A multimodal geological feature fusion method and system based on discrete wavelet transform and CLIP-StableDiffusion model
By employing a multimodal geological feature fusion method combining discrete wavelet transform and the CLIP-StableDiffusion model, the problems of cross-scale physical correlation and semantic feature fusion were solved, achieving high-precision multimodal geological data fusion and improving the accuracy and professional interpretability of geological modeling.
Patent Information
- Application Number
- CN202511271467.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-08
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2045-09-08
AI Technical Summary
Existing multimodal geological feature fusion methods cannot effectively establish cross-scale physical connections and fail to effectively integrate geological semantic features, resulting in low participation of geological language knowledge and limited generalization ability.
A multimodal geological feature fusion method based on discrete wavelet transform and CLIP-StableDiffusion model is adopted. Feature data is reconstructed through multi-scale discrete wavelet transform, and a CLIP-StableDiffusion network model is constructed, including a text encoder, an image encoder and a U-net diffusion generator, to achieve the fusion of multimodal data.
It significantly improves the ability to capture abrupt changes in geological body boundaries, achieves dynamic alignment between text descriptions and image features, generates sedimentary simulation images that clearly present the distribution patterns of sand bodies, accurately restores porosity and permeability parameters from well logging curves, and outputs results that meet the industrial-grade accuracy requirements of geological modeling.
Smart Images

Figure CN121167609B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and geological information modeling technology, and in particular to a multimodal geological feature fusion method and system based on discrete wavelet transform and CLIP-StableDiffusion model. Background Technology
[0002] Sedimentary reservoirs exhibit diverse characteristics due to the complex dynamic conditions during their formation, particularly in terms of geological and sedimentary properties, resulting in abundant multimodal data, including sedimentary models, seismic data, well logging data, sedimentary numerical simulations, and geological textual reports. Utilizing multimodal deep learning frameworks can integrate different types of reservoir data, and by uncovering deep-seated correlations hidden behind complex data through cross-modal learning, sedimentary reservoir geological modeling can be effectively optimized. Integrating large-scale multimodal models with geological and sedimentary data, and using different data sources, integrated data, and knowledge, along with training structures based on the large-scale multimodal models, not only improves model performance but also enhances its interpretability. Currently, in geological data fusion prediction, geological knowledge and expert experience are not effectively utilized; a single data source cannot fully reflect the internal structure of geological bodies, mainly due to data limitations, model complexity, and the lack of effective fusion mechanisms. By constructing a multimodal geological knowledge graph, developing interactive modeling tools, introducing adaptive learning mechanisms, and integrating multimodal data, the accuracy and practicality of geological data prediction and evaluation can be significantly improved. This allows geologists' professional knowledge and long-term accumulated experience to play a key role in objective geological data, thereby providing more precise support for fields such as oil and gas exploration and mineral resource assessment.
[0003] Currently, the methods for fusing multimodal geological features include spatiotemporal alignment of different resolution modalities such as seismic data and well logging data through time series interpolation or spatial gridding, or processing well logging time series data through long short-term memory networks, extracting spatial features of seismic profiles through convolutional neural networks, and using fully connected layers to achieve fusion.
[0004] The disadvantages of existing technologies include:
[0005] 1) Existing multimodal geological feature fusion is based on a single scale. Geological data often includes modes with different spatiotemporal resolutions, such as seismic waves and well logging curves. Feature-level fusion makes it difficult to establish cross-scale physical correlations.
[0006] 2) Existing multimodal geological feature fusion does not effectively integrate semantic features, i.e., text learning, with other modal data, resulting in low participation of geological language knowledge in the geological feature fusion process. Summary of the Invention
[0007] This invention addresses the shortcomings of existing multimodal geological feature fusion technologies, which are limited by different data scales, making it impossible to effectively establish cross-scale physical relationships, only able to fuse objective geological data, unable to effectively incorporate geological semantic features, and have limited generalization capabilities. It provides a multimodal geological feature fusion method and system based on discrete wavelet transform and CLIP-StableDiffusion model.
[0008] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows:
[0009] A multimodal geological feature fusion method based on discrete wavelet transform and CLIP-StableDiffusion model includes the following steps:
[0010] S1: Input multimodal geological data, including 3D seismic data, 3D sedimentary simulation data, 2D sedimentary numerical simulation images, well logging curves, and geological text descriptions;
[0011] S2: Standardize the data for each type: image data is normalized using Z-score, and text data is filtered using WordPiece word segmentation.
[0012] S3: Divide the training set and the test set into an 8:2 ratio;
[0013] S4: Reconstructing feature data using multi-scale discrete wavelet transform:
[0014] S41: Seismic data and sedimentary model data were reconstructed using Daubechies wavelet transform;
[0015] S42: Reconstruction of sedimentary numerical simulation images and well logging data using Coiflets wavelet transform.
[0016] S43: Geological texts are vectorized using wavelet transform of orthogonal basis matrices;
[0017] S5: Construct the CLIP-StableDiffusion network model, including: text encoder, image encoder, U-net diffusion generator, image decoder, and text decoder;
[0018] Configure the text encoder structure, including: tokenizer, embedding layer, Transformer encoding layer, and projection layer;
[0019] Configure the image encoder structure, including: image splitter, fully connected layer, and Transformer encoding layer;
[0020] Configure the U-net diffusion generator, including: downsampling module and upsampling module.
[0021] The downsampling module contains convolutional layers and Transformer blocks, and inserts Gaussian noise;
[0022] The upsampling module contains deconvolutional layers and Transformer blocks to perform denoising diffusion;
[0023] S6: Input the reconstructed data into the CLIP-StableDiffusion network model for training, and optimize the generated results through the diffusion noise addition / denoising process;
[0024] S7: Calculate the root mean square error loss. Terminate training when the root mean square error (RMSE) between the output and the actual data is less than the preset error; otherwise, return to S6.
[0025] S8: Perform multi-scale wavelet inverse transform on the model output to generate fused seismic data volume, sedimentary numerical simulation image, well logging curve and geological text.
[0026] Furthermore, the type of the discrete wavelet transform is dynamically selected based on the data type:
[0027] Seismic data and sedimentary model data were reconstructed using a 4-level Daubechies wavelet transform;
[0028] The scaling equation is:
[0029] The wavelet equation is:
[0030] Where h k For the low-pass filter coefficients, g k Here, N represents the high-pass filter coefficients, N is the wavelet order, and k is the filter coefficient index.
[0031] Sedimentary numerical simulation images and well logging data were reconstructed using a 3-layer Coiflets wavelet transform.
[0032] The scaling equation is:
[0033] The wavelet equation is:
[0034] Where L = 6N-1 is the filter length;
[0035] Geological texts are vectorized using wavelet transform with two-level orthogonal basis matrices;
[0036] Furthermore, the text encoder of the CLIP-StableDiffusion model specifically includes a vocabulary of 49,408 tokens and an input length of 77 tokens;
[0037] The embedding dimension is 768, and the positional encoding is aligned with the input length.
[0038] Transformer encoding layer: 12 attention heads, 3072 intermediate layer dimensions, 64 attention dimensions, and output projected to a 512-dimensional shared space;
[0039] According to the method of claim 1, the image encoder specifically comprises:
[0040] The input image size is 512×512, the block size is 16×16, and the number of blocks is M=1024;
[0041] The fully connected layer has an input dimension of 768 and an output dimension of 768.
[0042] In convolution operations, the number of feature map channels doubles layer by layer, while in deconvolution operations, the number of channels halves layer by layer.
[0043] Furthermore, the structure of the U-net diffusion generator is as follows:
[0044] Downsampling part: 4 iterations, each containing 3 3×3 convolutional layers and 1 Transformer block, with 4 cross-attention heads and 512 temporal embedding dimensions;
[0045] Upsampling: 4 iterations, each containing 3 x 3 deconvolutional layers and 1 Transformer block, preserving high-frequency details through skip connections;
[0046] Furthermore, the upsampling diffusion process employs a denoising diffusion probability model:
[0047]
[0048] Where, α t ,β t σ is the noise dispatch coefficient. t is the variance term that controls randomness, ∈ is the newly added random noise, and c is the text condition.
[0049] Furthermore, the Adam optimizer is used to update the parameters, and the weights and biases of the network are updated according to the gradient. The training is set to 1000 epochs, with each epoch consisting of 50 loops, until the model achieves optimal performance on the validation set.
[0050] This invention also discloses a multimodal geological feature fusion system, which can be used to implement the above-mentioned multimodal geological feature fusion method, specifically including:
[0051] The multimodal data input interface module is equipped with a dedicated data access port for receiving 3D seismic data volumes, 3D sedimentary simulation data volumes, 2D sedimentary numerical simulation images, various types of well logging curves, and structured geological text descriptions. The geological text contains key elements for lithological classification and sedimentary facies delineation, ensuring the integrity of the original data input.
[0052] The data standardization and partitioning module performs image data standardization to eliminate dimensional differences, and uses a geological dictionary to drive word segmentation and filter redundant information for text data. After standardization, the data is divided into training and testing sets according to a preset ratio. The training set is used for model parameter optimization, while the testing set is used for final generalization capability verification.
[0053] The multi-scale feature reconstruction module incorporates a differentiated wavelet processor: seismic and sedimentary model data undergo multi-level feature decomposition using Daubechies wavelets, sedimentary numerical simulation and well logging data are extracted using Coiflets wavelets, and geological text is mapped to semantic vectors through orthogonal basis matrices. Each processor adaptively adjusts the decomposition level according to the data type.
[0054] The CLIP-StableDiffusion engine integrates a dual-path encoding architecture: the text encoding unit performs word segmentation embedding and semantic vector transformation, while the image encoding unit performs block processing and spatial feature extraction. The core U-net diffusion generator includes a two-stage process of downsampling noise injection and upsampling semantic-guided denoising, and uses a cross-attention mechanism to fuse geological text conditions to control the generation direction.
[0055] The training control and output generation module dynamically monitors the training loss value and employs a threshold termination mechanism driven by geological principles to control the iteration process. Training automatically stops when the multimodal joint error falls below the threshold experienced by geological experts. The output data is then restored through inverse wavelet transform to form fused seismic data volumes, sedimentary simulation images, well logging curves, and structured geological reports.
[0056] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned multimodal geological feature fusion method.
[0057] The present invention also discloses a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described multimodal geological feature fusion method.
[0058] Compared with the prior art, the advantages of the present invention are as follows:
[0059] 1. By employing differentiated discrete wavelet transform technology, a dedicated decomposition strategy is designed for high-frequency signals such as seismic bodies and sedimentary simulation data, significantly improving the ability to capture abrupt changes in geological body boundaries. The synergistic analysis of high-frequency details and low-frequency trends effectively addresses the problem of missing cross-scale physical correlations, overcoming the ambiguity limitations of traditional methods in lithological interface identification.
[0060] 2. This innovative approach integrates the semantic understanding capabilities of the CLIP model with the generative capabilities of Stable Diffusion, enabling dynamic alignment between textual descriptions (such as "tidal deltaic sediments") and image features (such as seismic profiles). Geological expert knowledge is embedded and diffused through orthogonal basis vectorization, endowing the model with professional interpretability and overcoming the technical bottleneck of single-data-driven approaches.
[0061] 3. The U-net diffusion generator based on the cross-attention mechanism simultaneously optimizes the consistency between image spatial structure and text semantics during the denoising process. The generated sedimentary simulation images clearly present the distribution pattern of sand bodies, and the well logging curves accurately restore the trends of porosity and permeability parameters. The output results meet the industrial-grade accuracy requirements of geological modeling.
[0062] 4. The modular design of the entire process, from data standardization and wavelet reconstruction to diffusion training, significantly reduces the intensity of manual intervention. The training termination mechanism intelligently controls the number of iterations through an error threshold constrained by geological laws, avoiding the risk of overfitting and shortening the reservoir evaluation cycle by more than 40%.
[0063] 5. The proposed CLIP-StableDiffusion framework can be applied to fields such as mineral exploration and geothermal reservoir assessment. The configurable characteristics of wavelet transform parameters and diffusion model structure support adaptation to geophysical datasets of different resolutions, providing a standardized intelligent solution for energy exploration. Attached Figure Description
[0064] Figure 1 This is a technical roadmap of the multimodal geological feature fusion method according to an embodiment of the present invention.
[0065] Figure 2 This is a schematic diagram of the CLIP-stableDiffusion network model structure according to an embodiment of the present invention. Detailed Implementation
[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and examples.
[0067] like Figure 1 As shown, this invention provides a multimodal geological feature fusion method based on discrete wavelet transform and CLIP-StableDiffusion model, comprising the following steps:
[0068] Step 1: Input Data Classification and Integration: Geological data is classified according to data characteristics into seismic data (3D time-series images with a size of 128×128×128), sedimentary simulation learning mode (3D data volume with a size of 128×128×128), sedimentary numerical simulation (2D profile images with a size of 128×128), well logging data (including 9 types of well logging curves such as GR / SP / RT), and geological text descriptions (including 15 types of geological elements such as lithology and sedimentary facies).
[0069] Step 2: Data Standardization: Before inputting the data into the model, it needs to be standardized to reduce the dimensional differences between different features. Image signal data is normalized using the Z-score method. The numerical range is scaled to near 0 without altering the data distribution. x represents a data point, μ is the mean of the data points, and σ is the standard deviation of the data points. Text data is segmented using WordPiece to retain words with geological information while filtering out irrelevant words.
[0070] Step 3: After data standardization, the dataset is divided into a training set and a test set in a ratio of 8:2. The training set is used to train the model, and the test set is used to evaluate the model's generalization ability.
[0071] Step 4: Multi-scale discrete wavelet transform processing: using the Daubechies wavelet transform method with 4-level decomposition: scaling equation: Wavelet equation: Where h k For the low-pass filter coefficients, g k The high-pass filter coefficients are used to reconstruct seismic and sedimentary pattern features from the seismic and sedimentary pattern data volumes. A three-level Coiflets wavelet transform method is employed. The scaling equation is as follows: Wavelet equation: Where L = 6N-1 is the filter length (N is the wavelet order), sedimentary numerical simulation features and well logging features are reconstructed from sedimentary numerical simulation images and well logging data; a two-layer orthogonal basis matrix wavelet transform method is used to vectorize the geological text features;
[0072] Step 5: Construct the CLIP-Stable Diffusion network model ( Figure 2 It mainly consists of five parts: text encoder, image encoder, U-net diffusion generator, image decoder, and text decoder;
[0073] Step 6: Construct a CLIP-Stable Diffusion network model, which mainly includes five parts: a text encoder, an image encoder, a U-net diffusion generator, an image decoder and a text decoder. The text encoder specifically further includes a tokenizer, an embedding layer, a Transformer encoding layer, and a projection layer. The tokenizer adopts the BPE (Byte-Pair Encoding) algorithm to split the text into high-frequency sub-word units (such as "tide-dominated" + "delta" + "deposition"); the embedding layer converts discrete sub-word units into 768-dimensional continuous vectors (Base model) through semantic vector mapping, each discrete word unit corresponds to a learnable embedding parameter, and through a position encoding mechanism, learnable position embeddings (non-fixed trigonometric function encoding) assign an independent vector to each discrete word unit position, so that the model perceives the sequence order; the Transformer encoding layer splits the input vector into multiple groups of Q / K / V vectors through a multi-head attention mechanism, calculates attention in different semantic spaces in parallel, inserts residual connections to directly add the input to the output to avoid information loss, and performs layer normalization to normalize the input of each layer, which accelerates the network convergence speed and improves the generalization ability of the model. Finally, an activation function is added, which introduces a nonlinear transformation to map the parameter features to the corresponding output end, and performs nonlinear mapping on the output result; the projection layer linearly compresses the 768-dimensional text features output by the Transformer encoding layer to a 512-dimensional shared space, to align the dimension with the image encoder.
[0074] In the text encoder, the vocabulary size of the tokenizer is set to 49,408, the input length is 77, the discrete sub-word embedding dimension of the embedding layer is D_t=768, the position encoding length P is aligned with the input length; in the Transformer encoding layer, the number of single-layer attention heads is H=12, the middle layer dimension is M=4×D_t=3072, the attention dimension is D_a=D_t / H, the total number of inserted multi-head attention layers is L=12, and the output dimension of the projection layer is K=512. In the image encoder, the input image size of the image splitter is 512×512, the patch size is 16×16, the number of patches M=(S / G)²=16; in the fully connected layer, the input dimension is D_patch=3×16×16, the output dimension D_embed=D_patch remains consistent with the input, and the parameter settings of the Transformer encoding layer and the projection layer are consistent with those in the text encoder. After each convolution operation, the number of channels of the feature map becomes twice the original number.
[0075] Step 7: The image encoder includes a splitter, a fully connected layer, a Transformer encoder layer, and a projection layer. The splitter segments the input image into fixed-size non-overlapping blocks, converting the two-dimensional image into serialized data to adapt to the sequence processing capabilities of the Transformer. The fully connected layer linearly transforms the input vector through a weight matrix, mapping each image block processed by the splitter into a D-dimensional vector, realizing the vectorized representation of local image features, equivalent to the primary feature extraction of traditional convolution. The Transformer encoder layer and projection layer are the same as those in the text encoder. Finally, the image is linearly mapped to a 512-dimensional shared space through the projection layer, aligning with the dimensions of the text encoder.
[0076] Step 8: The U-net diffusion generator consists of diffusion training downsampling, diffusion generation upsampling, and LOSS update modules. The downsampling part specifically includes convolution operations and a Transformer part. In the convolutional layer, convolution kernels are used to extract data features. The convolution kernels are weight matrices with a specific distribution, which weight the original data according to a specific distribution to achieve image feature extraction. The upsampling part includes deconvolution operations and a Transformer part. Deconvolution restores the feature map size of the upsampling stage to the size of the input image. The concatenation operation connects the feature results obtained from each convolution operation in the downsampling stage with the corresponding deconvolution results in the upsampling stage to achieve multi-scale prediction. Gaussian noise following a standard normal distribution is inserted at each time step in the diffusion training downsampling stage. The noise proportion is gradually increased through weight coefficients to achieve the effect of continuously adding noise to the image in this stage. In the diffusion generation upsampling stage, a denoising diffusion probability model algorithm is used.
[0077]
[0078] Where α t ,β t It is the preset noise dispatch coefficient (α) t =1-β t ), c is the text condition, σ tThis is the variance term controlling randomness, and ∈ represents newly added random noise (inserted only during training), achieving continuous denoising of the image in this stage. Throughout the U-net diffusion generation process, the Q / K / V matrix output by the text encoder is extracted and fed into the U-net diffusion generation model as a condition through cross-attention, in order to obtain the interrelationship information at different positions in the text prompts. The LOSS update module calculates the loss between the generated image and the real image for each training iteration using the root mean square error method during network training and incorporates it as a condition into the diffusion generation upsampling stage, dynamically fine-tuning the network to continuously optimize the generated images until sufficient training iterations or sufficient training accuracy are achieved. The image decoder and text encoder operate on the same principle as their corresponding decoders, but in opposite directions, gradually expanding the spatial dimension to convert the latent variables after denoising in the diffusion process into images and text.
[0079] Step 9: In the text encoder of the CLIP-Stable Diffusion network model, the vocabulary size of the word segmenter is set to V, the input length is N, the discrete word embedding dimension of the embedding layer is D_t = 768, the positional encoding length P is aligned with the input length, in the Transformer encoding layer, the number of attention heads in a single layer is H = 12, the intermediate layer dimension is M = 4 × D_t, the attention dimension is D_a = D_t / H, the total number of multi-head attention layers is L = 12, and the output dimension of the projection layer is K = 512. In the image encoder, the input image size of the image segmenter is S × S, the block size is G × G (typically 16), and the number of blocks is M = (S / G). 2 In the fully connected layer, the input dimension is D_patch = 3 × G × G, and the output dimension is D_embed = D_patch, consistent with the input. The Transformer encoding layer and projection layer have the same parameter settings as in the text encoder. After each convolution operation, the number of channels in the feature map doubles.
[0080] Step 10: The diffusion training downsampling part of the U-net diffusion generator. The first downsampling uses 3 convolutional layers with a kernel size of K×K and 1 Transformer block, with a stride of S_d. Each convolutional layer is followed by a normalization layer and an activation function. The number of text embedding cross-attention heads is R, and the temporal embedding dimension is T_dim. A total of Y downsampling operations are performed.
[0081] Step 11: The upsampling part of the U-net diffusion generator's diffusion training. The first upsampling uses three deconvolutional layers with a kernel size of K×K and one Transformer block, with a stride of S_d. Each deconvolutional layer is followed by a normalization layer and an activation function. The number of cross-attention heads for text embedding is R, the temporal embedding dimension is T_dim, and the skip connection dimension D_skip is associated with the upsampling part to preserve high-frequency details. A total of Y upsampling operations are performed. After each deconvolution operation, the number of channels in the feature map becomes half of its original value.
[0082] Step 12: Steps 5 to 11 involve constructing the CLIP-Stable Diffusion network model. Then, the seismic reconstruction data, sedimentary pattern reconstruction data, sedimentary numerical simulation feature reconstruction data, well logging feature reconstruction data, and text feature vectors from Step 4 are input into the constructed CLIP-Stable Diffusion network model to complete model training.
[0083] Step 13: Multi-scale inverse wavelet transform processing: According to step 4, the geological reconstruction data output by the model is processed by inverse wavelet transform according to their respective scales to output simulated seismic data, sedimentary simulation learning mode, sedimentary numerical simulation, well logging data, and geological text description.
[0084] Step 14: For fine-tuning the network parameters, refer to steps 9 to 11. Continue propagating forward through the calculations and processing of each layer to obtain the prediction results. Calculate the value of the loss function based on the actual results and the prediction results (the loss function uses the mean squared error). Where X obs,i X is the actual result of the i-th iteration. model,i Given the model training result of the i-th iteration, the gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm. The weights and biases of the network are updated based on the gradient. The training is set to 1000 epochs, with each epoch consisting of 50 loops, until the model achieves optimal performance on the validation set.
[0085] Step 15: When the difference between the reconstructed structure output by the CLIP-Stable Diffusion network model and the actual result is less than 0.01, the fine-tuning of the geological reconstruction data model is completed; otherwise, return to the training network to continue training.
[0086] Step 16: Input the test set data into the network model after wavelet transform.
[0087] The CLIP-Stable Diffusion network model was tested and its generation accuracy was evaluated to be over 80%. Finally, the geological multimodal data fusion prediction based on the CLIP-Stable Diffusion network using discrete wavelet transform was completed, and the network model was output.
[0088] In another embodiment of the present invention, a multimodal geological feature fusion system is provided. This system can be used to implement the above-described multimodal geological feature fusion method, specifically including:
[0089] The multimodal data input interface module is equipped with a dedicated data access port for receiving 3D seismic data volumes, 3D sedimentary simulation data volumes, 2D sedimentary numerical simulation images, various types of well logging curves, and structured geological text descriptions. The geological text contains key elements for lithological classification and sedimentary facies delineation, ensuring the integrity of the original data input.
[0090] The data standardization and partitioning module performs image data standardization to eliminate dimensional differences, and uses a geological dictionary to drive word segmentation and filter redundant information for text data. After standardization, the data is divided into training and testing sets according to a preset ratio. The training set is used for model parameter optimization, while the testing set is used for final generalization capability verification.
[0091] The multi-scale feature reconstruction module incorporates a differentiated wavelet processor: seismic and sedimentary model data undergo multi-level feature decomposition using Daubechies wavelets, sedimentary numerical simulation and well logging data are extracted using Coiflets wavelets, and geological text is mapped to semantic vectors through orthogonal basis matrices. Each processor adaptively adjusts the decomposition level according to the data type.
[0092] The CLIP-StableDiffusion engine integrates a dual-path encoding architecture: the text encoding unit performs word segmentation embedding and semantic vector transformation, while the image encoding unit performs block processing and spatial feature extraction. The core U-net diffusion generator includes a two-stage process of downsampling noise injection and upsampling semantic-guided denoising, and uses a cross-attention mechanism to fuse geological text conditions to control the generation direction.
[0093] The training control and output generation module dynamically monitors the training loss value and employs a threshold termination mechanism driven by geological principles to control the iteration process. Training automatically stops when the multimodal joint error falls below the threshold experienced by geological experts. The output data is then restored through inverse wavelet transform to form fused seismic data volumes, sedimentary simulation images, well logging curves, and structured geological reports.
[0094] In another embodiment of the present invention, a terminal device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used for the operation of a multimodal geological feature fusion method.
[0095] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory). This computer-readable storage medium is a memory device in a terminal device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the terminal device and extended storage media supported by the terminal device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device.
[0096] One or more instructions stored in a computer-readable storage medium can be loaded and executed by a processor to implement the corresponding steps of the multimodal geological feature fusion method in the above embodiments; one or more instructions in the computer-readable storage medium are loaded and executed by a processor.
[0097] 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.
[0098] 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, and 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0099] 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.
[0100] These computer program instructions may 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.
[0101] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the implementation methods of the present invention, and should be understood that the scope of protection of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of the present invention.
Claims
1. A multimodal geological feature fusion method based on discrete wavelet transform and CLIP-StableDiffusion model, characterized in that, Includes the following steps: S1: Input multimodal geological data, including 3D seismic data, 3D sedimentary simulation data, 2D sedimentary numerical simulation images, well logging curves, and geological text descriptions; S2: Standardize the data for each type: image data is normalized using Z-score, and text data is filtered using WordPiece word segmentation. S3: Divide the training set and the test set; S4: Reconstructing feature data using multi-scale discrete wavelet transform: S41: Seismic data and sedimentary model data were reconstructed using Daubechies wavelet transform; S42: Sedimentary numerical simulation images and well logging data were reconstructed using Coiflets wavelet transform; S43: Geological texts are vectorized using wavelet transform of orthogonal basis matrices; S5: Construct the CLIP-StableDiffusion network model, including: text encoder, image encoder, U-net diffusion generator, image decoder, and text decoder; Configure the text encoder structure, including: tokenizer, embedding layer, Transformer encoding layer, and projection layer; Configure the image encoder structure, including: image splitter, fully connected layer, and Transformer encoding layer; Configure the U-net diffusion generator, including: downsampling module and upsampling module; The downsampling module contains convolutional layers and Transformer blocks, and inserts Gaussian noise; The upsampling module contains deconvolutional layers and Transformer blocks to perform denoising diffusion; S6: Input the reconstructed data into the CLIP-StableDiffusion network model for training, and optimize the generated results through the diffusion noise addition / denoising process; S7: Calculate the root mean square error loss. Terminate training when the root mean square error (RMSE) between the output and the actual data is less than the preset error; otherwise, return to S6. S8: Perform multi-scale wavelet inverse transform on the model output to generate fused 3D seismic data, 3D sedimentary simulation data, 2D sedimentary numerical simulation images, well logging curves, and geological text; The type of discrete wavelet transform is dynamically selected based on the data type: 3D seismic data and 3D sedimentary simulation data were reconstructed using a 4-layer Daubechies wavelet transform; The scaling equation is: ; The wavelet equation is: ; in These are the coefficients of the low-pass filter. These are the coefficients of the high-pass filter. The wavelet order is... For filter coefficient index; Two-dimensional sedimentary numerical simulation images and well logging data were reconstructed using a 3-layer Coiflets wavelet transform. The scaling equation is: ; The wavelet equation is: ; in This is the filter length; Geological texts are vectorized using wavelet transform with two-layer orthogonal basis matrices.
2. The method according to claim 1, characterized in that, The structure of the U-net diffusion generator is as follows: Downsampling part: 4 iterations, each containing 3 3×3 convolutional layers and 1 Transformer block, with 4 cross-attention heads and 512 temporal embedding dimensions; Upsampling: 4 iterations, each containing 3 x 3 deconvolutional layers and 1 Transformer block, preserving high-frequency details through skip connections.
3. The method according to claim 2, characterized in that, The upsampling diffusion process employs a denoised diffusion probability model: ; in, For noise dispatch coefficient, It is the variance term that controls for randomness. It is newly added random noise. For text conditions.
4. The method according to claim 1, characterized in that, The model training uses the Adam optimizer, which updates the network weights and biases according to the gradient, and sets 1000 epochs of training. Each epoch of training includes 50 loops until the model achieves optimal performance on the validation set.
5. A multimodal geological feature fusion system based on discrete wavelet transform and CLIP-StableDiffusion model, characterized in that, This system can be used to implement the multimodal geological feature fusion method according to any one of claims 1 to 4, specifically including: The multimodal data input interface module is equipped with a dedicated data access port for receiving 3D seismic data volumes, 3D sedimentary simulation data volumes, 2D sedimentary numerical simulation images, various types of well logging curves, and structured geological text descriptions. The geological text contains key elements of lithological classification and sedimentary facies division, ensuring the integrity of the original data input. The data standardization and partitioning module is used to perform image data standardization processing to eliminate dimensional differences, and text data is segmented and filtered for redundant information using a geological dictionary. After standardization, the data is divided into training set and test set according to a preset ratio. The training set is used for model parameter optimization, and the test set is responsible for the final generalization ability verification. The multi-scale feature reconstruction module has a built-in differential wavelet processor: seismic and sedimentary model data are decomposed into multi-level features using Daubechies wavelet, sedimentary numerical simulation and well logging data are extracted using Coiflets wavelet, and geological text is mapped into semantic vectors through orthogonal basis matrix; each processor adaptively adjusts the decomposition level according to the data type. The CLIP-StableDiffusion module integrates a dual-path encoding architecture: the text encoding unit implements word segmentation and embedding and semantic vector transformation, and the image encoding unit performs block processing and spatial feature extraction; the core U-net diffusion generator includes a dual-stage downsampling noise injection and upsampling semantic-guided denoising, and controls the generation direction by fusing geological text conditions through a cross-attention mechanism. The training control and output generation module is used to dynamically monitor the training loss value and adopts a threshold termination mechanism driven by geological laws to control the iteration process. When the multimodal joint error is lower than the threshold of geological experts' experience, the training is automatically stopped. The output data is restored to the fused seismic data volume, sedimentary simulation image, well logging curve and structured geological report through wavelet inverse transform.
6. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the multimodal geological feature fusion method according to any one of claims 1 to 4.
7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the multimodal geological feature fusion method according to any one of claims 1 to 4.
Citation Information
Patent Citations
Cross-scale nesting and fusion modeling method for multi-scale geological model
CN120579438A
