An attribute fusion method and system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder

Abstract

This invention discloses an attribute fusion method and system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder, belonging to the fields of artificial intelligence and seismic exploration. The method includes: constructing an unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder (UHM-AWFFA) network, where the encoder uses a multi-branch independent encoder to extract deep independent features of multiple seismic attributes and dynamically fuses them through a weighted feature fusion layer; the decoder performs hierarchical reconstruction of the fused features, simultaneously outputting the reconstruction results of each attribute and the fused attribute profile; training is performed using an interactive reconstruction loss function, and fine-tuning is performed using a ternary multimodal adaptive dynamic transfer learning (TE-MADT) strategy, including a gradient-driven adaptive thawing mechanism, a multimodal collaborative transfer strategy, and a cross-modal knowledge distillation mechanism. This invention achieves nonlinear intelligent fusion of seismic attributes, effectively uncovers deep coupling relationships between attributes, and improves the geological interpretation accuracy of the fused profile.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence technology and seismic exploration, specifically to an attribute fusion method and system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder. Background Technology

[0002] Multi-attribute fusion is an important method in seismic data processing and interpretation. Different attributes represent different characteristics in seismic data. Multi-attribute fusion can achieve complementary advantages between attributes, thereby providing an accurate data foundation for subsequent geological interpretation.

[0003] However, traditional seismic attribute fusion methods mostly employ linear fusion, such as image-based overlay, RGB fusion, and principal component analysis. These methods have limitations in handling complex geological structures, as seismic attributes exhibit nonlinear relationships with geological structures and between attributes themselves, which linear fusion struggles to accurately characterize. Furthermore, directly fusing seismic attributes is susceptible to noise interference, resulting in irrelevant information in the fused profile. This hinders the effective uncovering of deep coupling relationships between attributes and with geological structural features, thus interfering with subsequent geological interpretation.

[0004] In recent years, with the rapid development of computing theory and hardware, artificial intelligence has achieved leapfrog development across various industries, providing robust solutions to many complex problems that traditional methods cannot address. Deep learning, with its powerful nonlinear modeling capabilities, has offered new ideas and methods for seismic attribute fusion. By utilizing multi-layered neural networks to automatically extract features from data, it learns more representative high-dimensional features.

[0005] Multimodal neural networks (MMNs) are neural network structures capable of simultaneously processing data from different modalities. For data of different modalities, independent features are extracted through separate network branches, and then fused using a fusion layer. Since different seismic attributes have different physical meanings and represent different features, they can be considered as different types of data. Currently, MNNs are widely used in fields such as vision, medicine, and seismic exploration, and have significant advantages in seismic attribute fusion, especially in processing multimodal data and complex geological features, where they can achieve nonlinear fusion of seismic attributes.

[0006] However, most existing methods for attribute fusion using deep learning are single-level fusions, such as feature-level fusion or decision-level fusion, failing to fully combine the advantages of different levels of fusion. Feature-level fusion integrates features extracted from each branch, enabling the network to learn deep correlations between different attributes, but it ignores the independence of each attribute, causing interference between attributes. Decision-level fusion combines the reconstruction results of each branch at the output stage. While it retains the independent information features of each attribute, it is insufficient in mining deep attribute correlations and cannot fully utilize the complementary information between attributes. Therefore, how to construct a network framework that simultaneously achieves feature-level fusion and decision-level fusion, taking into account the advantages of both, so as to leverage the ability of feature-level fusion to deeply mine the correlations between attributes and improve the network's expressive power, while using decision-level fusion to retain the independent information of each attribute and avoid the loss of feature information that may be caused by feature-level fusion. For the future development direction of earthquake attribute fusion, achieving accurate and efficient multimodal feature fusion is an urgent problem to be solved. Summary of the Invention

[0007] To extract the independent features of each seismic attribute and achieve feature-level fusion and decision-level fusion, thereby realizing attribute fusion, this invention proposes an attribute fusion method and system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder.

[0008] To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides an attribute fusion method based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder, comprising: Obtain various seismic attribute data about the target area; An unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder UHM-AWFFA network is constructed, the network including an encoder and a decoder; The various seismic attribute data are input into the UHM-AWFFA network, wherein: In the encoder, a multi-branch independent encoder is used to extract depth features from different seismic attribute data independently to obtain multiple independent features; and a weighted feature fusion layer is used to dynamically fuse the multiple independent features to generate a weighted fused feature. In the decoder, a multi-branch decoder is used to perform hierarchical reconstruction of the weighted fusion features, reconstructing each type of seismic attribute data separately; one branch decoder decodes the weighted fusion features to obtain the fused seismic attribute profile; The training process of the UHM-AWFFA network is iteratively optimized using the interactive reconstruction loss function until the loss function converges. The multiple seismic attribute data to be fused are input into the trained UHM-AWFFA network, and the fused seismic attribute profile is output.

[0009] Furthermore, each feature extraction branch in the encoder includes multiple convolutional modules connected in sequence. Each convolutional module consists of a convolutional layer, a batch normalization layer, and a ReLU activation function, and a Dropout layer is connected after the first convolutional layer. The weighted feature fusion layer assigns weights to features extracted from different coding branches through an adaptive weighting mechanism; The decoder includes a shared decoding layer and multiple branch decoding layers; the shared decoding layer is used to integrate weighted fusion features from the encoder; the branch decoding layer includes an attribute reconstruction branch with the same number of feature extraction branches and a fusion attribute output branch; each attribute reconstruction branch receives the context-enhanced fusion features output by the shared decoding layer and reconstructs a reconstructed profile of the corresponding input seismic attribute; the fusion attribute output branch receives the context-enhanced fusion features and directly outputs the fused seismic attribute profile.

[0010] Furthermore, the interaction reconstruction loss function is expressed as: in, and These represent the input and refactored attributes on different branches. This is the output of the fourth decoding layer, where N represents the total number of training samples.

[0011] Furthermore, the method also includes: after obtaining the trained UHM-AWFFA network as a pre-trained model, applying the ternary multimodal adaptive dynamic transfer learning (TE-MADT) strategy to fine-tune the pre-trained model to obtain a final fusion model adapted to the geological conditions of the target area; the TE-MADT strategy includes a gradient-driven adaptive thawing mechanism, a multimodal collaborative transfer strategy, and a cross-modal knowledge distillation mechanism.

[0012] Furthermore, the adaptive thawing mechanism dynamically determines the thawing timing of the next layer by monitoring the gradient stability of the thawed layer. The adaptive thawing criterion is defined as follows: Let the gradient of the Lth layer at the t-th training step be... Define the gradient stability index S L : Where W is the size of the sliding window. The mean of the gradient within the window; The thawing of the next layer L-1 is triggered when the deepest thawed layer L meets the following conditions: in, As the stability threshold, This is the gradient magnitude threshold; When unfreezing a new layer, dynamically set the learning rate based on the distance between that layer and the input layer: Where L is the total number of layers. As the attenuation factor, It is a regulating factor.

[0013] Furthermore, the multimodal cooperative migration strategy includes: Calculate the fitness index A for each mode m. m The calculation formula is: in, The reconstruction loss for this modality on the new task, The baseline loss for the pre-trained model; A collaborative unfreezing mechanism is adopted, and the unfreezing of the Lth layer is triggered only when the fitness index of all modes in the Lth layer exceeds the preset fitness threshold. Introducing a mode balance regularization term into the loss function ,in, It is the variance function. This is the balance coefficient.

[0014] Furthermore, the cross-modal knowledge distillation mechanism includes: using a modality that has quickly adapted to the new task as the teacher modality and a modality that adapts slowly as the student modality, and using the teacher modality to guide the feature learning of the student modality, so as to promote the collaborative transfer of knowledge and consistent feature expression between modalities.

[0015] Secondly, an attribute fusion system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder includes: The data acquisition module is used to acquire various seismic attribute data; The network construction module is used to build an unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder UHM-AWFFA network to perform multi-branch independent feature extraction, adaptive weighted feature-level fusion, hierarchical reconstruction and decision-level fusion of multimodal seismic attributes; The training module uses the interactive reconstruction loss function to perform unsupervised training on the UHM-AWFFA network, resulting in a pre-trained fusion model. The transfer learning module is used to perform gradient-driven adaptive unfreezing, multimodal collaborative transfer, and cross-modal knowledge distillation fine-tuning on the pre-trained fusion model to obtain the final fusion model adapted to the target region. The fusion module is used to input the seismic attributes to be fused into the final fusion model and output the nonlinearly fused seismic attribute profile.

[0016] Thirdly, an electronic device 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 method.

[0017] Fourthly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method.

[0018] Based on the above technical solution, the embodiments of the present invention can produce at least the following technical effects: This invention constructs an unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder (UHM-AWFFA) and adopts a hybrid multimodal fusion strategy of feature-level interactive fusion and decision-level integration. This overcomes the shortcomings of traditional linear fusion, which struggles to characterize nonlinear relationships, and the limitations of single-level fusion, which suffers from feature interference or insufficient correlation mining. It can deeply mine the potential coupling relationships between multiple seismic attributes while preserving the independent feature contributions of each attribute, significantly improving the geological interpretation accuracy of the fused profile. Furthermore, the proposed interactive reconstruction loss function effectively drives the network to fully extract the key features of each attribute by simultaneously constraining the reconstruction quality of each attribute and the consistency of the fused attributes. In addition, the introduced ternary multimodal adaptive dynamic transfer learning (TE-MADT) strategy, through gradient-driven adaptive thawing, multimodal collaborative transfer, and cross-modal knowledge distillation, can quickly adapt and transfer under new regions or new data conditions, enhancing the model's generalization ability and universality, reducing the dependence on manual parameter tuning, and providing an efficient and robust solution for the nonlinear intelligent fusion of seismic attributes under complex geological conditions. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0020] Figure 1 This is a flowchart of a method according to an embodiment of the present invention; Figure 2 This is a diagram of the unsupervised hybrid multimodal feature autoencoder network structure of the present invention; Figure 3 This is a graph showing the loss function results of an embodiment of the present invention; Figure 4This is a cross-sectional view of the multi-attribute nonlinear intelligent fusion of the UHM-AWFFA network of this invention; Figure 5 This is a schematic diagram of the system structure according to an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. In addition, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0022] like Figure 1 As shown, this embodiment discloses an attribute fusion method based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder, the method comprising: S100: Obtain multiple sets of seismic attribute data of different types for the target area.

[0023] S200: Construct an unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder (UHM-AWFFA) network and input the various seismic attribute data into the UHM-AWFFA network. The UHM-AWFFA network structure constructed in this embodiment is as follows: Figure 2 As shown, the network consists of an encoder and a decoder, and adopts a three-stage fusion architecture of multimodal feature extraction, weighted feature-level fusion, hierarchical reconstruction and decision-level fusion.

[0024] The encoder contains multiple independent coding branches, each corresponding to one of the number of seismic attributes to be fused. In this embodiment, three seismic attributes are preferably fused, corresponding to three independent coding branches to ensure the independence of feature extraction. Each branch's convolutional layer includes convolution, batch normalization, and the ReLU activation function. The feature dimension of each layer increases with the number of convolutional layers. Batch normalization reduces internal covariate shift, helping to accelerate the training process and improve model stability. The ReLU activation function, by introducing non-linearity, enhances the network's learning ability and alleviates gradient vanishing. A Dropout layer is added after the first convolutional layer. Dropout randomly drops some nodes in the network during training, enhancing the model's generalization ability and enabling the network to learn more robust and generalized features. Specific network parameters are shown in Table 1 below, where p represents (1, 2, 3).

[0025] The outputs of all independent encoding branches are fed into a weighted feature fusion layer. By weightedly fusing features extracted from different convolutional layers, the network can more comprehensively capture and integrate important information from seismic attribute data, forming a more representative feature representation and improving the attribute fusion effect. Through the aforementioned encoder structure, different and independent features can be extracted from different input seismic attributes, thereby enhancing the nonlinear feature extraction capability of the unsupervised hybrid multimodal network. This ensures that the network can comprehensively extract key features from seismic attributes, laying the foundation for subsequent attribute reconstruction and fusion.

[0026] Table 1 Encoder Network Parameters Taking the first convolutional module in encoder 1_1 as an example, its structure is Conv+BN+ReLU, with a 3×3 kernel and a stride of 2, ensuring that the feature extraction capability of the UHM-AWFFA network is maintained while reducing feature dimensionality. The learnable parameters of this module include weights and biases. The number of weights is related to the number of input and output channels, ensuring that the UHM-AWFFA network can flexibly adjust the feature representation. This structure extracts local features from attribute data while using batch normalization to improve the training stability of the model.

[0027] The weighted feature fusion layer in the encoder fuses features extracted from different coding branches through an adaptive weighting mechanism, assigning significant weights to key features during training. The fully connected layer maps the fused features to the final output space. The fully connected layer has a large number of weights (512×4096) and a bias parameter (512), reflecting the comprehensive utilization of features by the UHM-AWFFA network.

[0028] The main task of the decoder is to reconstruct and fuse the obtained weighted fused features. The decoder consists of a shared decoding layer and multiple branch decoding layers (decoding layer 1, decoding layer 2, decoding layer 3, and decoding layer 4), utilizing transposed convolutions (TransConv) and ReLU activation functions. The convolution kernel is 3×3 with a stride of 2. Specific network parameters are shown in Table 2, where C represents (1, 2, 3, 4). The main function of the shared decoding layer is to further integrate the weighted fused features from the encoder. By inputting the weighted fused features into the shared decoding layer, the network can effectively utilize the feature information of different attributes. Furthermore, during reconstruction, the feature representation after passing through the shared decoding layer provides richer contextual information, allowing each branch to consider more features during reconstruction. When processing complex seismic data, this shared decoding layer helps improve the reconstruction quality. The first three branch decoding layers reconstruct different seismic attributes, ensuring that the features of each attribute are fully extracted and utilized during network iteration, and ensuring that the weighted fused features contain key feature information of the three seismic attributes. If the features of a certain attribute are not effectively extracted, the seismic fusion attribute obtained by passing the weighted fusion features through the branch decoding layer (decoding layer 4) will lose important feature information, resulting in distortion of the fused profile. Moreover, through the reconstruction of each attribute by each decoding layer, and by utilizing the loss function feedback mechanism, the network is continuously optimized during the iteration process, promoting the encoder to further extract deep features from the branch attributes.

[0029] Table 2 Decoder Network Parameters S300: To achieve nonlinear intelligent fusion of multiple seismic attributes, an interactive reconstruction loss function is innovatively proposed and adopted to guide the training process during iterative optimization. This ensures the accuracy and effectiveness of seismic attribute reconstruction and fusion results, and comprehensively evaluates the network's performance. The specific loss calculation is as follows: in, and These represent the input and refactored attributes on different branches. This is the output of the fourth decoding layer, where N represents the total number of training samples. (Part 1) The mean square error between the reconstructed attributes and the true attributes in the first three decoding layers is calculated. By minimizing this loss, the network can effectively and independently extract and utilize the features of each attribute, ensuring the quality of the seismic attributes reconstructed in the first three decoding layers. Part Two This is the loss function for fused attributes. It calculates the mean squared error between the reconstructed weighted fused features and the reconstruction results of the first three decoding layer branches, thus ensuring that the UHM-AWFFA network can fully utilize the feature information of different attributes. This promotes consistency in the UHM-AWFFA network when processing different modalities, ensuring effective attribute fusion between different modalities.

[0030] Example of UHM-AWFFA network application: To evaluate the feasibility and effectiveness of an unsupervised hybrid multimodal feature autoencoder, a deep reflection seismic profile of a certain region was selected for multi-attribute nonlinear intelligent fusion. The optimized seismic attributes of various types were segmented to construct a training set. The aforementioned UHM-AWFFA network and interactive reconstruction loss function were used for training and optimization, and the loss function results are shown below. Figure 3 As shown. By Figure 3 It can be seen that the loss value drops rapidly in the initial stage, indicating that the UHM-AWFFA network model quickly learns the main features in the initial stage, which has a significant effect on reducing the loss and is in a "rapid convergence period," indicating that the learning rate and initial parameter settings have a positive impact on the convergence speed. A slight fluctuation occurs when the iteration reaches 2300, with a noticeable "peak" on the loss function curve. This phenomenon may be due to the large variance of some batches in the training sample data, causing instantaneous fluctuations in the loss. This gradually stabilizes and converges as training progresses. As the number of iterations increases, the loss function curve approaches 0 and remains at a low level with slight fluctuations, indicating that the UHM-AWFFA network model is approaching saturation, the loss almost no longer decreases, the curve remains stable, and the model has completed fitting the training data. The final attribute fusion profile is as follows: Figure 4 As shown.

[0031] S400: To enhance the model's cross-region generalization ability, this invention proposes a Triple-Element Multi-Modal Adaptive Dynamic Transfer Learning (TE-MADT) strategy to fine-tune the pre-trained model. This strategy includes three parts: a gradient-driven adaptive unfreezing mechanism, a multi-modal collaborative transfer strategy, and a cross-modal knowledge distillation mechanism.

[0032] (1) Gradient-Driven Adaptive Unfreezing (GDAU) TE-MADT proposes an adaptive thawing criterion based on gradient flow characteristics. This mechanism dynamically determines the thawing timing of the next layer by monitoring the gradient stability of the thawed layer. The adaptive thawing criterion is defined as follows: Let the gradient of the Lth layer at the t-th training step be... Define the gradient stability index SL: Where W is the size of the sliding window. This represents the mean gradient within the window.

[0033] Thawing trigger condition: The thawing of the next layer L-1 is triggered when the deepest thawed layer L meets the following conditions: in, As the stability threshold, This is the gradient magnitude threshold. This mechanism ensures that deeper layers are only unfrozen after the current layer has fully adapted to the new task, avoiding feature destruction caused by premature unfreezing.

[0034] Adaptive learning rate adjustment: When unfreezing a new layer, the learning rate is dynamically set based on the distance between that layer and the input layer. Where L is the total number of layers. As the attenuation factor, It is a regulating factor.

[0035] Through a gradient-driven adaptive unfreezing mechanism, the model can automatically adjust the transfer learning pace according to the learning difficulty of new data, without the need for manual parameter tuning.

[0036] (2) Multi-Modal Collaborative Transfer (MMCT) MMCT is designed to address the issue of inconsistent adaptation speeds among different modes during the migration process, and mainly includes: Modal fitness assessment: For the m-th mode, define its fitness index A. m : in, The reconstruction loss for this modality on the new task, This is the baseline loss of the pre-trained model.

[0037] Collaborative unfreezing strategy: To ensure the consistency of multimodal feature fusion, a collaborative unfreezing mechanism is adopted, that is, unfreezing only occurs when the fitness of all modalities in layer L exceeds a threshold. Only when the time is right will the unfreezing of that layer be triggered.

[0038] Modal balance loss: Introducing a modal balance regularization term to prevent a single mode from dominating during migration. in, It is the variance function. This is the balance coefficient.

[0039] (3) Cross-modal knowledge distillation To address the asynchronous adaptation issue in multimodal learning during the transfer process, this invention introduces a cross-modal knowledge distillation (CMKD) mechanism. This mechanism aims to guide slower-adapting modalities ("student modalities") by having modalities that have quickly adapted to the new task ("teacher modalities") guide those that are slower to adapt, thereby promoting collaborative knowledge transfer and consistent feature expression among modalities, enhancing the consistency of multimodal fusion, suppressing the dominance of any particular modality, and ultimately comprehensively improving the model's robustness and generalization ability when facing unknown geological conditions and noise interference.

[0040] S500: Input the multiple seismic attribute data to be fused into the trained UHM-AWFFA network and output the fused seismic attribute profile.

[0041] like Figure 5 As shown, the implementation method discloses an attribute fusion system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder, including: The data acquisition module is used to acquire various seismic attribute data; The network construction module is used to build an unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder UHM-AWFFA network to perform multi-branch independent feature extraction, adaptive weighted feature-level fusion, hierarchical reconstruction and decision-level fusion of multimodal seismic attributes; The training module uses the interactive reconstruction loss function to perform unsupervised training on the UHM-AWFFA network, resulting in a pre-trained fusion model. The transfer learning module is used to perform gradient-driven adaptive unfreezing, multimodal collaborative transfer, and cross-modal knowledge distillation fine-tuning on the pre-trained fusion model to obtain the final fusion model adapted to the target region. The fusion module is used to input the seismic attributes to be fused into the final fusion model and output the nonlinearly fused seismic attribute profile.

[0042] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An attribute fusion method based on unsupervised mixed multi-modal and ternary adaptive transfer auto-encoder, characterized in that, include: Obtain various seismic attribute data about the target area; An unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder UHM-AWFFA network is constructed, the network including an encoder and a decoder; The various seismic attribute data are input into the UHM-AWFFA network, wherein: In the encoder, a multi-branch independent encoder is used to extract depth features from different seismic attribute data independently to obtain multiple independent features; and a weighted feature fusion layer is used to dynamically fuse the multiple independent features to generate a weighted fused feature. In the decoder, a multi-branch decoder is used to perform hierarchical reconstruction of the weighted fusion features, reconstructing each type of seismic attribute data separately; one branch decoder decodes the weighted fusion features to obtain the fused seismic attribute profile; The training process of the UHM-AWFFA network is iteratively optimized using the interactive reconstruction loss function until the loss function converges. The multiple seismic attribute data to be fused are input into the trained UHM-AWFFA network, and the fused seismic attribute profile is output.

2. The attribute fusion method based on unsupervised hybrid multi-modal and ternary self-adaptive transfer auto-encoder according to claim 1, characterized in that, Each feature extraction branch in the encoder includes multiple convolutional modules connected in sequence. Each convolutional module consists of a convolutional layer, a batch normalization layer, and a ReLU activation function, and a Dropout layer is connected after the first convolutional layer. The weighted feature fusion layer assigns weights to features extracted from different coding branches through an adaptive weighting mechanism; The decoder includes a shared decoding layer and multiple branch decoding layers; the shared decoding layer is used to integrate weighted fusion features from the encoder; The branch decoding layer includes an attribute reconstruction branch with the same number of feature extraction branches and a fusion attribute output branch; Each attribute reconstruction branch receives the context-enhanced fusion features output by the shared decoding layer and reconstructs the corresponding input seismic attribute profile; the fused attribute output branch receives the context-enhanced fusion features and directly outputs the fused seismic attribute profile. 3.The attribute fusion method based on unsupervised hybrid multi-modal and ternary self-adaptive transfer auto-encoder according to claim 1, characterized in that, The interactive reconstruction loss function is expressed as follows: ; wherein, and represent the input and reconstructed attributes on different branches, is the output of the fourth decoding layer, and N represents the total number of training samples.

4. The attribute fusion method based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder according to claim 1, characterized in that, The method further includes: after obtaining the trained UHM-AWFFA network as a pre-trained model, applying the ternary multimodal adaptive dynamic transfer learning (TE-MADT) strategy to fine-tune the pre-trained model to obtain a final fusion model adapted to the geological conditions of the target area; the TE-MADT strategy includes a gradient-driven adaptive thawing mechanism, a multimodal collaborative transfer strategy, and a cross-modal knowledge distillation mechanism.

5. The attribute fusion method based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder according to claim 4, characterized in that, The adaptive thawing mechanism dynamically determines the timing of thawing the next layer by monitoring the gradient stability of the thawed layer. The adaptive thawing criterion is defined as follows: Let the gradient of the Lth layer at the t-th training step be... Define the gradient stability index S L : ; Where W is the size of the sliding window. The mean of the gradient within the window; The thawing of the next layer L-1 is triggered when the deepest thawed layer L meets the following conditions: ; in, As the stability threshold, This is the gradient magnitude threshold; When unfreezing a new layer, dynamically set the learning rate based on the distance between that layer and the input layer: ; Where L is the total number of layers. As the attenuation factor, It is a regulating factor.

6. The attribute fusion method based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder according to claim 4, characterized in that, The multimodal cooperative migration strategy includes: A fitness indicator A is calculated for each modality m m with the formula: ; in, The reconstruction loss for this modality on the new task, The baseline loss for the pre-trained model; A collaborative unfreezing mechanism is adopted, and the unfreezing of the Lth layer is triggered only when the fitness index of all modes in the Lth layer exceeds the preset fitness threshold. Introducing a mode balance regularization term into the loss function ,in, It is the variance function. This is the balance coefficient.

7. The attribute fusion method based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder according to claim 4, characterized in that, The cross-modal knowledge distillation mechanism includes: using a modality that has adapted quickly to the new task as the teacher modality and a modality that adapts slowly as the student modality, and using the teacher modality to guide the feature learning of the student modality, so as to promote the collaborative transfer of knowledge and consistent feature expression between modalities.

8. An attribute fusion system based on unsupervised hybrid multimodal and ternary adaptive transfer autoencoder, used to implement the method described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire various seismic attribute data; The network construction module is used to build an unsupervised hybrid multimodal adaptive weighted feature fusion autoencoder UHM-AWFFA network to perform multi-branch independent feature extraction, adaptive weighted feature-level fusion, hierarchical reconstruction and decision-level fusion of multimodal seismic attributes; The training module uses the interactive reconstruction loss function to perform unsupervised training on the UHM-AWFFA network, resulting in a pre-trained fusion model. The transfer learning module is used to perform gradient-driven adaptive unfreezing, multimodal collaborative transfer, and cross-modal knowledge distillation fine-tuning on the pre-trained fusion model to obtain the final fusion model adapted to the target region. The fusion module is used to input the seismic attributes to be fused into the final fusion model and output the nonlinearly fused seismic attribute profile.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images