Intelligent three-dimensional wrinkle modeling method based on recurrent neural network
Through the intelligent modeling method based on recurrent neural network, the problems of accurate simulation and low efficiency in three-dimensional fold modeling are solved, efficient and accurate fold model generation is achieved, and the development of geological scientific research and engineering applications are promoted.
Patent Information
- Application Number
- CN202411616597.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-13
- Publication Date
- 2025-02-28
AI Technical Summary
The existing three-dimensional fold modeling technology is difficult to accurately simulate the development process, deformation mechanism and morphological changes of folds at different scales. The modeling process is time-consuming and labor-intensive, and lacks automated and intelligent solutions.
Using an intelligent modeling method based on recurrent neural network (LSTM), we collect and preprocess geological data, use multi-point geological statistics simulation DS algorithm to generate training data, stack multiple LSTMs to form coding and prediction networks, generate diverse and multi-scale geological fold models, and predict three-dimensional fold data through two-dimensional sections.
Efficient extraction and transformation of geological fold laws are achieved, and diverse and multi-scale fold models are quickly generated, which improves modeling efficiency and accuracy, and reduces manual intervention and calculation costs.
Smart Images
Figure CN119540484A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of geological statistical simulation and intelligent modeling, and in particular to a three-dimensional wrinkle intelligent modeling method based on a recurrent neural network. Background Art
[0002] As an important branch of the deep integration of geological science and information technology, three-dimensional fold modeling has become a key means to reveal deep geological structures and resource distribution in recent years. The core of this technology is to reconstruct the complex underground geological fold structure in three dimensions through digital technology, so as to achieve quantitative and qualitative analysis of crustal movement, mineral occurrence and potential geological disasters. Researchers use a variety of information sources such as high-precision measurement data, remote sensing images and drilling data, combined with geomechanical principles and statistical methods, to strive to construct a three-dimensional fold model that is both in line with natural laws and practical, in order to meet the needs of multiple fields such as resource exploration, engineering design and environmental assessment.
[0003] In view of the nonlinearity, uncertainty and high complexity of geological fold phenomena, researchers have long been trying to find effective ways to reveal and quantify these complex geological laws. In terms of three-dimensional fold modeling, some breakthroughs have been made internationally, such as the use of multi-point geostatistical simulation technology and DS algorithms to try to capture the diversity and variability of folds. However, there are still huge challenges in converting field measured data, geological profile information and seismic data into accurate three-dimensional models. How to accurately simulate the development process of folds, deformation mechanisms and morphological changes at different scales is still a key issue that needs to be solved.
[0004] Faced with the huge amount of geological data and complex geological structure, improving the efficiency of three-dimensional fold modeling has become one of the key points of scientific research. Current modeling work often requires a lot of manual intervention and complicated calculation processes, especially when dealing with large-scale areas or complex tectonic areas, which is time-consuming and labor-intensive. Although some scholars and teams have made significant progress with the help of advanced algorithms and high-performance computing platforms, such as using parallel computing technology and machine learning algorithms to accelerate the modeling process, there is still much room for improvement in how to achieve automated and intelligent modeling, especially in real-time updating and dynamic adjustment of model parameters. Summary of the invention
[0005] In order to solve the problem of how to accurately simulate the development process, deformation mechanism and morphological changes of wrinkles at different scales, and how to achieve automatic and intelligent modeling, especially in real-time updating and dynamic adjustment of model parameters, the present invention provides a three-dimensional wrinkle intelligent modeling method based on a recurrent neural network, which mainly includes:
[0006] S1: Collect geological data and preprocess and clean them;
[0007] S2: Analyze and process the geological data after the S1 operation based on the DS algorithm to obtain fold data as training data. Multiple training data constitute a training library;
[0008] S3: stack multiple LSTMs to form an encoding and prediction network, generate a geological fold model, and train the geological fold model using training data;
[0009] S4: Based on the two-dimensional section of the geological image actually obtained, the three-dimensional fold data is predicted using the trained geological fold model.
[0010] A storage device stores instructions and data for implementing a three-dimensional wrinkle intelligent modeling method based on a recurrent neural network.
[0011] A three-dimensional wrinkle intelligent modeling device based on a recurrent neural network includes: a processor and a storage device; the processor loads and executes instructions and data in the storage device to implement a three-dimensional wrinkle intelligent modeling method based on a recurrent neural network.
[0012] The beneficial effects brought by the technical solution provided by the present invention are:
[0013] (1) The present invention uses multi-point geostatistics to simulate the DS algorithm to construct a pattern training library that fully represents the geological semantics of folds. The statistical method is applied to geological data analysis, thereby achieving effective extraction and transformation of geological fold laws.
[0014] (2) The present invention adopts an intelligent modeling method based on LSTM to quickly generate diverse and multi-scale fold models. Compared with traditional modeling methods, this method can better capture the complexity and diversity of geological folds and has higher modeling efficiency and accuracy.
[0015] (3) The present invention can provide an efficient and accurate fold model for various geological science research and engineering applications, and promote the development and application of the geological field. BRIEF DESCRIPTION OF THE DRAWINGS
[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:
[0017] Figure 1 It is a flow chart of a three-dimensional wrinkle intelligent modeling method based on a recurrent neural network in an embodiment of the present invention;
[0018] Figure 2 It is a flow chart for generating a geological fold model in the present invention;
[0019] Figure 3 It is a schematic diagram of a method for constructing a wrinkle pattern training library of the DS algorithm in the present invention;
[0020] Figure 4 It is a schematic diagram of a method for rapidly constructing a multi-scale wrinkle model based on RNN in the present invention;
[0021] Figure 5 The model effect of the experimental case in the present invention is demonstrated as follows: (a) is the input of the geological fold model, a two-dimensional section; (b) is the actual three-dimensional fold modeling where the two-dimensional section is located; (c) is the three-dimensional fold modeling generated by the geological fold model prediction;
[0022] Figure 6 It is a schematic diagram of the operation of the hardware device in the embodiment of the present invention. DETAILED DESCRIPTION
[0023] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.
[0024] Example 1
[0025] Please refer to Figure 1 , Figure 1 : is a flow chart of a three-dimensional wrinkle intelligent modeling method based on a recurrent neural network in an embodiment of the present invention, which specifically includes:
[0026] S1: Collect geological data, including various types of data such as geological strata, topography and geomorphology, and pre-process and clean the geological data to ensure the quality and availability of the data.
[0027] S2: Analyze and process the geological data after S1 operation based on multi-point geostatistical simulation DS (Direct Sampling) algorithm, such as Figure 3 As shown, the geological model is extracted, including simulation grids (a), e)) and training image grids (b), c), d)). Figure 3 a) in the figure is to define data events in the simulation grid. The question mark indicates the node to be simulated. The two white and black pixels indicate the nodes that have been simulated before. Figure 3 b) using the dimensions a, b, c, d of the data event, define a search window in the TI grid; Figure 3 c) linearly scan the search window starting at a random position until d) the simulated data event is satisfactorily matched, Figure 3 d) is a schematic diagram after simulation data matching; Figure 3e) assign the value of the central node of the first matching data event to the simulation node; finally, the fold data as training data is obtained, and the DS algorithm is executed several times by replacing the training image and conditional data to amplify the sample. After the sample is amplified, multiple training data form a training library; the DS algorithm is used to analyze the geological data, including the spatial distribution and correlation of the geological data, and the geological model is converted into a training library. This training library will fully represent the fold geological semantics and contain geological models of various types and scales. Combine field exploration data and geological expert knowledge to verify and correct the model to improve the richness and reliability of the training library.
[0028] S3: Stack multiple LSTMs to form an encoding and prediction network, quickly generate diverse and multi-scale geological fold models, and use training data to train the geological fold models; LSTM is suitable for the multi-scale characteristics of geological folds. In the model design, the diversity and multi-scale characteristics of geological folds are considered, and the appropriate network structure and parameter settings are designed to ensure that the model can fully express the complexity and diversity of geological folds.
[0029] S4: Based on the two-dimensional section of the geological image actually obtained, the three-dimensional fold data is predicted using the trained geological fold model.
[0030] The implementation process of S2-S3 is as follows Figure 2 As shown, it specifically includes two aspects:
[0031] (I) The process of using the DS algorithm to analyze and process the geological data after the S1 operation to obtain the fold data as training data is as follows:
[0032] (1): Load geological data and reasonably distribute known condition data to all relevant simulation grid nodes to ensure the validity of geological data.
[0033] (2): Define a simulation path to provide clear direction for subsequent steps.
[0034] (3): Start the simulation loop until all simulated grid SG nodes have been visited to ensure the comprehensiveness of the simulation.
[0035] (4): Define the node x to be simulated for subsequent data processing.
[0036] (5) Obtain the data pattern Nx at position x for in-depth analysis.
[0037] (6) Randomly scan the training image TI, and for each pattern Ny on TI, calculate the pattern distance d(Nx,Ny) to evaluate the similarity between the data.
[0038] The pattern distance is the difference a at each positioni Summing and dividing by the total number n is equivalent to calculating the mismatch ratio between the two patterns. The distance is a value between 0 and 1. The smaller the value, the more similar the two patterns are. The pattern distance is expressed as follows:
[0039]
[0040] Where d{Nx,Ny} represents the distance between patterns Nx and Ny; n represents the number of sampling points contained in the pattern; a i represents an indicator variable, which is used to determine whether the values of two patterns at position i are the same; represents the value of the i-th position in the pattern Nx, Represents the value of the i-th position in the pattern Ny;
[0041] when , it means that the two patterns have the same value at the i-th position. At this time, a i =0.
[0042] when , it means that the two patterns have different values at the i-th position. i =1.
[0043] (7) If d(Nx, Ny) is less than the threshold t, Z(x) is set to Z(y) and the current loop ends to achieve a fast response.
[0044] (8) If the scan ratio exceeds the maximum scan ratio f, the value at the position Z(y*) where the pattern distance d(Nx,Ny*) is the smallest is assigned to the node x to be simulated to optimize the result.
[0045] (9) End the current process and ensure that all steps have been completed.
[0046] (10) Figure 3 As shown, the final simulation results are output and a wrinkle pattern training library is obtained for further analysis and application.
[0047] (II) The process of generating geological fold model is as follows:
[0048] (11) Design the network structure of the intelligent modeling model based on the long short-term memory (LSTM) network, especially consider the complexity and diversity of geological folds, and construct a deep LSTM to capture geological characteristics and laws.
[0049] (12) Use the training data prepared in step (10) to train the LSTM-based network structure. LSTM includes a forget gate, an input gate, a memory unit, and an output gate. The key equation of LSTM is as follows:
[0050] Forget Gate: The Forget Gate determines how much state information from the previous moment needs to be discarded. The formula is as follows:
[0051]
[0052] ·f t : The output of the forget gate, which is a vector between 0 and 1, determines the memory state C of the previous moment t-1 How much of each part should be retained.
[0053] ·W xf *X t : Current input X t Impact on the forget gate.
[0054] ·W hf *H t-1 : The hidden state H at the previous moment t-1 Impact on the forget gate.
[0055] · The memory state C at the previous moment t-1 Impact on the forget gate. The introduction of this term enables the forget gate to refer to the memory state of the previous moment and further decide how much past information to retain.
[0056] b f : Bias term of forget gate.
[0057] σ: represents the sigmoid activation function, which compresses the result to between 0 and 1, indicating the proportion of forgetting.
[0058] Input Gate: The input gate determines how much input information at the current moment should be added to the memory state.
[0059]
[0060] ·i t : The output of the input gate, also a value between 0 and 1, determines which new information can be added to the memory state.
[0061] ·W xi *X t : Current input X t Impact on the input gate.
[0062] ·W hi *H t-1 : The hidden state H at the previous moment t-1 Impact on the input gate.
[0063] · The memory state C at the previous moment t-1Impact on the input gate.
[0064] b i : Bias term for the input gate.
[0065] Memory unit update C t :Memory unit C t It is the core of LSTM and is used to store long-term information. It combines the memory of the previous moment with the candidate memory information of the current moment.
[0066]
[0067] · The memory state C at the previous moment t-1 Multiply by the output f of the forget gate t , determines how much past memory is retained.
[0068] · Input Gate i t The candidate memory is limited between -1 and 1 by the tanh activation function, and then combined with the input gate i t The ratio value controls the extent to which new information is added to the memory.
[0069] ·C t : Current memory status.
[0070] ·W xc *X t : The impact of the current input on the memory unit.
[0071] · The hidden state H at the previous moment t-1 Impact on memory cells.
[0072] b c : The bias term of the memory unit;
[0073] Output Gate: The output gate determines the output of the LSTM unit, that is, the hidden state H at the current moment t :
[0074]
[0075] The final hidden state H t : The hidden state H at the current moment t is the final output of LSTM.
[0076] o t :Output gate o t It is used to adjust the strength of the hidden state and determines which information is passed to the next moment or the next layer of the network.
[0077] tanh(C t ): for the current memory state C t Perform a tanh transform and clamp the values to between -1 and 1.
[0078] ·W xo *X t : The impact of the current input on the output gate.
[0079] · The hidden state H at the previous moment t-1 Impact on the output gate.
[0080] b o : Bias term for the output gate.
[0081] (13) All states of the LSTM are initialized to zero before the first input arrives, indicating that there is no prior knowledge about the future.
[0082] (14) The future state of a cell in the grid is determined by combining the LSTM input with the past states of the local neighbors through a convolution operator. The state is padded before applying the convolution operation to ensure that the state has the same number of rows and columns as the input.
[0083] (15) Figure 4 As shown in the figure, the encoding and prediction networks are formed by stacking multiple LSTMs, where the encoding LSTM compresses the entire input sequence into a hidden state tensor, and the prediction LSTM connects all states and then expands the hidden state through a 1×1 convolutional layer to generate the final prediction.
[0084] (16) Through the back propagation algorithm and optimizer, the model parameters are continuously adjusted to improve the model's ability to model geological folds. Regularization and dropout techniques are also applied to prevent the model from overfitting.
[0085] (17) Use the trained LSTM to generate diverse and multi-scale geological fold models, and quickly generate rich models by inputting different geological data.
[0086] (18) Evaluate and verify the generated geological fold model and compare it with the actual geological data to assess the accuracy and realism of the model.
[0087] (19) Adjust and optimize the model based on the verification results to further improve performance and stability.
[0088] (20) Output the model and save the result. So far, a model that can generate three-dimensional wrinkle modeling based on two-dimensional wrinkle sections has been obtained.
[0089] Steps (18)-(19) evaluate and optimize the generated geological fold model, and improve the performance and applicability of the model by comparing and verifying it with actual geological data. By continuously optimizing and improving the model, rapid and accurate modeling of the morphological characteristics of geological folds can be achieved. This provides an efficient and accurate fold model for geological science research and engineering applications, and promotes the development and application of the geological field.
[0090] The geological fold model generation effect diagram of the present invention is as follows Figure 5 As shown, Figure 5 (a) is the input of the geological fold model, a two-dimensional slice. In this example, the size of the two-dimensional slice is 64*64; Figure 5 (b) is the output of the geological fold model, a real three-dimensional fold model; Figure 5 (c) in the figure is the predicted three-dimensional wrinkle model. From the implementation results, it can be seen that the actual modeling is not much different from the predicted modeling, which shows that the model can effectively generate three-dimensional modeling based on two-dimensional sections.
[0091] Example 2
[0092] A three-dimensional wrinkle intelligent modeling device 401 based on a recurrent neural network, such as Figure 6 As shown, it includes: a processor 402 and a storage device 403; the processor 402 loads and executes the instructions and data in the storage device 403 to implement the three-dimensional wrinkle intelligent modeling method based on recurrent neural network.
[0093] Example 3
[0094] A storage device, wherein the storage device stores instructions and data for implementing the three-dimensional wrinkle intelligent modeling method based on recurrent neural network.
[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims
1. A three-dimensional wrinkle intelligent modeling method based on recurrent neural network, characterized in that: include: S1: Collect geological data and preprocess and clean them; S2: Analyze and process the geological data after the S1 operation based on the DS algorithm to obtain fold data as training data. Multiple training data constitute a training library; S3: stack multiple LSTMs to form an encoding and prediction network, generate a geological fold model, and train the geological fold model using training data; S4: Based on the two-dimensional section of the geological image actually obtained, the three-dimensional fold data is predicted using the trained geological fold model.
2. The three-dimensional wrinkle intelligent modeling method based on recurrent neural network according to claim 1, characterized in that: In S1, the geological data includes geological strata and topography.
3. The three-dimensional wrinkle intelligent modeling method based on recurrent neural network according to claim 1, characterized in that: In S2, the DS algorithm is used to analyze the geological data, including obtaining the spatial distribution and correlation of the geological data, and converting the pattern of the geological data into the form of a training library.
4. The three-dimensional wrinkle intelligent modeling method based on recurrent neural network according to claim 1, characterized in that: In S2, the geological data after the S1 operation is analyzed and processed based on the DS algorithm to obtain the fold data as training data as follows: S2.1: Load geological data and distribute known condition data to all relevant simulation grid nodes; S2.2: Define a simulation path; S2.3: Start cyclic simulation until all simulated grid SG nodes are visited; S2.4: define the node x to be simulated, and obtain the data pattern Nx at the node x; S2.5: Randomly scan the training image TI, and for each pattern Ny on TI, calculate the pattern distance d(Nx,Ny) to evaluate the similarity between the data; S2.6: Determine whether d(Nx, Ny) is less than the threshold t. If so, set Z(X) to Z(Y), and then go to S2.
7. If not, determine whether the scan ratio exceeds the maximum scan ratio f. If so, assign the value at the position Z(y*) where the pattern distance d(Nx, Ny*) is the smallest to the node x to be simulated, and then go to S2.
7. If not, return to S2.
3. S2.7: End the current process, output the final simulation results, and obtain wrinkle data.
5. The three-dimensional wrinkle intelligent modeling method based on recurrent neural network according to claim 4, characterized in that: In S2.5, the calculation formula for the mode distance is as follows: Where d{Nx,Ny} represents the distance between patterns Nx and Ny; n represents the number of sampling points contained in the pattern; a i represents an indicator variable, which is used to determine whether the values of two patterns at position i are the same; represents the value of the i-th position in the pattern Nx, Represents the value of the i-th position in pattern Ny.
6. The three-dimensional wrinkle intelligent modeling method based on recurrent neural network according to claim 4, characterized in that: In S3, the process of generating the geological fold model is: S3.1: Design the network structure of the LSTM-based intelligent modeling model and train the network structure; S3.2: Initialize all states of LSTM to zero; S3.3: Determine the future state of a cell in the grid by combining the LSTM input with the past states of the local neighbors through a convolution operator; S3.4: Generate a geological fold model by stacking multiple LSTMs to form an encoding and prediction network.
7. The three-dimensional wrinkle intelligent modeling method based on recurrent neural network according to claim 1, characterized in that: LSTM includes forget gate, input gate, memory unit, and output gate; the forget gate determines how much state information of the previous moment needs to be discarded: Among them, f t represents the output of the forget gate, σ represents the sigmoid activation function, X t Indicates the current input, H t-1 represents the hidden state of the previous moment, C t-1 Indicates the memory state of the previous moment, W xf *X t Indicates the current input X t Impact on the forget gate, W hf *H t-1 Represents the hidden state H at the previous moment t-1 Impact on the forget gate, Represents the memory state C of the previous moment t-1 Impact on the forget gate, b f Represents the bias term of the forget gate; The input gate determines how much of the current input information should be added to the memory state: Among them, i t represents the output of the input gate, W xi *X t Indicates the current input X t Impact on the input gate, W hi *H t-1 Represents the hidden state H at the previous moment t-1 Impact on the input gate, Represents the memory state C of the previous moment t-1 Impact on the input gate, b i represents the bias term of the input gate; The memory unit is the core of LSTM and is used to store long-term information. It combines the memory of the previous moment and the candidate memory information of the current moment. The update formula is: Among them, C t Indicates the current memory status. Represents the memory state C of the previous moment t-1 Multiply by the output f of the forget gate t , determines how much past memory is retained; Represents the input gate i t The candidate memory is limited to between -1 and 1 by the tanh activation function, and then combined with the input gate i t The ratio value controls the extent to which new information is added to memory; W xc *X t Represents the impact of the current input on the memory unit, Represents the hidden state H at the previous moment t-1 Impact on memory cells, b c Represents the bias term of the memory unit; The output gate determines the output of the LSTM unit, that is, the hidden state H at the current moment. t : Among them, the output gate o t It is used to adjust the strength of the hidden state and determines which information is passed to the next moment or the next layer of the network; tanh(C t ) indicates the current memory state C t Perform tanh transformation, W xo *X t Indicates the impact of the current input on the output gate, Represents the hidden state H at the previous moment t-1 Impact on the output gate, b o Represents the bias term of the output gate.
8. A storage device, characterized in that: The storage device stores instructions and data for implementing the three-dimensional wrinkle intelligent modeling method based on recurrent neural network as described in any one of claims 1 to 7.
9. A three-dimensional wrinkle intelligent modeling device based on a recurrent neural network, characterized in that: include: Processor and storage device; the processor loads and executes instructions and data in the storage device to implement the three-dimensional wrinkle intelligent modeling method based on recurrent neural network as described in any one of claims 1 to 7.