An improved method for intelligent prediction of a formation model by machine learning and related devices

Abstract

This invention discloses an improved machine learning-based intelligent prediction method for stratigraphic models and related equipment. By fusing depth-aligned high-frequency active seismic source data and low-frequency passive seismic source data, a multi-dimensional complementary source feature vector is constructed, significantly improving the vertical resolution and deep penetration capability of stratigraphic prediction. A genetic algorithm is used to globally search and optimize the network parameters of the convolutional neural network. Combining binary encoding and genetic operations effectively avoids the pitfalls of traditional gradient descent methods, which are prone to getting trapped in local optima, thus improving the globality and stability of model parameter optimization. Through iterative fitness verification and threshold determination, the optimal network parameters are adaptively determined without manual intervention, significantly reducing the cost of hyperparameter tuning while effectively ensuring the generalization performance of the prediction model. The final prediction model can perform fast and accurate stratigraphic prediction based on real-time source data and can be widely applied in the field of data processing technology.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an improved intelligent prediction method and related equipment for geological models using machine learning. Background Technology

[0002] Accurate stratigraphic model prediction is a key technology in seismic exploration, oil and gas resource assessment, and geological disaster prevention. Traditional stratigraphic prediction methods mainly rely on manually extracting seismic attribute features and combining them with statistical models or shallow machine learning models (such as support vector machines and decision trees) for prediction. However, such methods have the following significant drawbacks: First, the manual feature extraction process is highly dependent on expert experience, making it difficult to guarantee the completeness and representativeness of feature design, and easily overlooking key information hidden in the original data, thus limiting prediction accuracy. Second, existing methods usually only utilize a single type of source data (such as active source data or passive source data), failing to fully integrate the complementary advantages of high-frequency active sources in shallow high-resolution detection and low-frequency passive sources in deep penetration, making it difficult to construct a full-depth, high-precision stratigraphic model. Furthermore, the hyperparameter settings of conventional machine learning models (such as traditional neural networks) often rely on manual trial and error or grid search, which is not only computationally expensive and inefficient, but also prone to getting trapped in local optima, making it difficult to obtain a globally optimal network structure and parameter configuration, resulting in insufficient model generalization ability. Furthermore, existing methods lack a dynamic adaptive adjustment mechanism during iterative optimization, resulting in poor adaptability to complex and ever-changing geological environments. Summary of the Invention

[0003] The main objective of this invention is to propose an improved intelligent prediction method, device, electronic device, storage medium, and program product for geological formation models using machine learning, aiming to solve at least one problem in the prior art.

[0004] To achieve the above objectives, one aspect of this invention proposes an improved intelligent prediction method for stratigraphic models using machine learning, the method comprising: Obtain the source dataset; each data sample in the source dataset includes source feature vector data and label data. The source feature vector data includes depth-aligned high-frequency active source data and low-frequency passive source data. Initialize the genetic population and the convolutional neural network; where each individual in the genetic population includes a set of network parameters; The network parameters of each individual in the genetic population are applied to a convolutional neural network, and the prediction error is verified using the seismic source dataset to obtain the fitness value of each individual. If a fitness value is greater than the iteration optimal threshold, the corresponding individual is taken as the optimal individual; otherwise, the network parameters of each individual in the genetic population are binary encoded based on the preset binary encoding length to obtain the first population; wherein, the network parameters of each individual in the first population are binary genes. Based on binary genes, genetic operations are performed on the first population to obtain the second population. Each individual in the second population is then binary-decoded to form a genetic population. Increment the number of iterations and return to the step of applying the network parameters of each individual in the genetic population to the convolutional neural network until the optimal individual is obtained or the number of iterations reaches the maximum number of iterations. If the number of iterations reaches the maximum number of iterations, the individual with the largest fitness value is selected as the optimal individual. The optimal individual network parameters are applied to a convolutional neural network to obtain a prediction model, which is then used to perform real-time geological prediction based on real-time seismic source data.

[0005] In some embodiments, obtaining the seismic source dataset includes the following steps: The original high-frequency active data of the study area was acquired and transformed into a two-dimensional time-frequency matrix with time step and frequency dimensions through short-time Fourier transform. The time-depth transformation of the time-frequency two-dimensional matrix is ​​performed to obtain a two-dimensional feature map of depth step and frequency dimension, and then the high-frequency single-channel feature map is obtained by normalization as high-frequency active source data. The raw low-frequency passive data of the study area and the reference source signal collected by the drilling vibration sensor were obtained, and the reflection energy sequence under different time shifts was obtained by cross-correlation. Based on the reflection energy sequence, the target feature values ​​of each depth segment are extracted by sliding at a preset depth interval and stacked to obtain a two-dimensional matrix of depth step and feature dimension. Then, the low-frequency single-channel feature map is obtained by normalization as low-frequency passive source data. High-frequency active source data and low-frequency passive source data are scaled and aligned to a depth benchmark, and then stitched together to obtain source feature vector data; The measured true values ​​of the study area were obtained as label data; the label data included lithology type, stratigraphic depth and acoustic velocity.

[0006] In some embodiments, the network parameters of each individual in the genetic population are applied to a convolutional neural network, and the prediction error is verified using a seismic source dataset to obtain the fitness value of each individual, including the following steps: The first individual in the genetic population was used as the test individual; The network parameters corresponding to the test individual are applied to the convolutional neural network to obtain the test model; The source feature vector data is used as input data to the test model to predict the stratigraphic prediction results for each data sample in the source dataset. The prediction errors of various stratigraphic features are obtained based on the stratigraphic prediction results and label data statistics corresponding to all data samples. The prediction errors of all stratigraphic features are normalized and weighted and fused, and then the fitness value corresponding to the test individual is obtained by taking the reciprocal. The next individual in the genetic population is used as the test individual. The process is repeated to apply the network parameters corresponding to the test individual to the convolutional neural network until the fitness value of each individual in the genetic population is obtained.

[0007] In some embodiments, the network parameters of each individual in the genetic population are binary encoded based on a preset binary encoding length to obtain a first population, including the following steps: The first individual in the genetic population is used as the coding individual; Use the first type of parameters in the network parameters of the encoded individual as the encoding parameters; Obtain the gene values ​​and extreme values ​​of the encoded parameters; where the extreme values ​​include the maximum and minimum values ​​of the parameters. Based on the parameter extrema and the binary code length, the encoding parameters are encoded into the corresponding binary genes; The expression for the binary gene is:

[0008] In the formula, X represents the gene value of the coding parameter. The binary gene representing the encoded parameters. This represents the minimum value of the encoding parameter. This represents the maximum value of the encoding parameter. Indicates the length of the binary code; The next type of parameter in the network parameters of the encoded individual is used as the encoded parameter. The steps to obtain the gene value and parameter extreme value of the encoded parameter are returned and executed until all types of network parameters are encoded. The next individual in the genetic population is used as the encoded individual. The process is repeated until all individuals in the genetic population are encoded, and the first population is obtained.

[0009] In some embodiments, before the step of binary encoding the network parameters of each individual in the genetic population based on a preset binary encoding length, the method further includes the following steps: All individuals in the genetic population are sorted in descending order of fitness value to obtain an individual sequence. A predetermined number of individuals in the first row of the individual sequence are selected as the first candidate population. Based on the selection probability corresponding to each individual, a predetermined number of individuals are randomly selected from the remaining individuals in the individual sequence as the second candidate population. The genetic population is updated by summing the first and second candidate populations; The preset quantity is half the number of genetic population initialization quantities.

[0010] In some embodiments, when the genetic operation is a crossover operation, the genetic operation is performed on the first population based on binary genes to obtain a second population, including the following steps: Initialize the individual crossover count to 0, and initialize an empty third candidate population; A pair of individuals is randomly selected from the first group as the crossover individuals; the crossover individuals include the first crossover individual and the second crossover individual; A first random number is generated randomly in the first interval. If the first random number does not exceed the preset crossover operation probability, the number of gene crossovers is initialized to 0. In the network parameters of the first crossover individual, a binary gene is randomly selected as the first crossover gene, and a binary code segment is randomly extracted from the first crossover gene based on the preset crossover length as the first crossover segment. In the network parameters of the second crossover individual, a binary gene is randomly selected as the second crossover gene, and a binary code segment is randomly extracted from the second crossover gene based on the crossover length as the second crossover segment. Perform crossover processing on the first and second crossover segments, increment the crossover count by 1, and return to execute the step of randomly selecting a type of binary gene from the network parameters of the first crossover individual as the first crossover gene until the crossover count reaches the first count. Then, add the first and second crossover individuals that have completed the crossover to the third candidate population. Increment the crossover count by 1, and return to the step of randomly selecting a pair of individuals from the first population as crossover individuals, until the crossover count reaches the second number; The third candidate population and the first population are combined into the second population.

[0011] In some embodiments, when the genetic operation is a mutation operation, the genetic operation is performed on the first population based on binary genes to obtain the second population, including the following steps: Initialize the number of individual mutations to 0, and initialize an empty fourth candidate population; One individual is randomly selected from the first population as the mutant individual; A second random number is randomly generated in the second interval. If the second random number does not exceed the preset mutation operation probability, the number of gene mutations is initialized to 0. A binary gene is randomly selected from the network parameters of the mutated individual as the mutated gene; Randomly select one bit of binary code in the mutated gene and invert the information. Increment the gene mutation count by 1. Return to the step of randomly selecting a type of binary gene as the mutated gene in the network parameters of the mutated individual. Continue until the gene mutation count reaches the third number. Add the mutated individual that has completed the mutation to the fourth candidate population. Increment the individual mutation count by 1, and return to the step of randomly selecting an individual from the first population as the mutated individual, until the individual mutation count reaches the fourth time; The fourth candidate population and the first population are combined into the second population.

[0012] In some embodiments, the method further includes the following steps: The system collects the real-time stratigraphic prediction results of each real-time stratigraphic prediction based on real-time seismic source data, as well as the actual stratigraphic characteristics corresponding to the real-time seismic source data. The prediction error is verified based on the real-time stratigraphic prediction results and the actual stratigraphic characteristics to obtain the real-time fitness value of the prediction model. If the real-time fitness value is less than the iterative optimal threshold, the real-time source data will be associated with the real stratigraphic features and added to the source dataset as a data sample. Return to the steps of initializing the genetic population and convolutional neural network to update the prediction model.

[0013] To achieve the above objectives, another aspect of the present invention proposes an improved machine learning-based intelligent prediction device for stratigraphic models, the device comprising: The first module is used to acquire the earthquake source dataset; each data sample in the earthquake source dataset includes earthquake source feature vector data and label data. The earthquake source feature vector data includes depth-aligned high-frequency active earthquake source data and low-frequency passive earthquake source data. The second module is used to initialize the genetic population and the convolutional neural network; each individual in the genetic population includes a set of network parameters; The third module is used to apply the network parameters of each individual in the genetic population to a convolutional neural network, use the seismic source dataset to verify the prediction error, and obtain the fitness value of each individual. The fourth module is used to select an individual as the optimal individual if the fitness value is greater than the iteration optimal threshold; otherwise, it performs binary encoding on the network parameters of each individual in the genetic population based on a preset binary encoding length to obtain the first population; wherein the network parameters of each individual in the first population are binary genes. The fifth module is used to perform genetic operations on the first population based on binary genes to obtain the second population, and to perform binary decoding on each individual in the second population as a genetic population. The sixth module is used to increment the number of iterations and return to execute the operation of the third module until the optimal individual is obtained or the number of iterations reaches the maximum number of iterations. If the number of iterations reaches the maximum number of iterations, the individual with the largest fitness value is selected as the optimal individual. The seventh module is used to apply the network parameters of the optimal individual to a convolutional neural network to obtain a prediction model, and then use the prediction model to perform real-time geological prediction based on real-time seismic source data.

[0014] In some embodiments, the apparatus further includes an eighth module for performing the following operations: All individuals in the genetic population are sorted in descending order of fitness value to obtain an individual sequence. A predetermined number of individuals in the first row of the individual sequence are selected as the first candidate population. Based on the selection probability corresponding to each individual, a predetermined number of individuals are randomly selected from the remaining individuals in the individual sequence as the second candidate population. The genetic population is updated by summing the first and second candidate populations; The preset quantity is half the number of genetic population initialization quantities.

[0015] In some embodiments, the apparatus further includes a ninth module for performing the following operations: The system collects the real-time stratigraphic prediction results of each real-time stratigraphic prediction based on real-time seismic source data, as well as the actual stratigraphic characteristics corresponding to the real-time seismic source data. The prediction error is verified based on the real-time stratigraphic prediction results and the actual stratigraphic characteristics to obtain the real-time fitness value of the prediction model. If the real-time fitness value is less than the iterative optimal threshold, the real-time source data will be associated with the real stratigraphic features and added to the source dataset as a data sample. Return to the second module to perform the operations, and update the prediction model.

[0016] To achieve the above objectives, another aspect of the present invention provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the aforementioned method.

[0017] To achieve the above objectives, another aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0018] To achieve the above objectives, another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0019] The embodiments of the present invention include at least the following beneficial effects: The present invention provides an improved intelligent prediction method, device, electronic device, storage medium, and program product for stratigraphic models using machine learning. This solution involves acquiring a seismic source dataset; wherein each data sample in the seismic source dataset includes seismic source feature vector data and label data, and the seismic source feature vector data includes depth-aligned high-frequency active seismic source data and low-frequency passive seismic source data; initializing a genetic population and a convolutional neural network; wherein each individual in the genetic population includes a set of network parameters; applying the network parameters of each individual in the genetic population to the convolutional neural network, using the seismic source dataset to verify the prediction error, and obtaining the fitness value of each individual; if there exists a fitness value greater than the iterative optimal threshold, the corresponding individual is taken as the optimal individual; otherwise, based on... A preset binary encoding length is used to binary encode the network parameters of each individual in the genetic population, resulting in the first population. The network parameters of each individual in the first population are represented by binary genes. Based on these binary genes, genetic operations are performed on the first population to obtain the second population. Each individual in the second population is then binary-decoded to form the genetic population. The iteration count is incremented, and the process of applying the network parameters of each individual in the genetic population to a convolutional neural network is repeated until the optimal individual is obtained or the maximum number of iterations is reached. If the maximum number of iterations is reached, the individual with the highest fitness value is selected as the optimal individual. The network parameters of the optimal individual are then applied to the convolutional neural network to obtain a prediction model. This prediction model is used to perform real-time seismic prediction based on real-time seismic source data. This invention constructs a multi-dimensional complementary source feature vector by fusing depth-aligned high-frequency active source and low-frequency passive source data, which significantly improves the vertical resolution and deep penetration capability of stratigraphic prediction. Furthermore, this invention employs a genetic algorithm to globally search and optimize the network parameters of a convolutional neural network. Combining binary encoding and genetic operations, it effectively avoids the pitfalls of traditional gradient descent methods, which are prone to getting trapped in local optima, thereby improving the globality and stability of model parameter optimization. Simultaneously, this invention adaptively determines the optimal network parameters through iterative fitness verification and threshold determination, eliminating the need for manual intervention and significantly reducing the cost of hyperparameter tuning while effectively ensuring the generalization performance of the prediction model. The final prediction model can perform fast and accurate stratigraphic prediction based on real-time source data. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an implementation environment for the intelligent prediction method for stratigraphic models using improved machine learning provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the improved machine learning-based intelligent prediction method for stratigraphic models provided in this embodiment of the invention. Figure 3This is a schematic diagram of the overall process of the improved machine learning-based intelligent prediction method for stratigraphic models provided in this embodiment of the invention. Figure 4 This is a schematic diagram of the structure of the improved machine learning-based intelligent prediction device for stratigraphic models provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of the structure of the electronic device provided in an embodiment of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of this invention; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this invention as detailed in the appended claims.

[0022] It is understood that the terms "first," "second," etc., used in this invention may be used to describe various concepts, but unless specifically stated otherwise, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of embodiments of this invention, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words "if" or "when" as used herein may be interpreted as "when," "in response to determination," or "in the event of a determination."

[0023] The terms “at least one,” “multiple,” “each,” “any,” etc., used in this invention, “at least one” includes one, two, or more than two; “multiple” includes two or more than two; “each” refers to each of the corresponding multiple; and “any” refers to any one of the multiple.

[0024] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this invention is for descriptive purposes only and is not intended to limit the invention.

[0025] To facilitate understanding of the technical solution of this invention, the key terms that may be involved in the technical solution of this invention will first be explained: Seismic while drilling (SWP) is a real-time detection technology that combines seismic exploration with drilling engineering. During drilling operations, it uses seismic wave sensors to collect seismic wave signals generated by drill bit rock-breaking vibrations or ground seismic sources to obtain vertical seismic profiles. It can identify formations in real time and predict information such as abnormal high pressure and reservoir distribution within hundreds of meters in front of the drill bit without interrupting drilling operations. It can effectively reduce drilling risks and improve the drilling rate of high-quality reservoirs. It has already been applied in the exploration of complex oil and gas reservoirs in China, such as the Tarim Basin and Yinggehai Basin.

[0026] Stratigraphic model prediction refers to the technology of constructing a three-dimensional stratigraphic structure / parameter model based on multi-source data such as geological exploration, borehole, and well logging, and combining mechanical, statistical, or artificial intelligence algorithms to infer unknown characteristics or evolution trends such as stratigraphic thickness, pore pressure, stress state, and geological hazard risk. It can significantly improve the scientific nature of decision-making in oil and gas exploration, mineral mining, and other scenarios, reduce the risk of engineering accidents, reduce resource waste, and provide quantitative support for geological research such as stratigraphic evolution and earthquake mechanisms.

[0027] Z-score normalization is one of the most commonly used data preprocessing methods. Its core is to perform a linear transformation on the original features so that the transformed features follow a standard normal distribution with a mean of 0 and a standard deviation of 1, thereby eliminating the interference of differences in feature dimensions and value ranges on subsequent algorithms.

[0028] Convolutional Neural Networks (CNNs) are deep learning models specifically designed for processing grid-like data (such as images and time series), inspired by the biological visual system. They significantly reduce the number of parameters and the risk of overfitting through two core characteristics: local perception and weight sharing. Their core structure includes convolutional layers (extracting local features), pooling layers (reducing dimensionality and enhancing robustness), and fully connected layers (outputting classification / regression results). Compared to traditional fully connected networks, CNNs better preserve the spatial structure information of images and are widely used in image classification, object detection, semantic segmentation, and face recognition.

[0029] Genetic Algorithms (GA) are stochastic global search optimization algorithms that mimic the mechanisms of biological evolution in nature. Drawing on Darwinian evolution and Mendelian genetics, they use an iterative process of population initialization, fitness evaluation, selection, crossover, and mutation to select approximate optimal solutions. They do not rely on gradient information from the objective function, are adaptable to complex optimization scenarios such as nonlinear and multimodal problems, and can be applied to machine learning hyperparameter tuning, engineering design parameter optimization, and production scheduling. By adjusting hyperparameters such as population size and crossover probability, global search capability and convergence speed can be balanced.

[0030] In related technologies, how to efficiently and intelligently achieve automatic prediction of stratigraphic models while improving prediction accuracy and robustness has become a technical problem that urgently needs to be solved in this field.

[0031] In view of this, this invention provides an improved machine learning-based intelligent prediction method and related equipment for stratigraphic models. This method involves acquiring a seismic source dataset; each data sample in the dataset includes seismic source feature vector data and label data, with the feature vector data including depth-aligned high-frequency active seismic source data and low-frequency passive seismic source data; initializing a genetic population and a convolutional neural network; each individual in the genetic population includes a set of network parameters; applying the network parameters of each individual in the genetic population to the convolutional neural network; using the seismic source dataset to verify the prediction error and obtain the fitness value of each individual; if a fitness value is greater than the iterative optimal threshold, the corresponding individual is selected as the optimal individual; otherwise, based on a preset binary encoding length... The network parameters of each individual in the genetic population are binary encoded to obtain the first population; the network parameters of each individual in the first population are represented by binary genes. Based on the binary genes, genetic operations are performed on the first population to obtain the second population, and each individual in the second population is binary decoded to form the genetic population. The number of iterations is incremented, and the step of applying the network parameters of each individual in the genetic population to a convolutional neural network is repeated until the optimal individual is obtained or the maximum number of iterations is reached. If the maximum number of iterations is reached, the individual with the highest fitness value is selected as the optimal individual. The network parameters of the optimal individual are applied to the convolutional neural network to obtain the prediction model, and the prediction model is used to perform real-time stratigraphic prediction based on real-time seismic source data. This invention constructs a multi-dimensional complementary source feature vector by fusing depth-aligned high-frequency active source and low-frequency passive source data, which significantly improves the vertical resolution and deep penetration capability of stratigraphic prediction. Furthermore, this invention employs a genetic algorithm to globally search and optimize the network parameters of a convolutional neural network. Combining binary encoding and genetic operations, it effectively avoids the pitfalls of traditional gradient descent methods, which are prone to getting trapped in local optima, thereby improving the globality and stability of model parameter optimization. Simultaneously, this invention adaptively determines the optimal network parameters through iterative fitness verification and threshold determination, eliminating the need for manual intervention and significantly reducing the cost of hyperparameter tuning while effectively ensuring the generalization performance of the prediction model. The final prediction model can perform fast and accurate stratigraphic prediction based on real-time source data.

[0032] It is understood that the improved machine learning-based intelligent prediction method for geological models provided by this invention can be applied to any computer device with data processing and computing capabilities, and this computer device can be various types of terminals or servers. When the computer device in the embodiments is a server, the server is an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms. Optionally, the terminal can be a smartphone, tablet, laptop, or desktop computer, but it is not limited to these.

[0033] like Figure 1 The diagram shown is a schematic representation of an implementation environment provided by an embodiment of the present invention. (Refer to...) Figure 1 The implementation environment includes at least one terminal 102 and a server 101. The terminal 102 and the server 101 can be connected via a network, either wirelessly or via a wired connection, to complete data transmission and exchange.

[0034] Server 101 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and big data and artificial intelligence platforms.

[0035] Additionally, server 101 can also be a node server in a blockchain network. Blockchain is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms.

[0036] Terminal 102 can be a smartphone, tablet computer, laptop computer, desktop computer, smart speaker, smartwatch, etc., but is not limited to these. Terminal 102 and server 101 can be directly or indirectly connected via wired or wireless communication, and this embodiment of the invention does not impose any limitations.

[0037] For example, based on Figure 1The implementation environment shown in this embodiment of the invention provides an improved machine learning-based intelligent prediction method for stratigraphic models. The following description uses the application of this improved machine learning-based intelligent prediction method for stratigraphic models in server 101 as an example. It can be understood that this improved machine learning-based intelligent prediction method for stratigraphic models can also be applied to terminal 102.

[0038] Reference Figure 2 , Figure 2 This is an optional flowchart of the improved machine learning-based intelligent prediction method for stratigraphic models provided in this embodiment of the invention. The executing entity of this improved machine learning-based intelligent prediction method for stratigraphic models can be any of the aforementioned computer devices (including servers or terminals). Figure 2 The method may include, but is not limited to, steps S100 to S700.

[0039] Step S100: Obtain the earthquake source dataset; Each data sample in the source dataset includes source feature vector data and label data. The source feature vector data includes depth-aligned high-frequency active source data and low-frequency passive source data. It should be noted that, in some embodiments, obtaining the seismic source dataset may include the following steps: acquiring the original high-frequency active data of the study area, and converting it into a time-frequency two-dimensional matrix with time step and frequency dimensions through short-time Fourier transform; performing time-depth transformation on the time-frequency two-dimensional matrix to obtain a two-dimensional feature map with depth step and frequency dimensions, and then normalizing it to obtain a high-frequency single-channel feature map as high-frequency active source data; acquiring the original low-frequency passive data of the study area and the reference source signal collected by the drilling vibration sensor, and obtaining the reflection energy sequence at different time shifts through cross-correlation; based on the reflection energy sequence, extracting the target feature values ​​of each depth segment by sliding at a preset depth interval to obtain a two-dimensional matrix with depth step and feature dimension, and then normalizing it to obtain a low-frequency single-channel feature map as low-frequency passive source data; performing scale unification and depth benchmark alignment on the high-frequency active source data and the low-frequency passive source data, and then stitching them together to obtain source feature vector data; acquiring the measured ground truth data of the study area as label data; wherein, the label data includes lithology type, layer depth, and acoustic velocity.

[0040] For example, in some specific implementations, extracting features from high-frequency active source data includes: For high-frequency active data (note that the data here has undergone noise reduction), a Short-Time Fourier Transform (STFT) is used to convert the one-dimensional time-domain signal into a two-dimensional time-frequency matrix. The window length is set to 256 sampling points, and the overlap rate is 50%. The final output two-dimensional matrix has the dimension of "time step × frequency dimension". Considering that the average propagation speed of seismic waves in the strata is 5000 m / s, the two-dimensional matrix of "time step × frequency dimension" can be transformed into a two-dimensional feature map of "depth step × frequency dimension". Finally, Z-score normalization is used to map the values ​​to the [0,1] interval, which can then be used as a high-frequency single-channel feature map. The formula is as follows: Where, Ω represents the timestamp, corresponding to the STFT time axis coordinate; Ω represents the angular frequency (unit: rad / s), corresponding to the STFT frequency axis coordinate. This represents the input data (i.e., high-frequency active data). This is the global time integration variable, the independent variable being integrated, and it only takes effect during the calculation process. This indicates a finite-length window function (the default is the Gaussian window function). This represents the propagation speed of seismic waves, with a value of 5000 m / s.

[0041]

[0042]

[0043] Extracting features from low-frequency passive source data includes: For low-frequency passive data (note that this data has undergone noise reduction processing), the data... Reference source signal acquired by vibration sensors installed near the drill bit By performing cross-correlation calculations, the reflection energy sequences at different time shifts are obtained, as shown in the following formula:

[0044]

[0045] in It refers to time, corresponding to the round-trip time of seismic waves traveling from the drill bit, reflecting off the strata, and returning to the top of the drill string. The velocity of the seismic wave is represented by a value of 5000 m / s. It can be directly mapped to the depth in front of the drill bit based on time. Then, the system slides at 1-meter depth intervals, extracting the following four types of features for each depth segment: peak cross-correlation coefficient, mean signal energy for that depth segment, dominant signal frequency, and root mean square acceleration value. Finally, these are stacked to obtain a two-dimensional matrix of <depth step × feature dimension (4 categories)>, which, after Z-score normalization, can be used as a low-frequency single-channel feature map.

[0046] Finally, by integrating heterogeneous data, a standard input is constructed: We have obtained the feature vector data of high-frequency active data and low-frequency passive data. The two need to be of the same size, aligned with the depth benchmark, and then concatenated into the multi-channel input data of CNN.

[0047] First, high-frequency and low-frequency feature maps are unified to the same size through interpolation / cropping, for example, to (200, 64), which corresponds to 200 depth steps (0–200 m) and 64 feature dimensions. Then, both types of feature maps are set with the drill bit's current position as depth 0 and the direction in front of the drill bit as positive, avoiding depth mismatch (i.e., all depths are positive numbers). Finally, they are concatenated into a dual-channel tensor along the channel dimension, preserving both the shallow, fine reflection features of the high-frequency data and the deep structural features of the low-frequency data, fully adapting to the input requirements of CNN convolution operations and providing multi-scale feature support for subsequent formation prediction.

[0048] The CNN used in this invention belongs to supervised learning. The standardized construction of the input data has already been completed; however, the label data still needs clarification. The label data here refers to the measured ground truth data obtained directly from the formation, including the actual lithology obtained from core sampling, the actual formation depth obtained from logging while drilling (LWD), and sonic velocity data. During training, the loss is calculated using the predicted values ​​output by the CNN (lithology, depth, velocity, etc.) and the aforementioned label values, and the CNN parameters are updated via backpropagation.

[0049] Step S200: Initialize the genetic population and convolutional neural network; Each individual in the genetic population includes a set of network parameters; For example, in some specific implementations, the initialization of the genetic population can be achieved as follows: The population consists of a large number of individuals, each possessing four genes, namely the number of convolutional kernels in a single convolutional layer. Learning rate Dropout rate L2 regularization coefficient That is, each individual represents a CNN prediction model.

[0050] Population initialization involves initializing the genes of each individual, as shown in the following formula. Represents a random value in the range [0,1]; MaxA and MinA represent the maximum and minimum number of convolution kernels; MaxB and MinB represent the maximum and minimum learning rates; MaxC and MinC represent the maximum and minimum dropout rates; MaxD and MinD represent the maximum and minimum L2 regularization coefficients. This represents the i-th individual in the first generation; This refers to the first generation of the population.

[0051]

[0052]

[0053]

[0054]

[0055]

[0056]

[0057] Step S300: Apply the network parameters of each individual in the genetic population to a convolutional neural network, use the seismic source dataset to verify the prediction error, and obtain the fitness value of each individual. It should be noted that, in some embodiments, step S300 may include the following steps: taking the first individual of the genetic population as a test individual; applying the network parameters corresponding to the test individual to a convolutional neural network to configure a test model; inputting the source feature vector data as input data into the test model to predict the stratigraphic prediction results for each data sample in the source dataset; statistically obtaining the prediction errors of various stratigraphic features based on the stratigraphic prediction results and label data corresponding to all data samples; normalizing and weighting the prediction errors of all stratigraphic features, and then obtaining the fitness value corresponding to the test individual by taking the reciprocal; taking the next individual of the genetic population as a test individual, and returning to execute the step of applying the network parameters corresponding to the test individual to a convolutional neural network, until the fitness value of each individual in the genetic population is obtained.

[0058] For example, in some specific implementations, the calculation of the prediction model error can be achieved as follows: Based on the current population (representing the j-th generation population), all individuals have now completed gene decoding, and each individual is iterated through sequentially. Let (i) represent an individual in the j-th generation population. The number of convolutional kernels, learning rate, Dropout rate, and L2 regularization coefficient for each convolutional layer are already determined. These are input into the CNN model, meaning the CNN prediction model representing this individual is ready. For each individual's CNN prediction model, using the current data, the corresponding predicted value is calculated. The predicted data includes two categories: classification labels (lithology) and regression labels (depth, velocity). The errors of these three labels need to be normalized and then weighted and fused to obtain an appropriate evaluation metric, md. The specific formula is as follows:

[0059]

[0060]

[0061]

[0062]

[0063]

[0064] Where md represents the fitness value of an individual, the larger the value, the better the individual performs; , , , representing the weights of the three labels, satisfying =1, with default values ​​of 0.3, 0.4, and 0.3 respectively; ER represents the error rate of lithological attributes, ErrCnt represents the number of incorrect predicted attributes, and AllCnt represents the total number of current samples; HER represents the mean absolute error rate of depth prediction. , HER represents the predicted and actual depths of the formation, respectively. MaxH and MinH represent the maximum and minimum actual depths of the formations in the study area (obtained from research data). The core idea of ​​this formula is to calculate the relative error rate of each data point and then take the average as HER. VER represents the normalized velocity average error. , These represent the predicted and actual velocity values, respectively. MaxV and MinV represent the maximum and minimum actual velocities of the strata in the study area (obtained from research data). The core idea of ​​this formula is to calculate the relative error rate of each data point and then take the average as VER.

[0065] In some preferred implementations, considering that the current population includes individuals who performed well in the previous generation and whose fitness values ​​have already been calculated, these individuals will not have their fitness values ​​calculated again.

[0066] Step S400: If there is an individual with a fitness value greater than the iteration optimal threshold, the corresponding individual is taken as the optimal individual; otherwise, the network parameters of each individual in the genetic population are binary encoded based on the preset binary encoding length to obtain the first population. In the first population, the network parameters of each individual are represented by binary genes. It should be noted that in some embodiments, the network parameters of each individual in the genetic population are binary encoded based on a preset binary encoding length to obtain the first population. This may include the following steps: taking the first individual in the genetic population as the encoded individual; taking the first type of parameter in the network parameters of the encoded individual as the encoded parameter; obtaining the gene value and parameter extreme value of the encoded parameter; wherein, the parameter extreme value includes the maximum value and the minimum value of the parameter; encoding the encoded parameter into the corresponding binary gene based on the parameter extreme value and the binary encoding length; taking the next type of parameter in the network parameters of the encoded individual as the encoded parameter, and returning to execute the step of obtaining the gene value and parameter extreme value of the encoded parameter, until all types of network parameters are encoded; taking the next individual in the genetic population as the encoded individual, and returning to execute the step of taking the first type of parameter in the network parameters of the encoded individual as the encoded parameter, until all individuals in the genetic population are encoded, and summarizing to obtain the first population.

[0067] For example, in some specific implementations, binary encoding of individual genes can be implemented as follows: This invention uses binary encoding for genes because the network hyperparameters are encoded in decimal, which needs to be converted to binary. The reason for using binary encoding is that the value ranges of different genes vary greatly. If decimal encoding were used, genes with smaller value ranges would have an uneven proportion in the search space, leading to reduced algorithm search efficiency and accuracy.

[0068] Here, EL is used to represent the binary encoding length, that is, EL bits are used to represent genes. The value of EL is chosen according to requirements; the larger the EL, the higher the data precision, but the greater the computational load, and vice versa. EL is customized according to requirements; based on experience, EL should not be less than 25, and the default value is 25. The encoding formula is as follows:

[0069] In the formula, X represents the gene value of the coding parameter. The binary gene representing the encoded parameters. This represents the minimum value of the encoding parameter. This represents the maximum value of the encoding parameter. Indicates the length of the binary code; It should be noted that, in some embodiments, before the step of binary encoding the network parameters of each individual in the genetic population based on a preset binary encoding length, the method may further include the following steps: sorting all individuals in the genetic population in descending order of fitness value to obtain an individual sequence; selecting a preset number of individuals from the top row of the individual sequence as a first candidate population; randomly selecting a preset number of individuals from the remaining individuals in the individual sequence as a second candidate population based on the selection probability corresponding to each individual; and updating the genetic population by summing the first candidate population and the second candidate population; wherein, the preset number is half of the initial number of the genetic population.

[0070] For example, in some specific implementations, the population can perform the selection operation as follows: The core objective of this step is to select superior individuals to participate in subsequent crossover and mutation operations. However, for the sake of population diversity (essentially to avoid getting trapped in local optima), individuals with poor performance may still have the potential to enter the next generation. The specific method is as follows: all individuals are sorted in descending order of fitness value. Before sorting... Individuals that are selected directly enter the next generation; for all individuals, the remaining individuals are randomly selected. The individuals are collectively formed into MP selected individuals. Among these MP selected individuals, the same individual may be selected multiple times, because the higher the fitness of an individual, the greater the probability of being selected.

[0071] Note that considering the current population includes new individuals generated by crossover and mutation from the previous generation, the number of individuals may exceed the maximum population size MP, but here we only need to select MP individuals.

[0072] Step S500: Based on binary genes, perform genetic operations on the first population to obtain the second population, and perform binary decoding on each individual in the second population to obtain the genetic population. It should be noted that in some embodiments, when the genetic operation is a crossover operation, based on binary genes, performing genetic operations on the first population to obtain a second population may include the following steps: initializing the individual crossover count to 0, initializing an empty third candidate population; randomly selecting a pair of individuals from the first population as crossover individuals; the crossover individuals include a first crossover individual and a second crossover individual; randomly generating a first random number in a first interval, and if the first random number does not exceed a preset crossover operation probability, initializing the gene crossover count to 0; randomly selecting a type of binary gene from the network parameters of the first crossover individual as the first crossover gene, and randomly extracting a segment of binary code from the first crossover gene based on a preset crossover length as the first crossover segment; in the second crossover individual... A binary gene is randomly selected from the network parameters of the individual as the second crossover gene. Based on the crossover length, a binary code segment is randomly extracted from the second crossover gene as the second crossover fragment. Crossover processing is performed on the first and second crossover fragments, and the gene crossover count is incremented by 1. The process returns to the step of randomly selecting a binary gene from the network parameters of the first crossover individual as the first crossover gene, until the gene crossover count reaches the first count. The first and second crossover individuals that have completed the crossover are added to the third candidate population. The individual crossover count is incremented by 1, and the process returns to the step of randomly selecting a pair of individuals from the first population as crossover individuals, until the individual crossover count reaches the second count. The third candidate population and the first population are combined into the second population.

[0073] For example, in some specific implementations, selecting individuals to perform the crossover operation can be achieved as follows: Given the selected individuals, randomly select a pair of individuals, and then calculate a random number in the range [0,1]. ,if If the probability of the crossover operation does not exceed CR, then the crossover operation should be performed on the individuals; otherwise, the current crossover process should be skipped. A total of [number] operations are performed. This means at most one random draw. Cross-interaction is completed on individuals.

[0074] Assume a pair of individuals < , >It has been clearly determined that cross-operation is required, so we will use < , Taking crossover as an example, each individual has 4 genes, so there are 4 opportunities for crossover. Randomly select a gene from the list, and then from... A gene is randomly selected from the dataset, and these two genes are crossed. This results in a new individual derived from the crossover. and New individuals will join the current population.

[0075]

[0076]

[0077]

[0078]

[0079] The default length of the intersection in this invention is Here we assume genes and genes Intersect. From Two starting points are randomly selected, let's say they are... and This is how the genes were determined. of Fragments and genes of Fragments are exchanged. It's important to note that, considering each pair of individuals completes crossover through four random gene selections, it's possible for a single gene to undergo multiple crossovers. For example, a pair of individuals... , During the first crossover operation, genes are randomly selected. , The crossover was completed, resulting in a new gene. <N , During the second crossover operation, < is randomly selected again. , >Crossing occurs, at which point the genes This situation, involving two crossover operations, is by design.

[0080] It should be noted that in some embodiments, when the genetic operation is a mutation operation, based on binary genes, performing genetic operations on the first population to obtain the second population may include the following steps: initializing the individual mutation count to 0, initializing an empty fourth candidate population; randomly selecting an individual from the first population as a mutated individual; randomly generating a second random number in the second interval, and if the second random number does not exceed the preset mutation operation probability, initializing the gene mutation count to 0; randomly selecting a type of binary gene as a mutated gene from the network parameters of the mutated individual; randomly selecting one bit of binary code in the mutated gene for information inversion processing, incrementing the gene mutation count by 1, and returning to execute the step of randomly selecting a type of binary gene as a mutated gene from the network parameters of the mutated individual until the gene mutation count reaches the third count, adding the mutated individual that has completed the mutation to the fourth candidate population; incrementing the individual mutation count by 1, and returning to execute the step of randomly selecting an individual from the first population as a mutated individual until the individual mutation count reaches the fourth count; and summing the fourth candidate population and the first population into the second population.

[0081] For example, in some specific implementations, selecting individuals to perform mutation operations can be achieved as follows: Given the selected individuals, randomly select one individual and then calculate a random number in the range [0,1]. ,if If the mutation probability MR is not exceeded, the individual undergoes mutation; otherwise, the mutation process is skipped. A total of [number] mutation operations are performed. times, that is, at most one time. The crossover of individuals. One point that needs to be emphasized here is that the individuals selected in step 10 are mutated, and the new individuals generated in step 11 are not mutated.

[0082]

[0083]

[0084] Assuming individual It has been clearly determined that mutation operations are necessary; here we will focus on an individual. Taking mutation as an example, each individual has 4 genes, so there are 4 opportunities for mutation. From Select one gene and perform a mutation operation, repeating this process a total of 4 times. Similar to the crossover operation, a gene can undergo a maximum of 4 mutation operations.

[0085] This invention selects one bit of information for mutation by default, and sets the information bit to its inverse (if it is 0, set it to 1; if it is 1, set it to 0).

[0086] For example, in some specific implementations, binary decoding of individual genes can be achieved as follows: Decoding and encoding are opposite processes and are reversible. Because crossover and mutation operations are based on binary encoding, decoding is required before iterating to the next fitness value calculation after completing the genetic operations. The formula is as follows:

[0087] Step S600: Increment the number of iterations and return to the step of applying the network parameters of each individual in the genetic population to the convolutional neural network until the optimal individual is obtained or the number of iterations reaches the maximum number of iterations. If the number of iterations reaches the maximum number of iterations, the individual with the largest fitness value is taken as the optimal individual. For example, in some specific implementations, steps S300 to S500 are a cyclical iterative process. Each execution represents one iteration of population reproduction, until the population reaches the maximum number of iterations (MGC). When the iteration is complete, it means that based on the current data, the genetic algorithm has found the optimal individual. The optimal CNN prediction model will be used as input to predict geological structures based on real-time data, and the results will be displayed on a 3D model.

[0088] Step S700: Apply the network parameters of the optimal individual to the convolutional neural network to obtain the prediction model, and use the prediction model to perform real-time stratigraphic prediction based on real-time seismic source data. For example, in some specific implementations, the formation is ultimately predicted based on the latest model: After the iteration is completed, the optimal CNN prediction model is ready, and then the latest real-time data is obtained. At this time, the processed real-time data is input into the CNN prediction model to predict the geological structure.

[0089] Furthermore, a 3D visualization digital twin can be used to showcase the formation model (structure, lithology, velocity) information output by the optimized prediction model, combined with downhole tool status data, and input into a lightweight 3D visualization digital twin platform. This platform uses digital twin technology to create a high-precision static model of downhole equipment, which, combined with the latest formation model output by the prediction model, can dynamically and in real-time display the downhole environment and analyze the working efficiency of tools (such as coring tools) based on model predictions.

[0090] It should be noted that, in some embodiments, the method may further include the following steps: collecting the real-time stratigraphic prediction results of the prediction model based on real-time seismic source data each time, as well as the actual stratigraphic features corresponding to the real-time seismic source data; verifying the prediction error based on the real-time stratigraphic prediction results and the actual stratigraphic features to obtain the real-time fitness value of the prediction model; if the real-time fitness value is less than the iterative optimal threshold, associating the real-time seismic source data with the actual stratigraphic features as data samples and adding them to the seismic source dataset; and returning to the steps of initializing the genetic population and convolutional neural network to update the prediction model.

[0091] For example, in some specific implementations, as the amount of real-time data increases, the predictive power of the obtained CNN prediction model may decline. When the prediction results do not meet expectations, it is necessary to update the existing dataset and then find the optimal prediction model again. If the model fitness value... If the value falls below the optimal threshold MD, the population is reconfigured for iteration to update the optimal model.

[0092] To explain in detail the principle of the technical solution of the present invention, the overall process of the present invention will be described below with reference to some specific embodiments. It is easy to understand that the following is an explanation of the technical principle of the present invention and should not be regarded as a limitation of the present invention.

[0093] To address the shortcomings of existing technologies, this invention proposes an improved machine learning-based intelligent prediction method for stratigraphic models. This method integrates multi-source seismic data and uses a genetic algorithm to optimize CNN parameters, achieving real-time stratigraphic prediction and dynamic model updates. Figure 3 As shown, the method of the present invention is implemented as follows: Step 1: Extract features from high-frequency active seismic source data. For high-frequency active data (note that the data here has undergone noise reduction), a Short-Time Fourier Transform (STFT) is used to convert the one-dimensional time-domain signal into a two-dimensional time-frequency matrix. The window length is set to 256 sampling points, and the overlap rate is 50%. The final output two-dimensional matrix has the dimension of "time step × frequency dimension". Considering that the average propagation speed of seismic waves in the strata is 5000 m / s, the two-dimensional matrix of "time step × frequency dimension" can be transformed into a two-dimensional feature map of "depth step × frequency dimension". Finally, Z-score normalization is used to map the values ​​to the [0,1] interval, which can then be used as a high-frequency single-channel feature map. The formula is as follows: Where, Ω represents the timestamp, corresponding to the STFT time axis coordinate; Ω represents the angular frequency (unit: rad / s), corresponding to the STFT frequency axis coordinate. This represents the input data (i.e., high-frequency active data). This is the global time integration variable, the independent variable being integrated, and it only takes effect during the calculation process. This indicates a finite-length window function (the default is the Gaussian window function). This represents the propagation speed of seismic waves, with a value of 5000 m / s.

[0094]

[0095]

[0096] Here's an example: Assume a sliding window length of 256 sampling points and a sampling rate of 1000Hz (each window corresponds to a 0.256s signal): When t=0.01s, that is, when the original signal's t falls within [0.01...],... Perform a Fourier transform on the local signal in the interval [0.128s, 0.01+0.128s] to obtain the frequency distribution of the reflected signal near t=0.01s (calculated to correspond to a depth of 25m). This analysis result corresponds to the feature value at a depth of 25m in the two-dimensional feature map. Proceed to step 2.

[0097] Step 2: Extract features from low-frequency passive source data. For low-frequency passive data (note that the data here has undergone noise reduction processing), extract the data... Reference source signal acquired by vibration sensors installed near the drill bit By performing cross-correlation calculations, the reflection energy sequences at different time shifts are obtained, as shown in the following formula:

[0098]

[0099] in It refers to time, corresponding to the round-trip time of seismic waves traveling from the drill bit, reflecting off the strata, and returning to the top of the drill string. The velocity of the seismic wave is represented by a value of 5000 m / s. It can be directly mapped to the depth in front of the drill bit based on time. Then, the system slides at 1-meter depth intervals, extracting the following four types of features for each depth segment: peak cross-correlation coefficient, mean signal energy for that depth segment, dominant signal frequency, and root mean square acceleration value. Finally, these are stacked to obtain a two-dimensional matrix of <depth step × feature dimension (4 categories)>, which is then Z-score normalized to serve as a low-frequency single-channel feature map. Proceed to step 3.

[0100] Step 3: Merge heterogeneous data to construct a standard input. We have now obtained feature vector data from high-frequency active data and low-frequency passive data. These need to be standardized in size, aligned with the depth reference, and then concatenated to form the multi-channel input data for the CNN.

[0101] First, high-frequency and low-frequency feature maps are unified to the same size through interpolation / cropping, for example, to (200, 64), which corresponds to 200 depth steps (0–200 m) and 64 feature dimensions. Then, both types of feature maps are set with the drill bit's current position as depth 0 and the direction in front of the drill bit as positive, avoiding depth mismatch (i.e., all depths are positive numbers). Finally, they are concatenated into a dual-channel tensor along the channel dimension, preserving both the shallow, fine reflection features of the high-frequency data and the deep structural features of the low-frequency data, fully adapting to the input requirements of CNN convolution operations and providing multi-scale feature support for subsequent formation prediction.

[0102] The CNN used in this invention belongs to supervised learning. The standardized construction of the input data has already been completed; however, the label data still needs to be clarified. The label data here refers to the measured ground truth data obtained directly from the formation, including the actual lithology obtained from core sampling, the actual formation depth obtained from logging while drilling (LWD), and sonic velocity data. During training, the loss is calculated using the predicted values ​​output by the CNN (lithology, depth, velocity, etc.) and the aforementioned label values, and the CNN parameters are updated through backpropagation. Proceed to step 4. Step 4: Initialize key hyperparameters. The parameters are mainly divided into two parts: parameters searched by the genetic algorithm, and key parameters of the CNN network.

[0103] The genetic parameters are as follows: Population size (MP): The maximum number of individuals in the population, set according to requirements, default is 40; Maximum number of iterations (MGC): The maximum number of iterations of the algorithm, set according to requirements, default is 12; Crossover probability (CR): A parameter used by the crossover operator, with a value range of [0,1], set according to requirements, default is 0.8; Mutation probability (MR): A parameter used by the mutation operator, with a value range of [0,1], set according to requirements, default is 0.2; Iteration optimal threshold (MD): When an individual with a value higher than MD is found, the optimal solution is considered to have been found; Binary encoding length (EL): Used for encoding and decoding individual genes, with a value not less than 25, default value is 25.

[0104] The key network parameters for CNNs are as follows: number of convolutional kernels per convolutional layer (to be used later). (Representation), used to control the extraction dimensions of local stratigraphic features; too many dimensions can lead to overfitting; the value range is 16–128; learning rate (used later). (Indicated), used to affect the model's convergence speed; too large a value will cause oscillations and non-convergence, with a range of 1e-5 to 1e-2; Dropout rate (used later) The L2 regularization coefficient (represented by 0.1 to 0.5) is used to reduce the risk of model overfitting and adapt to small sample geological data scenarios. These four parameters (represented by ) are used to constrain the weight distribution and avoid interference from abnormal seismic signals in the prediction results, with values ​​ranging from 1e-6 to 1e-3. These are the key parameters of the CNN prediction model and also the parameters optimized by the genetic algorithm. Proceed to step 5.

[0105] Step 5: Initialize the genetic population. The population consists of a large number of individuals, each with 4 genes, representing the number of convolutional kernels in a single convolutional layer. Learning rate Dropout rate L2 regularization coefficient That is, each individual represents a CNN prediction model.

[0106] Population initialization involves initializing the genes of each individual, as shown in the following formula. Represents a random value in the range [0,1]; MaxA and MinA represent the maximum and minimum number of convolution kernels; MaxB and MinB represent the maximum and minimum learning rates; MaxC and MinC represent the maximum and minimum dropout rates; MaxD and MinD represent the maximum and minimum L2 regularization coefficients. This represents the i-th individual in the first generation; This represents the first generation of the population. Proceed to step 6.

[0107]

[0108]

[0109]

[0110]

[0111]

[0112]

[0113] Step 6: Binary Encoding of Individual Genes. This invention uses binary encoding for genes. Step 5 generates decimal encoded individuals, which need to be converted to binary. The reason for using binary encoding is that the value ranges of each gene vary greatly. If decimal encoding is used, genes with smaller value ranges will have an uneven proportion in the search space, leading to reduced algorithm search efficiency and accuracy.

[0114] Here, EL is used to represent the binary encoding length, that is, EL bits are used to represent genes. The value of EL is chosen according to requirements; the larger the EL, the higher the data precision, but the greater the computational load, and vice versa. EL is customized according to requirements; based on experience, EL should not be less than 25, and the default value is 25. The encoding formula is as follows:

[0115]

[0116]

[0117]

[0118] Where A represents a decimal value. This represents a binary numerical value, and is similar to A otherwise. Here, we take the encoding of the regularization coefficient D as an example, where EL is 25, and the value range of D is 1e-6 to 1e-3. A calculation example is shown below: When D is 1e-6 , When D is 1e-4

[0119] Step 7: Binary Decoding of Individual Genes. Decoding and encoding are opposite processes and are reversible. Because subsequent crossover and mutation operations are based on binary encoding, decoding is required before iterating to the next fitness value calculation after completing the genetic operations. The formula is as follows:

[0120]

[0121]

[0122]

[0123] Step 8: Calculate the prediction model error. Based on the current population. (representing the j-th generation population), all individuals have now completed gene decoding, and each individual is iterated through sequentially. Let (i) represent an individual in the j-th generation population. The number of convolutional kernels, learning rate, Dropout rate, and L2 regularization coefficient for each convolutional layer are already determined. These are input into the CNN model, meaning the CNN prediction model representing this individual is ready. For each individual's CNN prediction model, using the current data, the corresponding predicted value is calculated. The predicted data includes two categories: classification labels (lithology) and regression labels (depth, velocity). The errors of these three labels need to be normalized and then weighted and fused to obtain an appropriate evaluation metric, md. The specific formula is as follows:

[0124]

[0125]

[0126]

[0127]

[0128]

[0129] Where md represents the fitness value of an individual, the larger the value, the better the individual performs; , , , representing the weights of the three labels, satisfying =1, with default values ​​of 0.3, 0.4, and 0.3 respectively; ER represents the error rate of lithological attributes, ErrCnt represents the number of incorrect predicted attributes, and AllCnt represents the total number of current samples; HER represents the mean absolute error rate of depth prediction. , HER represents the predicted and actual depths of the formation, respectively. MaxH and MinH represent the maximum and minimum actual depths of the formations in the study area (obtained from research data). The core idea of ​​this formula is to calculate the relative error rate of each data point and then take the average as HER. VER represents the normalized velocity average error. , These represent the predicted and actual velocity values, respectively. MaxV and MinV represent the maximum and minimum actual velocities of the strata in the study area (obtained from research data). The core idea of ​​this formula is to calculate the relative error rate of each data point and then take the average as VER.

[0130] Note that, considering the current population includes individuals who performed well in the previous generation, their fitness values ​​have already been calculated, and these individuals will not have their fitness values ​​calculated again. Proceed to step 9.

[0131] Step 9: Find the optimal model. Iterate through all individuals in the population and check if the fitness value md of an individual exceeds MD. If it does, then the optimal individual has been found. Proceed to step 15. Otherwise, it indicates that the current population still has room for evolution, so proceed to step 10 and continue iterating.

[0132]

[0133] Step 10: The population performs a selection operation. The core purpose of this step is to select superior individuals to participate in the subsequent crossover and mutation operations. However, for the sake of population diversity (essentially to avoid getting trapped in local optima), individuals with poor performance may still be included in the next generation. The specific method is as follows: All individuals are sorted in descending order of fitness value. Before sorting... Individuals that are selected directly enter the next generation; for all individuals, the remaining individuals are randomly selected. The individuals are collectively formed into MP selected individuals. Among these MP selected individuals, the same individual may be selected multiple times, because the higher the fitness of an individual, the greater the probability of being selected.

[0134] Note that considering the current population includes new individuals generated by crossover and mutation from the previous generation, the number of individuals may exceed the maximum population size (MP). However, we only need to select MP individuals here. Proceed to step 11.

[0135] Step 11: Select individuals to perform the crossover operation. Given the selected individuals, randomly select a pair of individuals and then calculate a random number in the range [0,1]. ,if If the probability of the crossover operation does not exceed CR, then the crossover operation should be performed on the individuals; otherwise, the current crossover process should be skipped. A total of [number] operations are performed. This means at most one random draw. Cross-interaction is completed on individuals.

[0136] Assume a pair of individuals < , >It has been clearly determined that cross-operation is required, so we will use < , Taking crossover as an example, each individual has 4 genes, so there are 4 opportunities for crossover. Randomly select a gene from the list, and then from... A gene is randomly selected from the dataset, and these two genes are crossed. This results in a new individual derived from the crossover. and New individuals will join the current population.

[0137]

[0138]

[0139]

[0140]

[0141] The default length of the intersection in this invention is Here we assume genes and genes Intersect. From Two starting points are randomly selected, let's say they are... and This is how the genes were determined. of Fragments and genes of Fragments are exchanged. It's important to note that, considering each pair of individuals completes crossover through four random gene selections, it's possible for a single gene to undergo multiple crossovers. For example, a pair of individuals... , During the first crossover operation, genes are randomly selected. , The crossover was completed, resulting in a new gene. <N , During the second crossover operation, < is randomly selected again. , >Crossing occurs, at which point the genes This situation, involving two crossover operations, is by design.

[0142] Regarding the process of gene crossover, here is a concrete example. Suppose a gene... and genes After crossover, their genetic information is as follows: 0 1011 0110 0101 1000 11[00 0011] 1 1000 0101 1[111 011]0 1111 1001 If selected randomly of Fragment ([] content) and of If the fragments ([] content) are crossed, the resulting gene information is as follows: 0 1011 0110 0101 1000 11[111 011] 1 1000 0101 1[00 0011]0 1111 1001 Proceed to step 12.

[0143] Step 12: Select an individual to perform the mutation operation. Given the selected individuals, randomly select one individual and then calculate a random number in the range [0,1]. ,if If the mutation probability MR is not exceeded, the individual undergoes mutation; otherwise, the mutation process is skipped. A total of [number] mutation operations are performed. times, that is, at most one time. The crossover of individuals. One point that needs to be emphasized here is that the individuals selected in step 10 are mutated, and the new individuals generated in step 11 are not mutated.

[0144]

[0145]

[0146] Assuming individual It has been clearly determined that mutation operations are necessary; here we will focus on an individual. Taking mutation as an example, each individual has 4 genes, so there are 4 opportunities for mutation. From Select one gene and perform a mutation operation, repeating this process a total of 4 times. Similar to the crossover operation, a gene can undergo a maximum of 4 mutation operations.

[0147] This invention defaults to selecting one bit of information for mutation, which is simply inverted (if it's 0, set it to 1; if it's 1, set it to 0). Here's a concrete example, assuming a gene... The mutation occurs, and its genetic information is as follows, assuming the gene... If the 18th element is randomly selected and mutated (the content in []), then inverted, a new gene is obtained. .

[0148] 0 1011 01[1]0 0101 1000 1100 0011 0 1011 01[0]0 0101 1000 1100 0011 Proceed to step 13.

[0149] Step 13: Obtain the new generation population. By executing steps 10 to 12, one round of population reproduction is completed, resulting in new individuals through crossover and mutation. At this point, the population consists of the individuals selected in step 10 and the new individuals mutated through crossover in steps 11 and 12, and the population size will exceed MP. Excess individuals will be eliminated in the next selection process, thus ensuring that the population size remains at MP. Proceed to step 14.

[0150] Step 14: Satisfy the iteration limit. Steps 7 to 14 are a cyclical iterative process. Each execution represents one iteration of population reproduction, until the population reaches the maximum iteration limit (MGC). If it has not exceeded the limit, proceed to step 7; otherwise, proceed to step 15.

[0151] Step 15: Obtain the optimal prediction model. Reaching this point means that, based on the current data, the genetic algorithm has found the optimal individual. And the optimal CNN prediction model, which will then be used as input to predict the geological structure based on the optimal prediction model, and displayed on a 3D model. Proceed to step 16.

[0152]

[0153] Step 16: Acquire real-time data. Acquire real-time high-frequency active and low-frequency passive source data, and perform deep denoising using wavelet denoising combined with adaptive filtering (data denoising is the core point of other patents, omitted here, and will not be described in detail). Then, extract the features of the high-frequency active and low-frequency passive source data and construct a standard input (refer to steps 1 to 3). Finally, the standard input data is obtained. Proceed to step 17.

[0154] Step 17: Predict stratigraphy based on the latest model. Step 15 prepared the optimal CNN prediction model, and Step 16 obtained the latest real-time data. Now, the processed real-time data is input into the CNN prediction model to predict the stratigraphic structure. Simultaneously, the prediction model error is calculated based on the actual stratigraphic feature data. (Refer to step 8 for the calculation process).

[0155] Step 18: 3D Visualization Digital Twin Display. The formation model (structure, lithology, velocity) information output from the optimized prediction model is combined with downhole tool status data and input into a lightweight 3D visualization digital twin platform. This platform uses digital twin technology to create a high-precision static model of downhole equipment. Combined with the latest formation model output from the prediction model, it can dynamically and in real-time display the downhole environment and analyze the working efficiency of tools (such as coring tools) based on model predictions. Proceed to Step 16.

[0156] Step 19: Update the prediction model. Steps 1 to 15 are a loop of updating the optimal prediction model, and steps 16 to 18 are a loop of predicting stratigraphic structure based on real-time data. As the amount of real-time data increases, the predictive power of the optimal CNN prediction model obtained in step 15 may decline. When the prediction results do not meet expectations, it is necessary to update the existing dataset and then find a new optimal prediction model. If the model fitness value... If the value is below the iterative optimal threshold MD, proceed to step 1 and update the optimal model. Otherwise, proceed to step 16.

[0157] In summary, this invention proposes an improved machine learning-based intelligent prediction method for stratigraphic models. It integrates multi-source seismic data and uses a genetic algorithm to optimize CNN parameters, achieving real-time stratigraphic prediction and dynamic model updates. Compared to existing technologies, the embodiments of this invention offer at least the following beneficial effects: 1) A feature construction method for multi-source heterogeneous seismic source data fusion is proposed. For high-frequency active and low-frequency passive seismic source data, multi-dimensional features are extracted by STFT time-frequency transformation and cross-correlation calculation. After normalization, size unification and depth alignment, a dual-channel CNN input is constructed to provide multi-scale data support for stratigraphic prediction.

[0158] 2) A genetic algorithm-based approach is proposed to optimize key hyperparameters of CNNs. The number of convolutional kernels, learning rate, dropout rate, and L2 regularization coefficient of the CNN are used as optimization variables of the genetic algorithm. The optimal individuals are selected through population initialization, binary encoding and decoding, selection, crossover and mutation iterative screening, which solves the problems of low efficiency and easy getting trapped in local optima in traditional manual parameter tuning.

[0159] 3) A closed-loop iterative real-time prediction and model update mechanism is proposed. A closed-loop process of "model training - real-time prediction - result verification - model update" is constructed. Based on the optimal model, real-time prediction of the downhole formation and 3D visualization are realized. When the prediction accuracy decreases, the dataset is automatically updated and the model is retrained and optimized.

[0160] like Figure 4 As shown, this embodiment of the invention also provides an improved machine learning-based intelligent prediction device 900 for stratigraphic models, which can implement the above-described method. This device may include: The first module 901 is used to acquire the earthquake source dataset; each data sample in the earthquake source dataset includes earthquake source feature vector data and label data, and the earthquake source feature vector data includes depth-aligned high-frequency active earthquake source data and low-frequency passive earthquake source data. The second module 902 is used to initialize the genetic population and the convolutional neural network; wherein, each individual in the genetic population includes a set of network parameters; The third module 903 is used to apply the network parameters of each individual in the genetic population to a convolutional neural network, use the seismic source dataset to verify the prediction error, and obtain the fitness value of each individual. The fourth module 904 is used to select an individual as the optimal individual if the fitness value is greater than the iteration optimal threshold; otherwise, it performs binary encoding on the network parameters of each individual in the genetic population based on a preset binary encoding length to obtain the first population; wherein the network parameters of each individual in the first population are binary genes. The fifth module 905 is used to perform genetic operations on the first population based on binary genes to obtain the second population, and to perform binary decoding on each individual in the second population as a genetic population. Module 6, 906, is used to increment the number of iterations and return to the operation of Module 3 until the optimal individual is obtained or the number of iterations reaches the maximum number of iterations. If the number of iterations reaches the maximum number of iterations, the individual with the largest fitness value is selected as the optimal individual. Module 7, 907, is used to apply the network parameters of the optimal individual to a convolutional neural network to obtain a prediction model, and then use the prediction model to perform real-time geological prediction based on real-time seismic source data.

[0161] In some embodiments, the apparatus further includes an eighth module for performing the following operations: All individuals in the genetic population are sorted in descending order of fitness value to obtain an individual sequence. A predetermined number of individuals in the first row of the individual sequence are selected as the first candidate population. Based on the selection probability corresponding to each individual, a predetermined number of individuals are randomly selected from the remaining individuals in the individual sequence as the second candidate population. The genetic population is updated by summing the first and second candidate populations; The preset quantity is half the number of genetic population initialization quantities.

[0162] In some embodiments, the apparatus further includes a ninth module for performing the following operations: The system collects the real-time stratigraphic prediction results of each real-time stratigraphic prediction based on real-time seismic source data, as well as the actual stratigraphic characteristics corresponding to the real-time seismic source data. The prediction error is verified based on the real-time stratigraphic prediction results and the actual stratigraphic characteristics to obtain the real-time fitness value of the prediction model. If the real-time fitness value is less than the iterative optimal threshold, the real-time source data will be associated with the real stratigraphic features and added to the source dataset as a data sample. Return to the second module to perform the operations, and update the prediction model.

[0163] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0164] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0165] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0166] like Figure 5 As shown, Figure 5 The hardware structure of an electronic device 1000 according to another embodiment is illustrated. The electronic device 1000 includes: The processor 1001 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (aSIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of the present invention. The memory 1002 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RaM). The memory 1002 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 1002 and is called and executed by the processor 1001. Input / output interface 1003 is used to implement information input and output; The communication interface 1004 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 1005 transmits information between various components of the device (e.g., processor 1001, memory 1002, input / output interface 1003, and communication interface 1004); The processor 1001, memory 1002, input / output interface 1003 and communication interface 1004 are connected to each other within the device via bus 1005.

[0167] The electronic device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0168] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0169] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0170] This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0171] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0172] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0173] The improved machine learning-based intelligent prediction method, device, electronic device, storage medium, and program product for stratigraphic models provided in this invention acquire a seismic source dataset. Each data sample in the seismic source dataset includes seismic source feature vector data and label data. The seismic source feature vector data includes depth-aligned high-frequency active seismic source data and low-frequency passive seismic source data. A genetic population and a convolutional neural network are initialized. Each individual in the genetic population includes a set of network parameters. The network parameters of each individual in the genetic population are applied to the convolutional neural network. The prediction error is verified using the seismic source dataset to obtain the fitness value of each individual. If a fitness value is greater than the iterative optimal threshold, the corresponding individual is selected as the optimal individual; otherwise, a preset binary encoding is used. The length of the network parameters of each individual in the genetic population is binary encoded to obtain the first population; where the network parameters of each individual in the first population are represented by binary genes; based on the binary genes, genetic operations are performed on the first population to obtain the second population, and each individual in the second population is binary decoded to form the genetic population; the number of iterations is incremented, and the step of applying the network parameters of each individual in the genetic population to the convolutional neural network is returned until the optimal individual is obtained or the number of iterations reaches the maximum number of iterations. If the number of iterations reaches the maximum number of iterations, the individual with the highest fitness value is selected as the optimal individual; the network parameters of the optimal individual are applied to the convolutional neural network to obtain the prediction model, and the prediction model is used to perform real-time stratigraphic prediction based on real-time seismic source data. This invention constructs a multi-dimensional complementary source feature vector by fusing depth-aligned high-frequency active source and low-frequency passive source data, which significantly improves the vertical resolution and deep penetration capability of stratigraphic prediction. Furthermore, this invention employs a genetic algorithm to globally search and optimize the network parameters of a convolutional neural network. Combining binary encoding and genetic operations, it effectively avoids the pitfalls of traditional gradient descent methods, which are prone to getting trapped in local optima, thereby improving the globality and stability of model parameter optimization. Simultaneously, this invention adaptively determines the optimal network parameters through iterative fitness verification and threshold determination, eliminating the need for manual intervention and significantly reducing the cost of hyperparameter tuning while effectively ensuring the generalization performance of the prediction model. The final prediction model can perform fast and accurate stratigraphic prediction based on real-time source data.

[0174] The embodiments described in this invention are for the purpose of more clearly illustrating the technical solutions of the embodiments of this invention, and do not constitute a limitation on the technical solutions provided by the embodiments of this invention. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this invention are also applicable to similar technical problems.

[0175] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of the present invention, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0176] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0177] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0178] The preferred embodiments of the present invention have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be within the scope of the claims of the present invention.

Claims

1. An improved machine learning-based intelligent prediction method for stratigraphic models, characterized in that, The method includes the following steps: Obtain the earthquake source dataset; wherein, each data sample in the earthquake source dataset includes earthquake source feature vector data and label data, and the earthquake source feature vector data includes depth-aligned high-frequency active earthquake source data and low-frequency passive earthquake source data; Initialize the genetic population and the convolutional neural network; wherein, each individual in the genetic population includes a set of network parameters; The network parameters of each individual in the genetic population are applied to the convolutional neural network, and the prediction error is verified using the seismic source dataset to obtain the fitness value of each individual. If there exists an individual whose fitness value is greater than the iteration optimal threshold, the corresponding individual is taken as the optimal individual; otherwise, the network parameters of each individual in the genetic population are binary encoded based on a preset binary encoding length to obtain a first population; wherein, the network parameters of each individual in the first population are binary genes. Based on the binary gene, genetic operations are performed on the first population to obtain the second population, and each individual in the second population is binary decoded to obtain the genetic population; The iteration count is incremented, and the step of applying the network parameters of each individual in the genetic population to the convolutional neural network is returned until the optimal individual is obtained or the iteration count reaches the maximum iteration count. If the iteration count reaches the maximum iteration count, the individual with the largest fitness value is taken as the optimal individual. The network parameters of the optimal individual are applied to the convolutional neural network to obtain a prediction model, and the prediction model is used to perform real-time stratigraphic prediction based on real-time seismic source data.

2. The method according to claim 1, characterized in that, The process of obtaining the earthquake source dataset includes the following steps: The original high-frequency active data of the study area was acquired and transformed into a two-dimensional time-frequency matrix with time step and frequency dimensions through short-time Fourier transform. The time-frequency two-dimensional matrix is ​​converted from time to depth to obtain a two-dimensional feature map of depth step and frequency dimension, and then normalized to obtain a high-frequency single-channel feature map as the high-frequency active source data. The original low-frequency passive data of the study area and the reference source signal collected by the drilling vibration sensor were obtained, and the reflection energy sequence under different time shifts was obtained by cross-correlation. Based on the reflected energy sequence, the target feature values ​​of each depth segment are extracted by sliding at a preset depth interval and stacked to obtain a two-dimensional matrix of depth step and feature dimension. Then, a low-frequency single-channel feature map is obtained by normalization as the low-frequency passive source data. The high-frequency active source data and the low-frequency passive source data are scaled and aligned with depth references, and then stitched together to obtain the source feature vector data; The measured true values ​​of the study area are obtained as the label data; wherein, the label data includes lithology type, stratigraphic depth and acoustic velocity.

3. The method according to claim 1, characterized in that, The step of applying the network parameters of each individual in the genetic population to the convolutional neural network, using the seismic source dataset to verify the prediction error, and obtaining the fitness value of each individual includes the following steps: The first individual in the genetic population was used as the test individual; The network parameters corresponding to the test individual are applied to the convolutional neural network to obtain the test model; The source feature vector data is used as input data to the test model to predict the stratigraphic prediction result for each data sample in the source dataset. The prediction errors of various stratigraphic features are obtained based on the stratigraphic prediction results corresponding to all the data samples and the statistical analysis of the label data. The prediction errors of all the aforementioned stratigraphic features are normalized and weighted and fused, and then the fitness value corresponding to the test individual is obtained by taking the reciprocal. The next individual in the genetic population is used as the test individual, and the step of applying the network parameters corresponding to the test individual to the convolutional neural network is returned until the fitness value of each individual in the genetic population is obtained.

4. The method according to claim 1, characterized in that, The process of binary encoding the network parameters of each individual in the genetic population based on a preset binary encoding length to obtain a first population includes the following steps: The first individual in the genetic population is taken as the coding individual; The first type of parameters in the network parameters of the encoded individual are used as the encoding parameters; Obtain the gene values ​​and extreme values ​​of the encoded parameters; wherein, the extreme values ​​of the parameters include the maximum value and the minimum value of the parameters; Based on the extreme values ​​of the parameters and the binary encoding length, the encoding parameters are encoded into the corresponding binary genes; The expression for the binary gene is: ， In the formula, X represents the gene value of the coding parameter. The binary gene representing the encoded parameters. This represents the minimum value of the encoding parameter. This represents the maximum value of the encoding parameter. Indicates the length of the binary code; The next type of parameter in the network parameters of the encoded individual is used as the encoded parameter, and the step of obtaining the gene value and parameter extreme value of the encoded parameter is returned to be executed until all types of parameters of the network parameters are encoded. The next individual in the genetic population is used as the encoded individual, and the step of using the first type of network parameters of the encoded individual as the encoding parameters is repeated until all individuals in the genetic population are encoded, and the first population is obtained by summarizing.

5. The method according to claim 1, characterized in that, Before the step of binary encoding the network parameters of each individual in the genetic population based on a preset binary encoding length, the method further includes the following steps: All individuals in the genetic population are sorted in descending order of fitness values ​​to obtain individual sequences. A predetermined number of individuals in the first row of the individual sequence are selected as the first candidate population. Based on the selection probability corresponding to each individual, the predetermined number of individuals are randomly selected from the remaining individuals in the individual sequence as the second candidate population. The genetic population is updated by summing the first candidate population and the second candidate population; The preset quantity is half of the initial quantity of the genetic population.

6. The method according to claim 1, characterized in that, When the genetic operation is a crossover operation, the process of performing genetic operations on the first population based on the binary gene to obtain a second population includes the following steps: Initialize the individual crossover count to 0, and initialize an empty third candidate population; A pair of individuals is randomly selected from the first population as crossover individuals; the crossover individuals include a first crossover individual and a second crossover individual; A first random number is randomly generated in the first interval. If the first random number does not exceed the preset crossover operation probability, the number of gene crossovers is initialized to 0. In the network parameters of the first crossover individual, a type of binary gene is randomly selected as the first crossover gene, and a binary code segment is randomly extracted from the first crossover gene based on a preset crossover length as the first crossover segment. In the network parameters of the second crossover individual, a type of binary gene is randomly selected as the second crossover gene, and a binary code segment is randomly extracted from the second crossover gene based on the crossover length as the second crossover segment. The first crossover segment and the second crossover segment are crossover processed, the gene crossover count is incremented by 1, and the process is returned to the step of randomly selecting a type of binary gene from the network parameters of the first crossover individual as the first crossover gene, until the gene crossover count reaches the first count, and the first crossover individual and the second crossover individual that have completed the crossover are added to the third candidate population. Increment the crossover count of the individuals by 1, and return to the step of randomly selecting a pair of individuals from the first population as crossover individuals, until the crossover count of the individuals reaches the second number; The third candidate population and the first population are combined to form the second population.

7. The method according to claim 1, characterized in that, When the genetic operation is a mutation operation, the process of performing genetic operations on the first population based on the binary gene to obtain the second population includes the following steps: Initialize the number of individual mutations to 0, and initialize an empty fourth candidate population; One individual is randomly selected from the first population as the mutant individual; A second random number is randomly generated in the second interval. If the second random number does not exceed the preset mutation operation probability, the number of gene mutations is initialized to 0. Randomly select one type of binary gene from the network parameters of the mutated individual as the mutated gene; Randomly select one bit of binary code in the mutated gene and perform information inversion processing, increment the gene mutation count by 1, and return to execute the step of randomly selecting a type of binary gene as the mutated gene from the network parameters of the mutated individual until the gene mutation count reaches the third number, and add the mutated individual that has completed the mutation to the fourth candidate population. Increment the number of individual mutations by 1, and return to the step of randomly selecting an individual from the first population as the mutated individual, until the number of individual mutations reaches the fourth time; The fourth candidate population and the first population are combined into the second population.

8. The method according to claim 1, characterized in that, The method further includes the following steps: Collect the real-time stratigraphic prediction results of the prediction model each time it performs real-time stratigraphic prediction based on the real-time seismic source data, as well as the actual stratigraphic characteristics corresponding to the real-time seismic source data; Based on the real-time stratigraphic prediction results and the actual stratigraphic characteristics, the prediction error is verified to obtain the real-time fitness value of the prediction model. If the real-time fitness value is less than the iterative optimal threshold, the real-time seismic source data is associated with the real stratigraphic features and added as a data sample to the seismic source dataset. Return to the steps of initializing the genetic population and convolutional neural network to update the prediction model.

9. An intelligent prediction device for stratigraphic models with improved machine learning, characterized in that, The device includes: The first module is used to acquire the earthquake source dataset; wherein, each data sample in the earthquake source dataset includes earthquake source feature vector data and label data, and the earthquake source feature vector data includes depth-aligned high-frequency active earthquake source data and low-frequency passive earthquake source data; The second module is used to initialize the genetic population and the convolutional neural network; wherein, each individual in the genetic population includes a set of network parameters; The third module is used to apply the network parameters of each individual in the genetic population to the convolutional neural network, use the seismic source dataset to verify the prediction error, and obtain the fitness value of each individual. The fourth module is used to select the corresponding individual as the optimal individual if the fitness value is greater than the iteration optimal threshold; otherwise, it performs binary encoding on the network parameters of each individual in the genetic population based on a preset binary encoding length to obtain a first population; wherein the network parameters of each individual in the first population are binary genes. The fifth module is used to perform genetic operations on the first population based on the binary gene to obtain a second population, and to perform binary decoding on each individual in the second population as the genetic population; The sixth module is used to increment the number of iterations and return to execute the operation of the third module until the optimal individual is obtained or the number of iterations reaches the maximum number of iterations. If the number of iterations reaches the maximum number of iterations, the individual with the largest fitness value is taken as the optimal individual. The seventh module is used to apply the network parameters of the optimal individual to the convolutional neural network to obtain a prediction model, and to use the prediction model to perform real-time geological prediction based on real-time seismic source data.

10. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method according to any one of claims 1 to 8.

Images