of the logging data in the standard logging data by a mask to obtain a mask sequence.

[0116] The model training module 102 is configured to train a geophysical logging large model with a multi-head self-attention layer by using the mask sequence and aiming to generate the part of the logging data. In the process of training the geophysical logging large model, the multi-head self-attention layer is used to interact the features of different logging series in the mask sequence and establish cross-series correlation, so as to generate the part of the logging data.

[0117] The data generation module 103 is configured to acquire original logging data of a target well point, and generate target logging data missing in the original logging data by using a trained geophysical logging large model.

[0118] By using the device provided in the embodiments of the present application, standard logging data of a sample well point is acquired, and part of the logging data in the standard logging data is masked by using a mask to obtain a mask sequence; the geophysical logging large model with the multi-head self-attention layer is trained by using the mask sequence and aiming to generate the part of the logging data; original logging data of a target well point is acquired, and target logging data missing in the original logging data is generated by using the trained geophysical logging large model; in the process of training the geophysical logging large model, the features of different logging series in the mask sequence are interacted by using the multi-head self-attention layer to establish cross-series association so as to generate the part of the logging data.

[0119] In the present application, it is considered that, in the measurement process, the original logging data measured may be missing due to various factors. There are some association relationships between different logging data, therefore, it is proposed to generate the missing target logging data based on the existing original logging data. Specifically, the standard logging data of a sample well point is provided first, and part of the logging data in the standard logging data is masked by using a mask, that is, the logging data with part of the data missing is obtained as a mask sequence. Then, the geophysical logging large model with the multi-head self-attention layer is trained by using the mask sequence. In the process of training the geophysical logging large model to generate the part of the logging data masked by the mask, the features of different logging series in the mask sequence are interacted by using the multi-head self-attention layer to establish cross-series association. That is, by training, the geophysical logging large model can learn the association between different logging series based on the multi-head self-attention layer therein, so that the geophysical logging large model can generate the missing logging parameters. After the model training is completed, the target logging data missing in the original logging data acquired by the target well point can be generated by using the trained geophysical logging large model. In this way, the logging data can be generated by the model, so as to make up for the missing logging data caused by various interference factors in the data acquisition process, and further improve the accuracy of geological research.

[0120] In an embodiment of the present application, the mask sequence acquisition module is specifically configured to mask the data of a single logging series in the standard logging data in proportion to obtain an internal mask sequence; mask the data of different logging series in the standard logging data in proportion to obtain an interactive mask sequence; superimpose the internal mask sequence and the interactive mask sequence to obtain a comprehensive mask sequence; and determine the internal mask sequence, the interactive mask sequence, or the comprehensive mask sequence as the mask sequence.

[0121] In an embodiment of the present application, the mask sequence acquisition module is specifically configured to discard the mask corresponding to the internal mask sequence if the same data point is masked in both the internal mask sequence and the interactive mask sequence.

[0122] In an embodiment of the present application, the model building module is specifically configured to create a converter, and add a multi-head self-attention layer to the converter to capture global dependency relationships between different depth points in the logging curve by using the multi-head self-attention layer; add a feedforward network for nonlinear conversion of features of each depth point; use a residual structure and normalization to accelerate training convergence and relieve gradient vanishing; embed longitudinal position information of the logging curve through depth position coding to enable the geophysical logging large model to perceive relative positions of different depth points in the stratum; perform Fourier transform on the logging curve to extract periodic features and fuse the periodic features with original time domain features; splice or weightedly fuse the features coded in the time domain and the frequency domain to form multi-scale feature representation and input to the converter for global modeling.

[0123] In an embodiment of the present application, the data generation module is specifically configured to perform sequence processing on the original logging data to obtain a numerical sequence; and input the numerical sequence to the trained geophysical logging large model for data generation processing to obtain the target logging data.

[0124] In an embodiment of the present application, the data generation module is specifically configured to input the original logging data of a missing target structured logging data series to the trained geophysical logging large model to generate target logging data corresponding to the target structured logging data series; wherein the target structured logging data series is at least one of a lithology logging series, a resistivity logging series, and a porosity logging series, and the target logging data is at least one logging data in the target structured logging data series.

[0125] In an embodiment of the present application, the well logging interpretation module is further configured to: acquire original well logging data of a target well point, and after generating target well logging data missing from the original well logging data using the trained geophysical well logging large model, supplement the original well logging data with the target well logging data to obtain to-be-interpreted well logging data; fine-tune the trained geophysical well logging large model; and interpret the to-be-interpreted well logging data using the fine-tuned geophysical well logging large model.

[0126] Corresponding to the above method embodiments, the embodiments of the present application further provide an electronic device. The electronic device described below can be correspondingly referred to the data generation method based on a geophysical large model described above.

[0127] Referring to Figure 6 As shown in the figure, the electronic device includes:

[0128] The memory 332 is configured to store a computer program.

[0129] The processor 322 is configured to execute the computer program to implement the steps of the data generation method based on a geophysical large model in the above method embodiments.

[0130] Specifically, please refer to Figure 7 , Figure 7 A specific structural diagram of an electronic device provided by the present embodiment is shown in the figure. The electronic device can have great differences due to different configurations or performances, and can include one or more processors (central processing units, CPU) (for example, one or more processors) and a memory 332 storing one or more computer programs 342 or data 344. The memory 332 can be temporary storage or persistent storage. The program stored in the memory 332 can include one or more modules (not shown in the figure), each of which can include a series of instruction operations in the data processing device. Further, the processor 322 can be configured to communicate with the memory 332 to execute a series of instruction operations in the memory 332 on the electronic device 301.

[0131] The electronic device 301 can further include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341.

[0132] The steps in the above-described data generation method based on a geophysical large model can be implemented by the structure of the electronic device.

[0133] Corresponding to the above method embodiments, the embodiments of the present application also provide a readable storage medium. The readable storage medium described below can be correspondingly referred to the data generation method based on the geophysical large model described above.

[0134] A readable storage medium, the readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the data generation method based on the geophysical large model of the above method embodiments.

[0135] The readable storage medium can be a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various readable storage media that can store program codes.

[0136] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. For the device disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0137] Those skilled in the art can further realize that the units and algorithm steps of each example described in combination with the embodiments disclosed herein can be realized by electronic hardware, computer software or combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been described in general in the above description. Whether the functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0138] The steps of the method or algorithm described in combination with the embodiments disclosed herein can be directly implemented by hardware, software module executed by a processor, or combination of the two. The software module can be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0139] Finally, it needs to be mentioned that, in this document, relationships such as first and second, and the like, are intended to distinguish one entity or action from another, and do not necessarily require or imply any actual such relationship or order between entities or actions. Also, the term including, or other variations thereof, is intended to cover the non-exclusive inclusion such that a process, method, article, or apparatus that comprises elements not expressly listed is still within the scope of such including process, method, article, or apparatus.

[0140] The principles and implementation manners of the present application are described herein by using specific examples, and the above example descriptions are only used to help understand the method and core idea of the present application; meanwhile, according to the idea of the present application, a person skilled in the art can make changes in specific implementation manners and application ranges; in summary, the content of the present description should not be understood as a limitation on the present application.

Claims

1. A data generation method based on a large geophysical model, characterized in that, include: Obtain standard logging data for sample well points, and use a mask to cover part of the logging data in the standard logging data to obtain a mask sequence; The mask sequence is data obtained by masking a portion of the standard logging data; Using the mask sequence and with the goal of generating the aforementioned partial logging data, a large geophysical logging model with a multi-head self-attention layer is trained. Obtain the raw logging data of the target well point, and use the trained geophysical logging model to generate the target logging data that is missing from the raw logging data; In the process of training the geophysical logging model, the multi-head self-attention layer is used to interact with the features of different logging series in the mask sequence to establish cross-series associations in order to generate the partial logging data. Specifically, a mask sequence is obtained by using a mask to cover a portion of the standard logging data, including: For the data of a single well logging series in the standard well logging data, the data is masked proportionally to obtain an internal mask sequence; 80% of the data is retained: maintaining the integrity of most of the original data and providing contextual information; 10% is replaced with a mask: forcing the model to learn to reconstruct the masked key data points through context; 10% is randomly replaced: simulating random instrument errors; For the data from different logging series in the standard logging data, the data are masked proportionally to obtain an interactive mask sequence; 30% of the data series are randomly selected for combination; 10% of the data are parallel masked: data are synchronously masked at the same depth position of the selected series to simulate multi-instrument synchronization failure or data acquisition blind spots. By superimposing the internal mask sequence and the interactive mask sequence, a composite mask sequence is obtained; The internal mask sequence, the interactive mask sequence, or the composite mask sequence is determined as the mask sequence.

2. The data generation method based on a large geophysical model according to claim 1, characterized in that, The superposition of the inner mask sequence and the interaction mask sequence includes: If the same data point is masked in both the internal mask sequence and the interactive mask sequence, then the mask corresponding to the internal mask sequence is discarded.

3. The data generation method based on a large geophysical model according to claim 1, characterized in that, Constructing a large-scale geophysical logging model with multiple self-attention layers, including: Create a converter and add a multi-head self-attention layer to the converter to capture the global dependencies between different depth points in the logging curve. Add a feedforward network to perform non-linear transformations on the features of each depth point; The residual structure and normalization are used to accelerate training convergence and alleviate gradient vanishing. By embedding the vertical position information of the logging curve through depth position encoding, the geophysical logging big model can perceive the relative position of different depth points in the formation. Fourier transform is performed on the logging curves to extract periodic features, which are then fused with the original time-domain features. The time-domain and frequency-domain encoded features are concatenated or weighted and fused to form a multi-scale feature representation, which is then input into the converter for global modeling.

4. The data generation method based on a large geophysical model according to claim 1, characterized in that, Using a trained geophysical logging model, target logging data missing from the original logging data is generated, including: The original well logging data is serialized to obtain a numerical sequence; The numerical sequence is input into a trained geophysical logging model for data generation processing to obtain the target logging data.

5. The data generation method based on a large geophysical model according to claim 1, characterized in that, Using a trained geophysical logging model, target logging data missing from the original logging data is generated, including: The original logging data of the missing target structured logging data series is input into the trained geophysical logging big model to generate the target logging data corresponding to the target structured logging data series. The target structured logging data series is at least one of the following: lithology logging series, resistivity logging series, and porosity logging series. The target logging data is at least one logging data series from the target structured logging data series.

6. The data generation method based on a large geophysical model according to any one of claims 1 to 5, characterized in that, After acquiring the raw logging data for the target well point and generating the target logging data missing from the raw logging data using a trained geophysical logging model, the process also includes: The original logging data is supplemented using the target logging data to obtain logging data to be interpreted; Fine-tune the trained geophysical logging model; Interpretive information of the well logging data to be interpreted is obtained using a finely tuned geophysical logging big model.

7. A data generation device based on a large geophysical model, characterized in that, include: The mask sequence acquisition module is used to acquire standard logging data of sample well points and use a mask to cover part of the logging data in the standard logging data to obtain a mask sequence; The mask sequence is data obtained by masking a portion of the standard logging data; The model training module is used to train a large geophysical logging model with a multi-head self-attention layer using the mask sequence and with the goal of generating the partial logging data. During the training of the large geophysical logging model, the multi-head self-attention layer is used to interact with the features of different logging series in the mask sequence to establish cross-series associations in order to generate the partial logging data. The data generation module is used to acquire the original logging data of the target well point and generate the target logging data missing from the original logging data using a trained geophysical logging big model. Specifically, the mask sequence acquisition module is used to mask the data of a single logging series in the standard logging data proportionally to obtain an internal mask sequence; 80% data retention: maintaining the integrity of most of the original data and providing contextual information; 10% replacement with mask: forcing the model to learn and reconstruct the masked key data points through context; 10% random replacement: simulating random instrument errors; masking the data of different logging series in the standard logging data proportionally to obtain an interactive mask sequence; 30% data series inter-series: randomly selecting some series combinations; 10% data being parallel masked: synchronously masking data at the same depth position of the selected series to simulate multi-instrument synchronization failure or data acquisition blind spots; superimposing the internal mask sequence and the interactive mask sequence to obtain a comprehensive mask sequence; and determining the internal mask sequence, the interactive mask sequence, or the comprehensive mask sequence as the mask sequence.

8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the data generation method based on a geophysical large model as described in any one of claims 1 to 6.

9. A readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the steps of the data generation method based on a geophysical large model as described in any one of claims 1 to 6.

Images