Three-dimensional geologic body multi-scale representation modeling method, device and medium

Abstract

The invention relates to the field of geographic information systems, and discloses a three-dimensional geologic body multi-scale representation modeling method and device and a medium, and the method comprises the steps: obtaining multi-source heterogeneous geological observation data, and constructing a data set; a GM-RNN network is constructed; the GM-RNN network comprises a space-time LSTM (Long Short Term Memory) recurrent neural network ST-LSTM unit with a memory decoupling strategy; training the GM-RNN network by using a data set, a reverse scheduling strategy and a space-time memory flow propagation strategy to obtain a trained network; inputting a geological profile of any size and scale of geological observation data into the trained network to realize multi-scale automatic construction of a three-dimensional geologic body; according to the method, the spatial-temporal evolution law characteristics of the complex geologic body are effectively captured, the model can better predict a corresponding prediction result under a relatively long time step length, and the long-term modeling performance of the GMRNN is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of geographic information systems, and in particular to a three-dimensional geological body multi-scale characterization modeling method, equipment and medium. Background Art

[0002] The visualization of three-dimensional urban geological models is the basis for qualitative and quantitative evaluation of underground space and geological environment investigation, and is also an important part of urban informatization development. It plays an important role in describing the distribution laws of underground space and can assist in further risk assessment and decision-making in the practice and development of underground projects.

[0003] Therefore, how to accurately construct a three-dimensional urban geological model based on multi-source heterogeneous urban geological survey data and different modeling workflows so that it can be coupled with a series of different geological attributes such as strata, hydrology, rock mass, etc., and realize the integrated and detailed characterization of underground space structure attribute information has always been a research hotspot in the field of three-dimensional urban geological modeling.

[0004] According to their nature, three-dimensional geological modeling methods can be mainly divided into two categories: deterministic modeling and stochastic modeling. Among them, stochastic modeling methods effectively overcome the limitations of deterministic modeling and can quantify the uncertainty of geological parameter distribution patterns, so they have been widely used in the field of stochastic modeling. In recent years, under the dual drive of geological knowledge and big data, deep learning algorithms can model complex and nonlinear geological processes in nature. This method can capture the correlation between different information without any prior assumptions. Even under the premise of unknown rules, deep learning can analyze and create corresponding distribution patterns and laws from a large number of data samples. Therefore, a series of deep learning models provide a good foundation for the field of urban three-dimensional geological modeling. Common deep generative models, such as variational autoencoders (VAEs) and generative adversarial networks (GANs), can effectively extract complex geological patterns and fit the high-order distribution characteristics of their parameters. However, due to differences in network structure and similarity metrics, not only will the generated model collapse, resulting in the simulation results being unable to be faithful to the given conditional data constraints, but also unable to effectively characterize the multi-scale characteristics of the geological model. The precision and accuracy of three-dimensional geological modeling need to be improved.

[0005] As a neural network with short-term memory, recurrent neural network (RNN) can effectively process sequence data and capture its internal dependencies, and has a relatively stable training mode. Among them, convolutional long short-term memory network (ConvLSTM) can encode the information contained in the image sequence, can explicitly model and represent the visual appearance and time dynamics at the same time, and better maintain the fine-grained attribute information of the image. A three-dimensional geological body can be regarded as a time-series image composed of multiple continuous geological sections. The structural attribute characteristics corresponding to multiple continuous sections maintain certain correlation characteristics at a specific time and space scale. Therefore, referring to the prediction learning theory, ConvLSTM can be combined with the modeling process to enable it to comprehensively consider the temporal and spatial regularity of the geological body between different layers. After predicting the structural attribute information of multiple layers, they are stacked in a certain direction, and finally the corresponding three-dimensional geological model is established. However, ConvLSTM is mostly used in the prediction process of time-series images, and it is difficult to directly apply it to the three-dimensional geological modeling process. At the same time, ConvLSTM cannot take into account both low-level and high-level geological semantic information during the prediction process, which easily leads to the loss of information flow and makes it difficult to uniformly represent the complex spatiotemporal attribute change laws of three-dimensional geological bodies. In addition, since the parameters between the encoding and prediction stages are shared, the prediction process is more dependent on the contextual information of adjacent results, which will affect the network's long-term modeling and prediction capabilities.

[0006] In summary, the existing methods do not fully consider the spatiotemporal variation of three-dimensional geological bodies in the modeling process, and there are certain deficiencies in modeling performance and multi-scale expression capabilities. Summary of the invention

[0007] The purpose of the present invention is to propose a three-dimensional geological body multi-scale characterization modeling method, equipment and medium to solve the technical problem that the existing technology does not fully consider the internal spatiotemporal variation characteristics of the three-dimensional geological body during the modeling process, resulting in poor modeling performance and lack of multi-scale expression ability.

[0008] In order to fully consider the spatiotemporal variation of three-dimensional geological bodies in the modeling process and improve the modeling performance and multi-scale expression capabilities, the present invention makes certain improvements on the basis of RNN, adjusts the propagation mode of information flow in network units at different levels, ensures that the dynamic response mechanism of the geological model can be effectively communicated, improves the long-term and short-term comprehensive modeling and prediction capabilities of network units, realizes the accurate prediction of the distribution patterns of geological parameters at different levels, and finally realizes the fine characterization and multi-scale representation of three-dimensional geological bodies.

[0009] Specifically, the present invention provides a three-dimensional geological body multi-scale characterization modeling method, comprising the following steps:

[0010] S1: Acquire multi-source heterogeneous geological observation data and construct a data set;

[0011] S2: constructing a GM-RNN network; the GM-RNN network includes a spatiotemporal LSTM recurrent neural network ST-LSTM unit with a memory decoupling strategy;

[0012] S3: Train the GM-RNN network using the data set, reverse scheduling strategy, and spatiotemporal memory flow propagation strategy to obtain a trained network;

[0013] S4: Input geological profiles of any size of geological observation data into the trained network to achieve multi-scale automatic construction of three-dimensional geological bodies.

[0014] A storage medium stores instructions and data for implementing a three-dimensional geological body multi-scale characterization modeling method.

[0015] A three-dimensional geological body multi-scale characterization modeling device comprises: a processor and the storage medium; the processor loads and executes instructions and data in the storage medium to implement a three-dimensional geological body multi-scale characterization modeling method.

[0016] The beneficial effects provided by the present invention are:

[0017] (1) The present invention introduces a ST-LSTM module with a memory decoupling strategy. Among them, ST-LSTM can ensure the long-term consistency and short-term dynamic response of the hidden state under the action of the time memory unit. In addition, based on the dual memory flow conversion mechanism, the two memory units in ST-LSTM are implicitly trained, processed and unified by the output gate under the control of multiple signals, so as to perform long-term and short-term dynamic comprehensive modeling of the image sequence. At the same time, the memory decoupling strategy enables the state conversion mechanism to maintain a certain diversity, which can ensure that the untangled representation of long-term and short-term dynamics is intuitively pursued during the prediction learning process, so as to obtain reasonable prediction results in the future time step. Under the condition of spatiotemporal memory flow propagation, low-level geological semantic information gradually propagates to the top layer in the vertical direction, and in the next time step, the acquired high-level memory state will be transmitted to the bottom layer. The bottom-up and top-down tortuous memory flows can traverse each node in the neural network, and effectively learn the corresponding memory representation in different horizontal recursive layers, effectively capturing the spatiotemporal evolution law characteristics of complex geological bodies.

[0018] (2) Based on the stacked ST-LSTM units and spatiotemporal memory flow propagation, the training process of GM-RNN is optimized through the reverse scheduling sampling strategy, which effectively alleviates the problem of insufficient optimization of network parameters in the encoding and decoding stages. By designing different sampling probabilities to randomly hide different parts of the input sequence, the encoding process gradually changes from multi-step prediction to one-step prediction, and the long-term, non-Markov dynamic properties of the input image sequence are effectively preserved. Therefore, at a longer time step, the model can better predict the corresponding prediction results, and the long-term modeling performance of GM-RNN is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic flow chart of the method of the present invention;

[0020] Figure 2 represents the network structure, where Figure 2 (a) represents the propagation mode of spatiotemporal memory flow between network nodes at different levels; Figure 2 (b) in the figure represents the structure and operation flow diagram of the spatiotemporal LSTM recurrent neural network with memory decoupling strategy;

[0021] Figure 3 The representative is the multi-scale reconstruction results based on the continuous attribute fold model and its corresponding quantitative evaluation results, including multi-scale simulation based on the vertical direction and multi-scale simulation based on the horizontal direction;

[0022] Figure 4 It is a working schematic diagram of the hardware device of the present invention. DETAILED DESCRIPTION

[0023] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0024] Before formally describing the present invention, the scheme of the present invention is first generally described for easy understanding.

[0025] Please refer to Figure 1 The present invention provides a three-dimensional geological body multi-scale characterization modeling method, comprising the following steps:

[0026] S1: Acquire multi-source heterogeneous geological observation data and construct a data set;

[0027] As an embodiment, a three-dimensional training image pattern library of the study area is established using multi-source heterogeneous geological observation data. The pattern library can be regarded as an expression of prior geological knowledge that integrates geological condition data and expert experience. Then, for each sub-model in the training image library, the top-level two-dimensional geological profile is regarded as a conditional data, thereby establishing multiple sets of training data sets to support network training.

[0028] S2: Constructing a GM-RNN (Geological-Modeling-based Recurrent Neural Network) network; the GM-RNN network includes a spatiotemporal LSTM recurrent neural network ST-LSTM unit with a memory decoupling strategy; the GM-RNN network is a recurrent neural network based on three-dimensional geological modeling.

[0029] It should be noted that in order to avoid overfitting problems and excessive computational overhead during the modeling process, only four ST-LSTM units need to be stacked sequentially to establish a GM-RNN network.

[0030] Please refer to Figure 2 It should be noted that ST-LSTM can ensure the long-term consistency and short-term dynamic response of hidden states under the action of temporal memory units. Moreover, based on the dual memory flow conversion mechanism, the two memory units in ST-LSTM are implicitly trained, processed and unified by the output gate under the control of multiple signals, so as to comprehensively model the long-term and short-term dynamics of the image sequence. At the same time, the memory decoupling strategy enables the state conversion mechanism to maintain a certain diversity, which can ensure that the unentangled representation of long-term and short-term dynamics is intuitively pursued in the prediction learning process, so as to obtain reasonable prediction results in the future time step. Under the condition of spatiotemporal memory flow propagation, low-level geological semantic information gradually propagates to the top layer in the vertical direction, and in the next time step, the acquired high-level memory state will be transmitted to the bottom layer. The bottom-up and top-down tortuous memory flows can traverse every node in the neural network, and effectively learn the corresponding memory representation in different horizontal recursive layers, effectively capturing the spatiotemporal evolution law characteristics of complex geological bodies.

[0031] S3: Train the GM-RNN network using the data set, reverse scheduling strategy, and spatiotemporal memory flow propagation strategy to obtain a trained network;

[0032] It should be noted that step S3 is specifically as follows:

[0033] S31. Select a small batch of samples from the data set, apply a reverse scheduling sampling strategy based on exponential changes to them, and obtain a mixed tensor for training the network;

[0034] Specifically, in step S31, the reverse scheduling sampling strategy specifically refers to: randomly hiding part of the information in the input data at a certain ratio according to the number of training times. The process is as follows:

[0035]

[0036] in, represents the probability of application ε k ∈[0,1] and 1-ε k From the reference image and prediction results Random sampling is performed in k It usually grows as the training iteration k deepens, with the start and end positions being ε s and ε e , α e is a proportional factor used to control the growth trend of the exponential function; this process will eventually obtain a mixed tensor containing True and False, recorded as mask_true, which will be used as the training process of the subsequent network.

[0037] S32, initialize the GM-RNN network training hyperparameters and the state of the spatiotemporal LSTM recurrent neural network ST-LSTM unit with a memory decoupling strategy;

[0038] It should be noted that in step S32, the network training hyperparameters include: total training times epochs, batch training set size batch_size, number of ST-LSTM units num_layers, decoupling loss weight coefficient β, RSS sampling starting position ∈ s With the end position ∈ e , input sequence length input_length, growth factor α e , sampling probability form, learning rate lr and type of optimizer.

[0039] In addition, the state of the ST-LSTM unit includes: hidden state Time memory state C t Its unit time increment ΔC, space-time memory state and its unit increment Initialized, the above parameters are initialized to 0 for corresponding data storage.

[0040] S33, according to the preset prediction range L predict , determine the training time step t and the prediction range L predict The size relationship is based on which the spatiotemporal memory flow propagation strategy is used to traverse and update the state of the ST-LSTM unit to obtain the final state of the ST-LSTM unit.

[0041] As an embodiment, step S33 is specifically as follows:

[0042] S331, determine the time step t and L predict relationship; if t <L input , then the first image sequence in the input data is selected and recorded as net; otherwise, net is obtained by mixing and adding the input data and the prediction results with a certain weight ratio. The weight ratio is calculated using the mixed tensor mask_true, which can be written as:

[0043] net=mask_true*Geo_sec t +(1-mask_true t-1 )*Geo_sec gen ,t≥2;

[0044] Among them, Geo_set t is the geological profile input in the tth time period, starting from the second profile sequence data, Geo_sec gen The geological section predicted for the adjacent previous time period;

[0045] S332, time memory through ST-LSTM unit and space-time memory The paired memory states are finally merged by the output gate o with multiple control signals. t To process and unify;

[0046] It should be noted that the ST-LSTM unit realizes the comprehensive modeling of long-term and short-term dynamics of image sequences. At the same time, under the action of the dual mechanism, a reasonable inductive bias is used to construct the corresponding distribution representation). If the distribution representation conforms to the underlying data distribution logic, the decoupled loss representation can promote the simulation and prediction performance of the network, and make the memory state focus on different aspects of the spatiotemporal variation characteristics, effectively preventing the memory state from learning the redundant features of the image sequence.

[0047] S333, for the first ST-LSTM unit, net, C0 and will be updated for the first time, C0 and The increments during this phase will be recorded as ΔC and middle;

[0048] S334, traverse the remaining ST-LSTM units and update the corresponding C t and C corresponding to different ST-LSTM units t and The increments will be stored in Δc and middle.

[0049] As an example, a spatiotemporal memory flow propagation strategy is introduced during the traversal process, which first updates upward across layers and then advances to the next time step. The state transition path is extended by crossing the time dimension, and additional neurons are added between adjacent nodes at the same level of the network. This tortuous propagation method allows all nodes to jointly maintain a network composed of The storage body represented by ensures that the memory states in different ConvLSTM layers are closely connected, and the horizontal memory flow propagation and storage are improved. This process can be summarized as follows:

[0050]

[0051] in, represents time memory, l and t represent different neural unit layers and time steps, respectively. Represents space-time memory, is the hidden state, is the input image sequence, i t 、f t and g t is a set of gated units, namely the input gate, forget gate and adjustment gate in LSTM. t is the output gate, used to generate prediction results, W * is the weight coefficient corresponding to different items, σ(*) is the Sigmoid activation function calculation, and ⊙ is the dot product operation.

[0052] S34. Obtain the decoupling loss and total loss function according to the final state of the ST-LSTM unit;

[0053] It should be noted that the calculation of ΔC and The decoupled loss between the predicted image and the reference image is calculated at the same time, the two types of loss terms are combined with different weights, and the parameters of the network are iteratively updated using stochastic gradient descent and back propagation algorithms.

[0054] The decoupled loss and total loss function can be written as follows:

[0055]

[0056] Among them, W decouple is a 1×1 convolution operator, acting on all ST-LSTM units, <·,·> is a dot product operation, ||·|| represents the l2 norm, acting on each feature map on channel c. t is the unit time step, T is the total time step, K represents the length of the image sequence prediction, and They represent the prediction results of the network and the corresponding reference image within time t respectively, and ω is the weight coefficient of the decoupling loss term.

[0057] S35. Repeat steps S33 to S34 according to the preset hyperparameters until the preset training end condition is met to obtain a trained network.

[0058] As an embodiment, the input image sequence data is arranged in reverse order and used as the input of the GM-RNN network, and the network parameters are re-optimized according to steps S33 to S34.

[0059] It should be noted that after all training stages are completed, observe whether the loss function converges to a certain interval. If it converges, the network model parameters need to be saved and the simulation results need to be evaluated.

[0060] During the test, a geological profile data is randomly selected as the input of the network. By predicting the structural attribute information of multiple subsequent layers and stacking the continuously distributed two-dimensional geological profiles along a certain direction, the multi-scale automatic construction of the three-dimensional geological body is realized. If the reconstructed model meets the given geological profile constraints and maintains a similar distribution pattern with the reference model, and has little difference in spatial variability, attribute information ratio and spatial structure, the network parameters can be saved to achieve the fine characterization of the three-dimensional geological body under the constraints of the two-dimensional geological profile. Otherwise, the network parameters are reset and a new round of training process is started.

[0061] S4: Input geological sections of any scale of geological observation data into the trained network to realize multi-scale automatic construction of three-dimensional geological bodies.

[0062] Finally, please refer to Figure 3 , Figure 3 It represents the multi-scale reconstruction results based on the continuous attribute fold model and its corresponding quantitative evaluation results, including multi-scale simulation based on the vertical direction and multi-scale simulation based on the horizontal direction.

[0063] from Figure 3 It can be seen that for the pre-trained GM-RNN, when a 64*64 geological profile is input, the results of multiple subsequent layers can be predicted. Similarly, when the sizes of the input geological profiles are 80*80 and 96*96 respectively, GM-RNN can also achieve multi-scale representation and modeling of three-dimensional geological bodies in the horizontal direction.

[0064] See also Figure 4 , Figure 440 is a schematic diagram of the working of the hardware device of an embodiment of the present invention, wherein the hardware device specifically comprises: a three-dimensional geological body multi-scale characterization modeling device 401, a processor 402 and a storage medium 403.

[0065] A three-dimensional geological body multi-scale characterization modeling device 401: The three-dimensional geological body multi-scale characterization modeling device 401 implements the three-dimensional geological body multi-scale characterization modeling method.

[0066] Processor 402: The processor 402 loads and executes the instructions and data in the storage medium 403 to implement the three-dimensional geological body multi-scale characterization modeling method.

[0067] Storage medium 403: The storage medium 403 stores instructions and data; the storage medium 403 is used to implement the three-dimensional geological body multi-scale characterization modeling method.

[0068] The beneficial effects of the present invention are:

[0069] (1) The present invention introduces a ST-LSTM module with a memory decoupling strategy. Among them, ST-LSTM can ensure the long-term consistency and short-term dynamic response of the hidden state under the action of the time memory unit. In addition, based on the dual memory flow conversion mechanism, the two memory units in ST-LSTM are implicitly trained, processed and unified by the output gate under the control of multiple signals, so as to perform long-term and short-term dynamic comprehensive modeling of the image sequence. At the same time, the memory decoupling strategy enables the state conversion mechanism to maintain a certain diversity, which can ensure that the untangled representation of long-term and short-term dynamics is intuitively pursued during the prediction learning process, so as to obtain reasonable prediction results in the future time step. Under the condition of spatiotemporal memory flow propagation, low-level geological semantic information gradually propagates to the top layer in the vertical direction, and in the next time step, the acquired high-level memory state will be transmitted to the bottom layer. The bottom-up and top-down tortuous memory flows can traverse each node in the neural network, and effectively learn the corresponding memory representation in different horizontal recursive layers, effectively capturing the spatiotemporal evolution law characteristics of complex geological bodies.

[0070] (2) Based on the stacked ST-LSTM units and spatiotemporal memory flow propagation, the training process of GM-RNN is optimized through the reverse scheduling sampling strategy, which effectively alleviates the problem of insufficient optimization of network parameters in the encoding and decoding stages. By designing different sampling probabilities to randomly hide different parts of the input sequence, the encoding process gradually changes from multi-step prediction to one-step prediction, and the long-term, non-Markov dynamic properties of the input image sequence are effectively preserved. Therefore, at a longer time step, the model can better predict the corresponding prediction results, and the long-term modeling performance of GM-RNN is effectively improved.

[0071] 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 geological body multi-scale characterization modeling method, characterized by: The method comprises the following steps: S1: Acquire multi-source heterogeneous geological observation data and construct a data set; S2: constructing a GM-RNN network; the GM-RNN network includes a spatiotemporal LSTM recurrent neural network ST-LSTM unit with a memory decoupling strategy; S3: Train the GM-RNN network using the data set, reverse scheduling strategy, and spatiotemporal memory flow propagation strategy to obtain a trained network; S4: Input geological sections of any scale of geological observation data into the trained network to achieve multi-scale automatic construction of three-dimensional geological bodies.

2. A three-dimensional geological body multi-scale characterization modeling method according to claim 1, characterized in that: Step S3 is as follows: S31. Select a small batch of samples from the data set, apply a reverse scheduling sampling strategy based on exponential changes to them, and obtain a mixed tensor for training the network; S32, initialize the GM-RNN network training hyperparameters and the state of the spatiotemporal LSTM recurrent neural network ST-LSTM unit with a memory decoupling strategy; S33, according to the preset prediction range L predict , determine the training time step t and the prediction range L predict The size relationship is based on which the spatiotemporal memory flow propagation strategy is used to traverse and update the state of the ST-LSTM unit to obtain the final state of the ST-LSTM unit. S34. Obtain the decoupling loss and total loss function according to the final state of the ST-LSTM unit; S35. Repeat steps S33 to S34 according to the preset hyperparameters until the preset training end condition is met to obtain a trained network.

3. A three-dimensional geological body multi-scale characterization modeling method according to claim 2, characterized in that: In step S31, the reverse scheduling sampling strategy specifically refers to: randomly hiding part of the information in the input data at a certain ratio according to the number of training times. The process is as follows: in, represents the probability of application ε k ∈[0,1] and 1-ε k From the reference image and prediction results Random sampling is performed in k It usually grows as the training iteration k deepens, with the start and end positions being ε s and ε e , α e is a proportional factor used to control the growth trend of the exponential function; this process will eventually obtain a mixed tensor containing True and False, recorded as mask_true, which will be used as the training process of the subsequent network.

4. A three-dimensional geological body multi-scale characterization modeling method according to claim 1, characterized in that: In step S32, the network training hyperparameters include: total training times epochs, batch training set size batch_size, number of ST-LSTM units num_laters, decoupling loss weight coefficient β, RSS sampling starting position ∈ s With the end position ∈ e , input sequence length input_length, growth factor α e , sampling probability form, learning rate lr and type of optimizer.

5. A three-dimensional geological body multi-scale characterization modeling method according to claim 3, characterized in that: In step S32, the state of the ST-LSTM unit includes: hidden state Time memory state C t Its unit time increment ΔC, space-time memory state and its unit increment Initialized, the above parameters are initialized to 0 for corresponding data storage.

6. A three-dimensional geological body multi-scale characterization modeling method according to claim 5, characterized in that: Step S33 is specifically as follows: S331, determine the time step t and L predict relationship; if t <L input , then the first image sequence in the input data is selected and recorded as net; otherwise, net is obtained by mixing and adding the input data and the prediction results with a certain weight ratio. The weight ratio is calculated using the mixed tensor mask_true, which can be written as: net=mask_true*Geo_sec t +(1-mask_true t-1 )*Geo_sec gen ,t≥2; Among them, Geo_sec t is the geological profile input in the tth time period, starting from the second profile sequence data, Geo_sec gen The geological section predicted for the adjacent previous time period; S332, time memory through ST-LSTM unit and space-time memory The paired memory states are finally merged by the output gate o with multiple control signals. t To process and unify; S333, for the first ST-LSTM unit, net, C0 and will be updated for the first time, C0 and The increments during this phase will be recorded as ΔC and middle; S334, traverse the remaining ST-LSTM units and update the corresponding C t and C corresponding to different ST-LSTM units t and The increments will be stored in ΔC and middle.

7. A three-dimensional geological body multi-scale characterization modeling method according to claim 6, characterized in that: In the traversal process of step S334, a spatiotemporal memory flow propagation strategy is introduced, which first updates upward across layers and then advances to the next time step. The state transfer path is extended by crossing the time dimension, and additional neurons are added between adjacent nodes at the same level of the network, as shown in the following formula: in, represents time memory, l and t represent different neural unit layers and time steps, respectively. Represents space-time memory, is the hidden state, is the input image sequence, i t 、f t and g t is a set of gated units, namely the input gate, forget gate and adjustment gate in LSTM. t is the output gate, used to generate prediction results, W * is the weight coefficient corresponding to different items, σ(*) is the Sigmoid activation function calculation, and ⊙ is the dot product operation.

8. A three-dimensional geological body multi-scale characterization modeling method according to claim 7, characterized in that: The mathematical formula for decoupling loss and total loss function in step S34 is as follows: Among them, W decouple is a 1×1 convolution operator, which acts on all ST-LSTM units, <·,·> is a dot product operation, and ‖·‖ represents norm, acting on each feature map on channel c. t is the unit time step, T is the total time step, and K represents the length of the image sequence prediction. and They represent the prediction results of the network and the corresponding reference image within time t respectively, and ω is the weight coefficient of the decoupling loss term.

9. A storage medium, characterized in that: The storage medium stores instructions and data for implementing a three-dimensional geological body multi-scale characterization modeling method as described in any one of claims 1 to 8.

10. A three-dimensional geological body multi-scale characterization modeling device, characterized by: include: Processor and storage medium; the processor loads and executes instructions and data in the storage medium to implement a three-dimensional geological body multi-scale characterization modeling method as described in any one of claims 1 to 8.

Images