Seismic facies identification and reservoir prediction method and system based on convolutional neural network

Abstract

The application discloses a seismic facies identification and reservoir prediction method and system based on a convolutional neural network, relates to the technical field of seismic exploration data processing, and comprises the following steps: obtaining seismic attribute data of a target river channel work area, and generating a direction evolution feature representing the spatial evolution law of a geological body; processing the feature through a pre-trained convolutional neural network, and outputting a preliminary seismic facies classification and reservoir prediction result; comparing the prediction result with measured data, and generating reservoir prediction error data; calculating a direction weighted difference value of the reservoir prediction error and the direction evolution feature, and generating a first adjustment amount on the basis of the direction weighted difference value; and performing directional and intelligent adjustment on preset convolutional layer weight parameters of the convolutional neural network. Through iterative execution of the prediction, error analysis and weight adjustment process based on geological directionality knowledge, the model feature extraction capability is deeply combined with geological laws, and finally, a high-precision seismic facies classification and reservoir prediction result is output.

Description

Technical Field

[0001] This invention relates to the field of seismic exploration data processing technology, and more specifically, to a method and system for seismic facies identification and reservoir prediction based on convolutional neural networks. Background Technology

[0002] In the field of seismic exploration, particularly in the prediction of complex reservoirs with significant spatial orientation, such as channel sand bodies, deep learning methods based on convolutional neural networks have been widely applied. However, existing technologies face a fundamental challenge in practical applications: there is a disconnect between the inherent learning mechanism of the network model and the spatial evolution laws of geological bodies. Convolutional neural networks extract features by sliding across data using fixed convolutional kernels, and their kernel weights are updated during the training phase using a backpropagation algorithm based on a global loss function. While this mechanism is powerful, it is essentially a black-box optimization, lacking explicit embedding and guidance from prior geological knowledge.

[0003] Specifically, in river channel areas, spatial variations in seismic properties have clear geological significance—variations along the river channel direction typically reflect sedimentary continuity or grain size rhythms, while variations perpendicular to the channel may indicate facies transition boundaries or lithological abrupt changes. In existing convolutional neural networks, the initial weights of the convolutional kernels are usually randomly generated and optimized through data-driven processes; the relationships between the weights within the kernel are not assigned any specific geological directional meaning. During training, the network can only passively attempt to learn these patterns from large amounts of data, unable to actively and specifically enhance its sensitivity to feature variations along specific geological directions (such as the river channel axis or vertical axis). This results in the network's ability to identify key geological boundaries and internal structures being highly dependent on the completeness and representativeness of the training data; in areas with scarce data or complex geological conditions, its generalization ability and prediction accuracy will significantly decrease.

[0004] Furthermore, when model predictions encounter errors, existing backpropagation adjustment mechanisms employ a uniform global correction. The error signal propagates backward layer by layer from the output layer, affecting the weight updates of all convolutional kernels evenly, without distinguishing whether the error stems from a misjudgment of river channel direction features or the neglect of vertical features. This adjustment method contradicts the logic of geological interpretation—when geologists discover discrepancies in predictions, they first determine which directional correlation is misunderstood and then make targeted corrections. Current technology lacks a mechanism to correlate spatial prediction errors with directional sensitivity during feature extraction, thus failing to achieve targeted convolutional kernel weight adjustment similar to geological thinking. This makes the model optimization process inefficient and difficult to converge within a limited number of iterations to a geomechanical model that accurately characterizes both river channel direction and boundaries.

[0005] To address the aforementioned issues, there is an urgent need in this field for a technology that can explicitly embed the directional evolution law of geological bodies into the model and can orient and intelligently adjust the weights of the convolutional kernel based on the spatial directional distribution of prediction errors. This would solve the core defects of existing convolutional neural networks in predicting directional reservoirs such as river channels, where the sensitivity of feature extraction is disconnected from geological laws and the optimization of model parameters lacks geological orientation. Summary of the Invention

[0006] To address the aforementioned technical challenges, a method and system for seismic facies identification and reservoir prediction based on convolutional neural networks are provided.

[0007] In a first aspect, the present invention provides a method for seismic facies identification and reservoir prediction based on convolutional neural networks, comprising the following steps:

[0008] S1. Obtain seismic attribute data of several spatial sampling points in the target river channel work area, and calculate the variation amplitude of seismic attribute data between adjacent spatial sampling points along the preset river channel direction and perpendicular to the preset river channel direction, respectively, and calculate the generation direction evolution characteristics.

[0009] The specific generation process of the directional evolution feature in S1 is as follows:

[0010] For each spatial sampling point, within the target river channel work area, multiple consecutive sampling points are selected along the preset river channel direction and perpendicular to the preset river channel direction to form a sequence, denoted as the first direction sequence and the second direction sequence. The absolute forward difference of the seismic attribute values ​​of adjacent points in the first direction sequence and the second direction sequence is calculated, and the arithmetic mean of all absolute forward differences is taken to obtain the mean directional gradient of the spatial sampling point in the preset river channel direction and perpendicular to the preset river channel direction.

[0011] The mean directional gradient along the preset river channel direction is taken as the first component, and the mean directional gradient perpendicular to the preset river channel direction is taken as the second component. The first component and the second component are orthogonal to each other and a two-dimensional vector is constructed with the spatial sampling point as the origin, which is denoted as the directional evolution feature.

[0012] S2. The pre-trained convolutional neural network processes the directional evolution characteristics according to the preset convolutional layer weight parameters and outputs seismic facies classification results and reservoir prediction results.

[0013] S3. Obtain the measured reservoir parameter data of the target river channel area, and calculate the difference between the reservoir prediction results and the corresponding spatial sampling points to generate reservoir prediction error data.

[0014] The specific process for generating reservoir prediction error data is as follows:

[0015] The difference between the measured reservoir parameters and the predicted reservoir results at the spatial sampling points is calculated to obtain the error value of the spatial sampling points.

[0016] Based on the error values ​​of all adjacent spatial sampling points within a preset radius of the spatial sampling point, Kriging interpolation is used to generate continuous reservoir prediction error data within a preset radius, starting from the spatial sampling point.

[0017] S4. Calculate the directional weighted difference between the reservoir prediction error data and the directional evolution characteristic data respectively;

[0018] The specific calculation process for the direction-weighted difference is as follows:

[0019] For each spatial sampling point, the error value of the reservoir prediction error data at that point is multiplied by the magnitude of the two-dimensional vector of the directional evolution feature at that point, and the product is used as the directional weighted difference value of that spatial sampling point.

[0020] S5. When the directional weighted difference is greater than the preset error association threshold, the first adjustment amount is generated according to the directional weighted difference, and the preset convolutional layer weight parameters of the convolutional neural network are adjusted accordingly, and then proceed to S6. When the directional weighted difference is less than or equal to the preset error association threshold, the current seismic facies classification result and reservoir prediction result are directly output.

[0021] The specific process for generating the first adjustment is as follows:

[0022] If the directional weighted difference of the spatial sampling points is greater than the preset error association threshold, then the seismic facies analysis area is delineated with a preset radius centered on it.

[0023] Within each seismic facies analysis region, the average of the orientation-weighted differences of all spatial sampling points is calculated as the local orientation correlation strength of that seismic facies analysis region;

[0024] The weighted difference of the orientation of all spatial sampling points in the target river channel work area is statistically analyzed, and its mean is used as the global orientation correlation strength.

[0025] If the spatial sampling points do not belong to any seismic facies analysis area, the preset convolutional layer weight parameters will not be adjusted.

[0026] If a spatial sampling point belongs to only one seismic facies analysis region, then the first initial adjustment value is calculated using the first function and used as the first adjustment value for that spatial sampling point. The first function is as follows: ,in, This represents the first initial adjustment amount. Indicates the local directional correlation strength. This indicates the preset local adjustment coefficient;

[0027] If a spatial sampling point belongs to two or more seismic facies analysis regions, the ratio of the area of ​​the overlapping region to which the spatial sampling point belongs to the total area of ​​the seismic facies analysis regions to which the spatial sampling point belongs is calculated and denoted as the weight of the corresponding overlapping region. The weighted difference of the directions of all spatial sampling points in the overlapping region is statistically analyzed, and its mean is used as the local directional correlation strength of the overlap. The second initial adjustment amount is calculated through the second function, and the second initial adjustment amount is used as the first adjustment amount for the spatial sampling point. The second function is as follows: ,in, This indicates the second initial adjustment amount. This indicates the preset local intensity weight. The number representing the overlapping region of the seismic facies analysis area to which the spatial sampling point belongs. This represents the total number of overlapping regions within the seismic facies analysis area to which the spatial sampling points belong. Indicates the first The weights of the overlapping regions Indicates the first The local directional correlation strength of overlapping regions.

[0028] The preset local intensity weight It is obtained by calculating the coefficient of variation of the local directional correlation strength of the overlapping regions corresponding to its spatial sampling points.

[0029] The process of adjusting the preset convolutional layer weight parameters is as follows:

[0030] The preset convolutional layer weight parameters are used to characterize the kernel weights of the convolutional kernels of the target convolutional layer for spatial feature extraction in a convolutional neural network;

[0031] Obtain the preset weight mapping relationship between the target convolutional layer and the spatial sampling points. The preset weight mapping relationship defines the mapping relationship between the spatial sampling points and the weights of at least one convolutional kernel in the target convolutional layer.

[0032] The direction-sensitive weights of each convolution kernel are obtained by performing convolution operations with a preset one-dimensional gradient operator along the preset river channel direction and a preset one-dimensional gradient operator perpendicular to the river channel direction, and taking the absolute value.

[0033] The ratio of the first component to the second component in the directional evolution characteristics of the spatial sampling point is denoted as the directional adjustment scaling factor.

[0034] The weight update amount of the generated point is calculated by continuously multiplying the first adjustment amount of the spatial sampling point, the orientation adjustment scaling factor and the orientation sensitive weight of the corresponding convolution kernel;

[0035] For each convolutional kernel, the point weight update amounts of all corresponding spatial sampling points are summarized according to the preset weight mapping relationship. After weighted average calculation, the kernel weight update amount of the convolutional kernel is obtained. The kernel weight update amounts of all convolutional kernels in the target convolutional layer are normalized to obtain the normalized kernel weight update amount of the convolutional kernel, and the kernel weight of the convolutional kernel is updated accordingly.

[0036] The preset error correlation threshold is obtained by calculating the standard deviation of the directional weighted difference of all spatial sampling points in the target river channel work area.

[0037] S6. Re-execute steps S3 to S5 based on the adjusted convolutional neural network until the difference between two consecutive reservoir prediction error data is less than the preset convergence threshold or the number of iterations reaches the maximum threshold; S7. Output the seismic facies classification results and reservoir prediction results based on the adjusted convolutional neural network.

[0038] Secondly, the present invention provides a seismic facies identification and reservoir prediction system based on convolutional neural networks, comprising:

[0039] Data acquisition module: used to acquire seismic attribute data of several spatial sampling points in the target river channel work area, and based on this, calculate the variation amplitude of seismic attribute data between adjacent spatial sampling points along the preset river channel direction and perpendicular to the preset river channel direction, and calculate the generation direction evolution characteristics.

[0040] Preliminary prediction module: Used to process directional evolution features through a pre-trained convolutional neural network based on preset convolutional layer weight parameters, and output seismic facies classification results and reservoir prediction results;

[0041] Error calculation module: used to acquire measured reservoir parameter data of the target river channel area, and calculate the difference between the data and the reservoir prediction results of the corresponding spatial sampling points to generate reservoir prediction error data;

[0042] Orientation weighting module: used to calculate the orientation weighted difference between reservoir prediction error data and orientation evolution characteristic data, respectively;

[0043] Convolutional layer weight adjustment module: When the directional weight difference is greater than the preset error association threshold, it generates a first adjustment amount based on the directional weight difference and adjusts the preset convolutional layer weight parameters of the convolutional neural network accordingly, and enters the weight adjustment iteration module. When the directional weight difference is less than or equal to the preset error association threshold, it directly outputs the current seismic facies classification result and reservoir prediction result.

[0044] Weight adjustment iteration module: used to re-execute the error calculation module to the convolutional layer weight adjustment module based on the adjusted convolutional neural network until the difference between two consecutive reservoir prediction error data is less than the preset convergence threshold or the number of iterations reaches the maximum threshold.

[0045] Output module: Used to output seismic facies classification results and reservoir prediction results based on the adjusted convolutional neural network.

[0046] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for seismic facies identification and reservoir prediction based on convolutional neural networks.

[0047] This invention has at least the following technical effects or advantages:

[0048] 1. Existing convolution kernel weight updates are data-driven black-box optimizations, disconnected from geological understanding. This invention calculates directional evolution characteristics, combines reservoir prediction error data with these characteristics to generate a directional weighted difference, and calculates a first adjustment amount based on this difference, which is ultimately used to adjust the preset convolution layer weight parameters. This process essentially correlates spatial prediction errors with the directional sensitivity of feature extraction, guiding the convolution kernel weights to adjust in a direction that better captures specific directional change patterns in the river channel.

[0049] 2. Traditional backpropagation uses global averaging for error correction. This invention identifies regions with large prediction errors and strong correlations to geological directional changes through directional weighted difference. Based on this, seismic facies analysis regions are divided, and the local and global directional correlation strengths are calculated. A differentiated strategy is used to generate the first adjustment variable based on the location of the spatial sampling points. This mechanism allows model parameter updates to move beyond a one-size-fits-all approach, focusing on local areas with complex geological facies changes and difficult predictions, enabling targeted reinforcement learning and significantly improving model optimization efficiency and prediction accuracy in these key areas.

[0050] 3. In complex regions where multiple seismic facies analysis areas overlap, this invention designs a more refined adjustment strategy. By calculating the weights of overlapping regions and the correlation strength of overlapping local directions, and using preset local intensity weights to balance the local information of overlapping regions with preset empirical coefficients, a second initial adjustment amount is generated through a second function. This dynamic weight allocation mechanism enables the model to comprehensively consider multi-source local direction correlation information when dealing with complex scenarios involving multiple geological anomalies and intersecting seismic facies, avoiding over-adjustment caused by a single local anomaly. This ensures that the model parameter adjustment is both sensitive and robust, improving the reliability and generalization ability of the overall reservoir prediction results. Attached Figure Description

[0051] Figure 1 A flowchart illustrating the seismic facies identification and reservoir prediction method based on convolutional neural networks provided in this embodiment;

[0052] Figure 2A schematic diagram of the structure of a seismic facies identification and reservoir prediction system based on a convolutional neural network provided in an embodiment. Detailed Implementation

[0053] The embodiments address the technical problem in existing technologies where the accuracy of orienting and intelligently adjusting the weights of convolutional kernels based on the spatial directional distribution of prediction errors is insufficient, through a seismic facies identification and reservoir prediction method and system based on convolutional neural networks.

[0054] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0055] This solution addresses the disconnect between feature extraction and geological directionality, and the lack of geological orientation in weight adjustment, in convolutional neural networks (CNNs) for channel reservoir prediction. It achieves targeted optimization of the network model by explicitly embedding geological spatial evolution patterns and establishing a correlation mechanism between prediction errors and directional features. First, starting with seismic attribute data, the variation amplitude between adjacent sampling points is calculated along the preset channel direction and perpendicular to it, generating directional evolution features. These features quantify the spatial orientation of the geological body as prior input knowledge. Subsequently, the pre-trained CNN processes these directional evolution features based on the current weights, outputting seismic facies classification and reservoir prediction results.

[0056] To further guide the network to align with geological patterns, measured reservoir parameters are introduced to calculate prediction errors. These error data are then correlated with directional evolution characteristics using directional weighted differences to assess the intensity of error distribution in different directions. When the directional weighted difference exceeds a preset threshold, it indicates insufficient feature extraction capability in a specific direction. In this case, a first adjustment amount is generated based on the directional weighted difference. This adjustment amount is determined by identifying seismic facies abrupt change benchmarks, dividing the analysis area, and calculating the local and global directional correlation strength, ensuring that the adjustment focuses on key spaces at geological boundaries or lithological abrupt changes.

[0057] The adjustment process for convolutional layer weights combines the first adjustment amount, the direction adjustment scaling factor, and the direction-sensitive weights of the convolutional kernel to generate point weight updates, which are then aggregated into kernel weight updates. Finally, the kernel weights are normalized and adjusted. This step enhances the sensitivity of the convolutional kernel to changes in channel direction or vertical features during the update process, achieving geologically driven parameter optimization. Based on the updated network, the prediction error is recalculated and the above adjustment process is iterated until the error converges or the maximum number of iterations is reached, thereby gradually improving the network's ability to capture directional geological patterns. The final output is the optimized seismic facies classification and reservoir prediction results. By integrating the directional evolution law into the network learning and adjustment mechanism, the accuracy and generalization ability of channel-oriented reservoir prediction are effectively improved.

[0058] like Figure 1The diagram shows a flowchart of a seismic facies identification and reservoir prediction method based on a convolutional neural network, as provided in an embodiment. The method includes the following steps: S1. Obtaining seismic attribute data from several spatial sampling points in the target river channel area, and calculating the variation amplitude of seismic attribute data between adjacent spatial sampling points along and perpendicular to the preset river channel direction, respectively, to generate directional evolution features; S2. Processing the directional evolution features using a pre-trained convolutional neural network according to preset convolutional layer weight parameters, and outputting seismic facies classification results and reservoir prediction results; S3. Obtaining measured reservoir parameter data of the target river channel area, and calculating the difference between this data and the reservoir prediction results of the corresponding spatial sampling points to generate reservoir prediction results. S4. Calculate the directional weighted difference between the reservoir prediction error data and the directional evolution characteristic data respectively; S5. When the directional weighted difference is greater than the preset error association threshold, generate the first adjustment amount based on the directional weighted difference and adjust the preset convolutional layer weight parameters of the convolutional neural network accordingly, and proceed to S6. When the directional weighted difference is less than or equal to the preset error association threshold, directly output the current seismic facies classification result and reservoir prediction result; S6. Re-execute steps S3 to S5 based on the adjusted convolutional neural network until the difference between two consecutive generated reservoir prediction error data is less than the preset convergence threshold or the number of iterations reaches the maximum threshold; S7. Output the seismic facies classification result and reservoir prediction result based on the adjusted convolutional neural network.

[0059] Furthermore, the specific generation process of the directional evolution feature is as follows: For each spatial sampling point, within the target river channel work area, several consecutive sampling points are selected along the preset river channel direction and perpendicular to the preset river channel direction to form a sequence, denoted as the first directional sequence and the second directional sequence. The absolute forward difference of the seismic attribute values ​​of adjacent points in the first directional sequence and the second directional sequence is calculated, and the arithmetic mean of all absolute forward differences is taken to obtain the mean directional gradient of the spatial sampling point in the preset river channel direction and perpendicular to the preset river channel direction. The mean directional gradient along the preset river channel direction is taken as the first component, and the mean directional gradient perpendicular to the preset river channel direction is taken as the second component. The first component and the second component are orthogonal to each other and a two-dimensional vector is constructed with the spatial sampling point as the origin, denoted as the directional evolution feature.

[0060] In this embodiment, this step addresses the problem of how to quantize a continuous seismic attribute field into a stable and noise-resistant directional index. By calculating the mean of the differences within the local sequence, the mean of the directional gradient smooths out random disturbances and robustly reflects the overall trend of attribute changes in that direction.

[0061] For example, for a given point, taking five consecutive points along a predetermined river channel direction, such as 30 degrees east of north, calculating the absolute value of the amplitude difference between adjacent points and averaging them, yields a "mean directional gradient" of 5.2 along the river channel direction. Performing the same operation in its vertical direction, such as 120 degrees east of north, yields a value of 12.8. The "directional evolution characteristic" of this point is then represented by a vector. This vector not only has a magnitude of approximately 13.8, characterizing the overall intensity of change, but its direction also indicates the dominant orientation of the change, providing a precise quantitative basis for subsequent correlation analysis.

[0062] Furthermore, the specific process for generating reservoir prediction error data is as follows: the difference between the measured reservoir parameters and the reservoir prediction results at the spatial sampling points is calculated to obtain the error value of the spatial sampling points; based on the error values ​​of all adjacent spatial sampling points within the preset radius of the spatial sampling points, the Kriging interpolation method is used to generate continuous reservoir prediction error data within the preset radius, starting from the spatial sampling points.

[0063] In this embodiment, this step expands the error information of discrete points into a spatially continuous error field, solving the problem that sparse points alone cannot comprehensively analyze the spatial distribution pattern of errors. Kriging interpolation is a geostatistical method that considers spatial correlation; it can estimate the regional error based on known point error values ​​and a spatial variation function model.

[0064] Furthermore, the specific calculation process of the directional weighted difference is as follows: for each spatial sampling point, the error value of the reservoir prediction error data at that point is multiplied by the magnitude of the two-dimensional vector of the directional evolution feature at that point, and the product is used as the directional weighted difference of that spatial sampling point.

[0065] In this embodiment, this step is a key calculation for achieving "geological guidance adjustment," establishing the product relationship between the prediction error and the intensity of local geological changes. If a location has a large prediction error (i.e., a large reservoir prediction error value) and also experiences drastic changes in seismic attributes (i.e., a large directional evolution characteristic modulus), its "directional weighted difference" will become very large, strongly indicating that this location is an area that the model needs to focus on.

[0066] For example, if point A has an error of 0.05 and its directional evolution feature modulus is 10, then the directional weighted difference value is 0.5; point B also has an error of 0.05, but its modulus is only 2, so the directional weighted difference value is 0.1. Clearly, point A has a stronger correlation signal and will be given higher priority in subsequent weight adjustments. This calculation method effectively distinguishes between global biases and specific errors related to drastic local geological changes.

[0067] Furthermore, the specific generation process of the first adjustment amount is as follows: If the directional weighted difference of the spatial sampling point is greater than the preset error correlation threshold, then the seismic facies analysis area is delineated with the spatial sampling point as the center and a preset radius; within each seismic facies analysis area, the average value of the directional weighted difference of all spatial sampling points is calculated as the local directional correlation strength of that seismic facies analysis area; the directional weighted difference of all spatial sampling points in the target river channel area is statistically analyzed, and its average value is obtained as the global directional correlation strength; if the spatial sampling point does not belong to any seismic facies analysis area, then the preset convolutional layer weight parameters are not adjusted; if the spatial sampling point belongs to only one seismic facies analysis area, then the first initial adjustment amount is calculated through the first function and used as the first adjustment amount of that spatial sampling point. The first function is as follows: ,in, This represents the first initial adjustment amount. Indicates the local directional correlation strength. This represents the preset local adjustment coefficient. If a spatial sampling point belongs to two or more seismic facies analysis regions, the ratio of the area of ​​the overlapping region of the seismic facies analysis regions to which the spatial sampling point belongs to the total area of ​​the seismic facies analysis regions to which the spatial sampling point belongs is calculated and recorded as the weight of the corresponding overlapping region. The directional weighted difference of all spatial sampling points in the overlapping region is statistically analyzed, and its mean is obtained as the overlapping local directional correlation strength. The second initial adjustment amount is calculated through the second function, and the second initial adjustment amount is used as the first adjustment amount for the spatial sampling point. The second function is as follows: ,in, This indicates the second initial adjustment amount. This indicates the preset local intensity weight. The number representing the overlapping region of the seismic facies analysis area to which the spatial sampling point belongs. This represents the total number of overlapping regions within the seismic facies analysis area to which the spatial sampling points belong. Indicates the first The weights of the overlapping regions Indicates the first The local directional correlation strength of overlapping regions.

[0068] In this embodiment, the process addresses the problem of how to finely generate adjustment quantities at complex geological junctions. It divides the space into "stable zones," which are unadjusted, "single anomaly zones," which are adjusted based on local intensity, and "complex overlapping zones," which are adjusted based on weighted local intensity.

[0069] For example, a point located at the intersection of a river boundary and a fault belongs to the overlapping part of two analysis regions. Region A, with a weight of 0.6 for the river boundary, has a local directional correlation strength of 0.8; Region B, with a weight of 0.4 for the fault, has a local directional correlation strength of 0.5. This reflects that the point is affected by both geological factors, but the influence of the river boundary is dominant.

[0070] Furthermore, the specific adjustment process of the preset convolutional layer weight parameters is as follows: the preset convolutional layer weight parameters are used to characterize the kernel weights of the convolutional kernels of the target convolutional layer for spatial feature extraction in the convolutional neural network; the preset weight mapping relationship between the target convolutional layer and the spatial sampling points is obtained, which defines the mapping relationship between the spatial sampling points and the weights of at least one convolutional kernel in the target convolutional layer; by performing convolution operations on each convolutional kernel in the target convolutional layer with a preset one-dimensional gradient operator along the preset river channel direction and a preset one-dimensional gradient operator perpendicular to the river channel direction, and taking the absolute value, the direction-sensitive weights of each convolutional kernel are obtained; the spatial feature extraction process is then performed. The ratio of the first component to the second component in the directional evolution feature of the spatial sampling points is denoted as the directional adjustment scaling factor. The point weight update amount is calculated by continuously multiplying the first adjustment amount of the spatial sampling points, the directional adjustment scaling factor, and the directional sensitive weights of the corresponding convolutional kernels. For each convolutional kernel, the point weight update amounts of all corresponding spatial sampling points are summarized according to the preset weight mapping relationship, and the kernel weight update amount of the convolutional kernel is obtained by weighted averaging. The kernel weight update amounts of all convolutional kernels in the target convolutional layer are normalized to obtain the normalized kernel weight update amount of the convolutional kernel, and the kernel weight of the convolutional kernel is updated accordingly.

[0071] In this embodiment, the orientation-sensitive weight of each convolutional kernel is obtained by performing convolution operations with a preset one-dimensional gradient operator along the preset river channel direction and a preset one-dimensional gradient operator perpendicular to the river channel direction, and taking the absolute value. This operation evaluates the "inherent" responsiveness of each convolutional kernel to gradient features in a specific direction.

[0072] For each convolutional kernel, the point weight updates of all corresponding spatial sampling points are summarized according to a preset weight mapping relationship. A weighted average is then used to calculate the kernel weight update of that convolutional kernel. The kernel weight updates of all convolutional kernels in the target convolutional layer are normalized to obtain the normalized kernel weight update of that kernel, and the kernel weight of that kernel is updated accordingly. This process summarizes the geologically semantic update suggestions of the corresponding spatial points into specific adjustment instructions for the weights of each convolutional kernel, thus achieving a closed loop from macroscopic geological error correlation to microscopic model parameter adjustment. The final effect is that the convolutional kernel weights in the network are guided to become more sensitive to important geological directional patterns in the target work area.

[0073] Furthermore, the preset mutation threshold is obtained by calculating the standard deviation of the directional weighted difference of all spatial sampling points in the target river channel area.

[0074] In this embodiment, when the influence intensity differences between overlapping regions are large, the coefficient of variation is high, the preset local intensity weight increases, and the result of the second function depends more on the weighted local information of each overlapping region; when the differences are small, the preset local intensity weight decreases, and the result of the second function tends to favor the preset local adjustment coefficient. This dynamic mechanism makes the adjustment strategy in complex regions more flexible and reasonable.

[0075] Furthermore, the preset local intensity weight is obtained by calculating the coefficient of variation of the overlapping local directional correlation intensity of each overlapping region corresponding to its spatial sampling point.

[0076] Figure 2 This is a schematic diagram of the structure of a seismic facies identification and reservoir prediction system based on a convolutional neural network provided in an embodiment. The system includes: a data acquisition module for acquiring seismic attribute data from several spatial sampling points in the target river channel area, and calculating the variation amplitude of seismic attribute data between adjacent spatial sampling points along and perpendicular to the preset river channel direction, respectively, to generate directional evolution features; a preliminary prediction module for processing the directional evolution features using a pre-trained convolutional neural network based on preset convolutional layer weight parameters, and outputting seismic facies classification results and reservoir prediction results; an error calculation module for acquiring measured reservoir parameter data of the target river channel area, calculating the difference between this data and the reservoir prediction results of the corresponding spatial sampling points, and generating reservoir prediction error data; and a directional weighting module. The system is used to calculate the directional weighted difference between reservoir prediction error data and directional evolution feature data, respectively. The convolutional layer weight adjustment module generates a first adjustment value based on the directional weighted difference when it exceeds a preset error correlation threshold. This adjustment value is then used to adjust the preset convolutional layer weight parameters of the convolutional neural network and proceed to the weight adjustment iteration module. When the directional weighted difference is less than or equal to the preset error correlation threshold, the current seismic facies classification result and reservoir prediction result are directly output. The weight adjustment iteration module re-executes the error calculation module and the convolutional layer weight adjustment module based on the adjusted convolutional neural network until the difference between two consecutively generated reservoir prediction error data is less than a preset convergence threshold or the number of iterations reaches the maximum threshold. The output module outputs the seismic facies classification result and reservoir prediction result based on the adjusted convolutional neural network.

[0077] The embodiments also provide a computer-readable storage medium for storing a computer program that, when executed by a processor, implements a seismic facies identification and reservoir prediction method based on a convolutional neural network.

[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for seismic facies identification and reservoir prediction based on convolutional neural networks, characterized in that, Includes the following steps: S1. Obtain seismic attribute data of spatial sampling points in the target river channel work area, and calculate the variation amplitude of seismic attribute data between adjacent spatial sampling points along the preset river channel direction and perpendicular to the preset river channel direction, respectively, and calculate the generation direction evolution characteristics. S2. The pre-trained convolutional neural network processes the directional evolution characteristics according to the preset convolutional layer weight parameters and outputs seismic facies classification results and reservoir prediction results. S3. Obtain the measured reservoir parameter data of the target river channel area, and calculate the difference between the reservoir prediction results and the corresponding spatial sampling points to generate reservoir prediction error data. S4. Calculate the directional weighted difference between the reservoir prediction error data and the directional evolution characteristic data respectively; S5. When the directional weighted difference is greater than the preset error association threshold, the first adjustment amount is generated according to the directional weighted difference, and the preset convolutional layer weight parameters of the convolutional neural network are adjusted accordingly, and then proceed to S6. When the directional weighted difference is less than or equal to the preset error association threshold, the current seismic facies classification result and reservoir prediction result are directly output. S6. Based on the adjusted convolutional neural network, re-execute steps S3 to S5 until the difference between two consecutive reservoir prediction error data is less than the preset convergence threshold or the number of iterations reaches the maximum threshold. S7. Output seismic facies classification results and reservoir prediction results based on the adjusted convolutional neural network.

2. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 1, characterized in that, The specific generation process of the directional evolution feature in S1 is as follows: For each spatial sampling point, within the target river channel work area, multiple consecutive sampling points are selected along the preset river channel direction and perpendicular to the preset river channel direction to form a sequence, denoted as the first direction sequence and the second direction sequence. The absolute forward difference of the seismic attribute values ​​of adjacent points in the first direction sequence and the second direction sequence is calculated, and the arithmetic mean of all absolute forward differences is taken to obtain the mean directional gradient of the spatial sampling point in the preset river channel direction and perpendicular to the preset river channel direction. The mean directional gradient along the preset river channel direction is taken as the first component, and the mean directional gradient perpendicular to the preset river channel direction is taken as the second component. The first component and the second component are orthogonal to each other and a two-dimensional vector is constructed with the spatial sampling point as the origin, which is denoted as the directional evolution feature.

3. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 1, characterized in that, The specific process for generating reservoir prediction error data in S3 is as follows: The difference between the measured reservoir parameters and the predicted reservoir results at the spatial sampling points is calculated to obtain the error value of the spatial sampling points. Based on the error values ​​of all adjacent spatial sampling points within a preset radius of the spatial sampling point, Kriging interpolation is used to generate continuous reservoir prediction error data within a preset radius, starting from the spatial sampling point.

4. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 1, characterized in that, The specific calculation process for the direction-weighted difference in S4 is as follows: For each spatial sampling point, the error value of the reservoir prediction error data at that point is multiplied by the magnitude of the two-dimensional vector of the directional evolution feature at that point, and the product is used as the directional weighted difference value of that spatial sampling point.

5. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 1, characterized in that, The specific process for generating the first adjustment in S5 is as follows: If the directional weighted difference of the spatial sampling points is greater than the preset error association threshold, then the seismic facies analysis area is delineated with a preset radius centered on it. Within each seismic facies analysis region, the average of the orientation-weighted differences of all spatial sampling points is calculated as the local orientation correlation strength of that seismic facies analysis region; The weighted difference of the orientation of all spatial sampling points in the target river channel work area is statistically analyzed, and its mean is used as the global orientation correlation strength. If the spatial sampling points do not belong to any seismic facies analysis area, the preset convolutional layer weight parameters will not be adjusted. If a spatial sampling point belongs to only one seismic facies analysis region, then the first initial adjustment value is calculated using the first function and used as the first adjustment value for that spatial sampling point. The first function is as follows: ,in, This represents the first initial adjustment amount. Indicates the local directional correlation strength. This indicates the preset local adjustment coefficient; If a spatial sampling point belongs to two or more seismic facies analysis regions, the ratio of the area of ​​the overlapping region to which the spatial sampling point belongs to the total area of ​​the seismic facies analysis regions to which the spatial sampling point belongs is calculated and denoted as the weight of the corresponding overlapping region. The weighted difference of the directions of all spatial sampling points in the overlapping region is statistically analyzed, and its mean is used as the local directional correlation strength of the overlap. The second initial adjustment amount is calculated through the second function, and the second initial adjustment amount is used as the first adjustment amount for the spatial sampling point. The second function is as follows: ,in, This indicates the second initial adjustment amount. This indicates the preset local intensity weight. The number representing the overlapping region of the seismic facies analysis area to which the spatial sampling point belongs. This represents the total number of overlapping regions within the seismic facies analysis area to which the spatial sampling points belong. Indicates the first The weights of the overlapping regions Indicates the first The local directional correlation strength of overlapping regions.

6. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 5, characterized in that, The preset local intensity weight It is obtained by calculating the coefficient of variation of the local directional correlation strength of the overlapping regions corresponding to its spatial sampling points.

7. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 1, characterized in that, The process of adjusting the preset convolutional layer weight parameters in S5 is as follows: The preset convolutional layer weight parameters are used to characterize the kernel weights of the convolutional kernels of the target convolutional layer for spatial feature extraction in a convolutional neural network; Obtain the preset weight mapping relationship between the target convolutional layer and the spatial sampling points. The preset weight mapping relationship defines the mapping relationship between the spatial sampling points and the weights of at least one convolutional kernel in the target convolutional layer. The direction-sensitive weights of each convolution kernel are obtained by performing convolution operations with a preset one-dimensional gradient operator along the preset river channel direction and a preset one-dimensional gradient operator perpendicular to the river channel direction, and taking the absolute value. The ratio of the first component to the second component in the directional evolution characteristics of the spatial sampling point is denoted as the directional adjustment scaling factor. The weight update amount of the generated point is calculated by continuously multiplying the first adjustment amount of the spatial sampling point, the orientation adjustment scaling factor and the orientation sensitive weight of the corresponding convolution kernel; For each convolutional kernel, the point weight update amounts of all corresponding spatial sampling points are summarized according to the preset weight mapping relationship. After weighted average calculation, the kernel weight update amount of the convolutional kernel is obtained. The kernel weight update amounts of all convolutional kernels in the target convolutional layer are normalized to obtain the normalized kernel weight update amount of the convolutional kernel, and the kernel weight of the convolutional kernel is updated accordingly.

8. The seismic facies identification and reservoir prediction method based on convolutional neural networks according to claim 1, characterized in that, The preset error correlation threshold in S5 is obtained by calculating the standard deviation of the directional weighted difference of all spatial sampling points in the target river channel work area.

9. A seismic facies identification and reservoir prediction system based on convolutional neural networks, characterized in that, The system for implementing the seismic facies identification and reservoir prediction method based on convolutional neural networks as described in any one of claims 8, wherein the system comprises: Data acquisition module: used to acquire seismic attribute data of spatial sampling points in the target river channel work area, and based on this, calculate the variation amplitude of seismic attribute data between adjacent spatial sampling points along the preset river channel direction and perpendicular to the preset river channel direction, and calculate the generation direction evolution characteristics. Preliminary prediction module: Used to process directional evolution features through a pre-trained convolutional neural network based on preset convolutional layer weight parameters, and output seismic facies classification results and reservoir prediction results; Error calculation module: used to acquire measured reservoir parameter data of the target river channel area, and calculate the difference between the data and the reservoir prediction results of the corresponding spatial sampling points to generate reservoir prediction error data; Orientation weighting module: used to calculate the orientation weighted difference between reservoir prediction error data and orientation evolution characteristic data, respectively; Convolutional layer weight adjustment module: When the directional weight difference is greater than the preset error association threshold, it generates a first adjustment amount based on the directional weight difference and adjusts the preset convolutional layer weight parameters of the convolutional neural network accordingly, and enters the weight adjustment iteration module. When the directional weight difference is less than or equal to the preset error association threshold, it directly outputs the current seismic facies classification result and reservoir prediction result. Weight adjustment iteration module: used to re-execute the error calculation module to the convolutional layer weight adjustment module based on the adjusted convolutional neural network until the difference between two consecutive reservoir prediction error data is less than the preset convergence threshold or the number of iterations reaches the maximum threshold. Output module: Used to output seismic facies classification results and reservoir prediction results based on the adjusted convolutional neural network.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the seismic facies identification and reservoir prediction method based on any one of claims 1-8.

Images