A seabed hydrate slope stability prediction method and system based on adversarial multi-source transfer learning

Abstract

This invention relates to a method for predicting the stability of seabed hydrate slopes based on adversarial multi-source transfer learning, comprising the following steps: S1, acquiring and constructing a multi-source domain slope dataset, including source domain data and target domain data; S2, preprocessing the samples in the multi-source domain dataset; S3, constructing an adversarial multi-source transfer learning model; S4, optimizing the parameters of the adversarial multi-source transfer learning model; S5, inputting the seabed hydrate slope samples to be evaluated into the parameter-optimized model, and outputting the slope stability prediction results and the contributions of key influencing factors. This invention can significantly improve the prediction accuracy and generalization ability of the safety factor of seabed hydrate slopes even under conditions of limited sample size, significant differences in the distribution of source and target domains, and strong interference from anomalous data. It has strong engineering applicability and can be widely applied to hydrate mining risk assessment, seabed slope stability analysis, and geological disaster monitoring and early warning.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of marine geotechnical engineering, geological hazard risk assessment and artificial intelligence, and specifically to a method and system for predicting the stability of seafloor hydrate slopes based on adversarial multi-source transfer learning. It can be used to output the predicted safety factor of seafloor hydrate slopes and provide the contribution or sensitivity ranking results of key influencing factors. Background Technology

[0002] Hydrates exist within the pores of marine sediments. Under altered temperature and pressure conditions, mining disturbances, and hydrate decomposition, the sediment structure and mechanical properties are prone to significant changes. This is accompanied by abnormally high pore water pressure and decreased effective stress, inducing a decline in slope stability and potentially leading to hazards such as submarine landslides, submarine foundation instability, tsunamis, and methane leaks. Predicting the stability of submarine hydrate slopes is a crucial foundation for safety assessments of hydrate mining and for monitoring and early warning of geological hazards.

[0003] In existing technologies, stability analysis of seabed slopes typically relies on numerical theoretical methods such as the limit equilibrium method and the strength-reduced finite element method. However, the high cost of obtaining seabed sediment parameters and the difficulty of experimental and in-situ testing often limit the available sample size for the target scenario (hydrate-containing sediment slopes). At the same time, historical data (source domain) from terrestrial slopes or conventional soil slopes differ significantly from data on seabed hydrate-containing sediment slopes (target domain) in terms of parameter distribution, mechanistic background, and noise structure.

[0004] In recent years, machine learning methods have been used for rapid prediction of slope stability, but they usually rely on a large number of identically distributed training samples. When there are insufficient samples in the target domain, significant differences in the distribution between the source and target domains, or outliers in the data, traditional machine learning models are prone to problems such as overfitting, weak generalization ability, and unstable prediction, making it difficult to meet the accuracy and robustness requirements of engineering applications. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method and system for predicting the stability of seafloor hydrate slopes based on adversarial multi-source transfer learning. This invention can effectively predict the stability of seafloor hydrate slopes under conditions of limited sample size and cross-domain distribution differences and abnormal data interference.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning includes the following steps: S1. Obtain and construct a multi-source domain slope dataset; the multi-source domain slope dataset includes source domain data and target domain data; S2. Perform data preprocessing on samples from multi-source domain datasets; S3. Construct an adversarial multi-source transfer learning model; S4. Optimize the parameters of the adversarial multi-source transfer learning model; S5. Input the sample of the seabed hydrate slope to be evaluated into the model after parameter optimization, and output the slope stability prediction results and the contribution of key influencing factors.

[0008] Preferably, step S1 specifically includes the following steps: S11. Source domain data construction: The source domain dataset is derived from land slope or conventional soil slope data, and includes at least slope geometric parameters, soil mechanical parameters and corresponding stability level data to form the source domain dataset. S12. Target Domain Data Construction: The target domain dataset consists of strength parameters obtained from triaxial or direct shear tests conducted in indoor similar material experiments, varying with hydrate state (hydrate saturation). S h Degree of hydrate decomposition S hd The changing patterns of pore pressure; establishing a pore pressure evolution model to calculate excess pore water pressure. u ex The input characteristic parameters are sampled within the set range of slope geometric parameters and working parameters; finally, the safety factor is calculated for each set of sampled parameters. F s This forms the target domain dataset.

[0009] Further preferred, the input features of the multi-source domain dataset include one or more of the following: soil unit weight. Cohesion c internal friction angle Elastic modulus E Poisson's ratio n Permeability coefficient k , slope height H slope angle SA , water depth D hydrate saturation S h Degree of hydrate decomposition S hd excess pore water pressure u ex .

[0010] Preferably, step S2 specifically includes the following steps: S21. Unit and dimension unification: Convert and unify the units and dimensions of two different source data, source domain data and target domain data, and perform consistent mapping on the fields; use completion or elimination strategies for missing data. S22. Abnormal Sample Identification and Processing: Abnormal data in the sample is identified and processed. Abnormal samples are judged based on the rules of geotechnical engineering physical constraints, and the identified abnormal samples are eliminated or corrected. S23. Feature standardization: Standardize or normalize the input features to eliminate the influence of different units on model training and improve the stability and convergence of the model training process. S24. Feature filtering or dimensionality reduction: Filter or reduce the dimensionality of the input features to obtain feature vectors for model training; S25. Training / Validation Set Partition and Target Domain Sample Augmentation: Divide the target domain samples into a training set and a validation set; when the number of target domain samples is insufficient, augment the target domain training set to alleviate the small sample problem and improve the model's generalization ability.

[0011] Further preferred methods include standardization using Z-score standardization; filtering or dimensionality reduction of input features using correlation analysis, principal component analysis (PCA), or feature selection methods based on variance thresholds; and target domain sample amplification methods including random oversampling or oversampling methods based on synthetic samples.

[0012] Preferably, step S3 specifically includes the following steps: S31. Constructing the model structure: The adversarial multi-source transfer learning model includes at least a feature extractor. G Domain discriminator C, regressor R Feature extractor G Used to map input feature vectors to latent feature representations. z = G ( x The domain discriminator C takes the latent feature representation z as input and determines the domain origin of a sample. Its output represents the result of whether the sample belongs to the source domain or the target domain. The regressor... R Used to represent latent features z Mapped to slope stability prediction results y = R ( z ); S32, Adversarial Learning Mechanism: Through adversarial training, the feature extractor... G By learning domain-invariant feature representations, the difference in feature distribution between the source and target domains can be reduced.

[0013] Furthermore, the adversarial training of the model employs multi-loss joint training optimization, and the joint loss function is expressed as:

[0014] in, This is the distribution difference loss, used to minimize the difference in the distribution of potential features between the source and target domains; This is the domain discrimination loss, used to train the domain discriminator to distinguish between the source domain and the target domain; To predict the loss, the stability label error is minimized; α , β , l These are the loss weighting coefficients, used to balance domain alignment and prediction accuracy.

[0015] Preferably, step S4 specifically includes the following steps: S41. Set a set of hyperparameters for the model to be optimized. The set of hyperparameters includes at least one or more of the following: learning rate, number of network layers, number of neurons per layer, dropout rate, batch size, number of training epochs, and loss weight coefficient. S42. Using the prediction error index on the validation set as the fitness function, an intelligent optimization algorithm is used to search for the optimal combination of hyperparameters that minimizes the fitness function of the validation set. S43. Train the adversarial multi-source transfer learning model using the optimal hyperparameter combination and output the optimal model.

[0016] More preferably, the intelligent optimization algorithm is the walrus optimizer algorithm. The walrus optimizer switches between global exploration and local development through an adaptive adjustment mechanism of danger signals and safety signals to achieve hyperparameter optimization. While training the adversarial multi-source transfer learning model, the validation set loss is monitored in real time, and training is automatically terminated when the validation loss no longer decreases within a preset number of rounds.

[0017] Preferably, step S5 specifically includes the following steps: S51. Obtain the input features of the seabed hydrate slope to be evaluated (e.g., , c , , E , n , k , H , SA , D , S h , S hd , u ex wait); S52. Input the above input features into the optimal model after parameter optimization, and output the predicted safety factor value. F s预 ; S53, Predicted value based on safety factor F s预The stability determination result is obtained by comparing with the preset threshold, and the contribution or sensitivity ranking result of the key influencing factors is output.

[0018] A stability prediction system for seafloor hydrate slopes based on adversarial multi-source transfer learning includes a dataset construction module, a data preprocessing module, an adversarial multi-source transfer learning module, a parameter optimization module, and a stability prediction output module. The dataset construction module acquires and constructs a multi-source domain dataset containing source domain data and target domain data. The data preprocessing module performs unit and dimension unification, anomaly and missing data processing, standardization, feature filtering or dimensionality reduction, and target domain sample partitioning on the multi-source domain data. The adversarial multi-source transfer learning module constructs and trains an adversarial transfer learning model composed of a feature extractor, a domain discriminator, and a regressor or classifier to achieve feature alignment between the source and target domains. The parameter optimization module adaptively optimizes the model's hyperparameters. The stability prediction output module inputs the seafloor hydrate slope samples to be evaluated into the parameter-optimized model and outputs the slope stability prediction results and the contributions of key influencing factors.

[0019] The advantages and positive effects of this invention are: This invention can significantly improve the prediction accuracy and generalization ability of the safety factor of seabed hydrate slopes even under conditions of limited sample size, significant differences in the distribution of source and target domains, and strong interference from anomalous data. It has strong engineering applicability and can be widely used in hydrate mining risk assessment, seabed slope stability analysis, and geological disaster monitoring and early warning. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0021] Figure 1 This is a flowchart of the prediction method of the present invention; Figure 2 This is a schematic diagram showing the interface of the prediction system of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0023] like Figure 1As shown, this invention provides a method for predicting the stability of seafloor hydrate slopes based on adversarial multi-source transfer learning. The hydrate is a methane-containing hydrate. The method includes the following steps: S1. Obtain and construct a multi-source domain slope dataset; the multi-source domain dataset includes source domain data and target domain data.

[0024] S11. Obtain one or more source domain datasets. The source domain dataset consists of historical cases of land slopes or conventional soil slopes, public databases, or engineering monitoring data. It includes at least slope geometric parameters, soil mechanical parameters, and corresponding stability level data to form the source domain dataset, which is used to provide prior knowledge for training the target domain model.

[0025] S12. Obtain the target domain dataset. The target domain dataset consists of strength parameters obtained from triaxial or direct shear tests conducted in indoor similar material experiments, as a function of hydrate saturation. S h Degree of hydrate decomposition S hd The changing patterns; combined with the pore pressure evolution model, the excess pore water pressure under hydrate decomposition or disturbance conditions is calculated. u ex Within a given range of slope geometric parameters and working parameters, the input characteristic parameters are sampled. Finally, the safety factor is calculated for each set of sampled parameters using either the limit equilibrium method or the strength reduction method. F s This forms the target domain dataset.

[0026] The input features of the multi-source domain dataset include one or more of the following: soil unit weight. Cohesion c internal friction angle Elastic modulus E Poisson's ratio n Permeability coefficient k , slope height H slope angle SA , water depth D hydrate saturation S h Degree of hydrate decomposition S hd excess pore water pressure u ex .

[0027] Specifically, based on indoor similar material tests, pore pressure evolution models, and engineering experience, the input characteristic parameters of the target domain slope in this embodiment are set as follows: Table 1 Input characteristic parameters of the target domain slope

[0028] The simplified Bishop method from the limit equilibrium method is used to calculate the slope stability. Safety factor. F s Defined as:

[0029] in: Let be the self-weight of the i-th soil strip, in kN; The pore water pressure acting on the bottom surface of the i-th soil strip includes excess pore water pressure. u ex The unit is kPa; The width of the i-th soil strip is in meters. The angle between the bottom surface of the i-th soil strip and the horizontal plane is expressed in degrees. This represents the cohesion of the i-th soil strip, in kPa. Let be the internal friction angle corresponding to the i-th soil strip, in degrees. The soil mass is divided into several strips using the strip method.

[0030] Substituting the above parameters into the calculation, the safety factor of the seabed hydrate slope under this working condition is obtained as follows: F s =1.12, this result is used as the true label value of the target domain dataset samples for subsequent model training and validation.

[0031] S2. Perform data preprocessing on samples from the multi-source dataset. S21. Unification of Units and Dimensions The units and dimensions of the source and target domain data from two different sources are converted and unified, and the fields are mapped in a consistent manner; a completion strategy or a removal strategy is adopted for missing data.

[0032] Specifically, first map the source domain data and the target domain data to the same set of fields: [ , c , , E , n , k , H , SA , D , S h , S hd , u ex The source domain data typically does not contain... S h ,S hd ,u exThese three items are set as missing labels (NA), and are treated uniformly during subsequent training. The unit is uniformly set to kN / m 3 If the source domain is kg / m 3 Then according to = ρg / 1000 conversion; c , u ex The unit is uniformly set to kPa; , SA Units are all in degrees. If it is in radians, multiply by 180 / π. E The unit is uniformly set to MPa; n Dimensionless; k Units are unified as m / s ; H , D The unit is uniformly m.

[0033] For continuous variables with a small number of missing values ​​(missing rate less than 5%), such as E The median was used for imputation; samples with missing key variables were directly removed; samples with naturally missing values ​​in the source domain were also removed. S h , S hd , u ex During training, a "missing mask" approach is used: a new three-dimensional mask feature is added, missing features are set to 0 and padded with 0, and existing features are set to 1 and the original value is retained, in order to avoid bias introduced by incorrect completion.

[0034] S22. Abnormal Sample Identification and Processing Abnormal data in the sample is identified and processed. Abnormal samples are judged based on the rules of geotechnical engineering physical constraints, and the identified abnormal samples are removed or corrected.

[0035] Specifically, if the input feature parameters are not within the corresponding range below, they are considered to be reference surface errors and will be uniformly rejected.

[0036] The specific range of input feature parameters is as follows:

[0037] S23, Feature Standardization Processing The Z-score standardization method is used to standardize or normalize the input features to eliminate the influence of different units on model training and improve the stability and convergence of the model training process.

[0038] Specifically, during the training phase, only the standardized parameters of the input features are fitted using the training set samples. The mean and standard deviation of each continuous feature are calculated, and the standardized parameters are used to perform a consistency transformation on the training set, validation set, and test set data to avoid information leakage and ensure the objectivity of model evaluation.

[0039] S24, Feature Filtering or Dimensionality Reduction The input features are filtered or reduced in dimensionality using correlation analysis, principal component analysis (PCA), or feature filtering methods based on variance thresholds to obtain feature vectors for model training.

[0040] Specifically, calculate the correlation coefficient matrix between features. If the absolute correlation coefficient between two features is >0.95, retain the feature with clearer engineering significance and remove the feature with the same source repetition density field.

[0041] S25. Training / Validation Partitioning and Target Domain Sample Amplification The target domain samples are divided into training and validation sets. When the number of target domain samples is insufficient, random oversampling or oversampling based on synthetic samples is used to augment the target domain training set to alleviate the small sample problem and improve the model's generalization ability.

[0042] Specifically, the target domain samples are randomly divided into training and validation sets, for example, in a ratio of 8:2, and a fixed random seed is used to ensure reproducibility. If the target domain training set has a small sample size (e.g., <200), random oversampling (e.g., SMOTE or random oversampling) is used to increase the target domain training set to a preset size (e.g., double the original size) to reduce the risk of overfitting caused by small samples.

[0043] S3. Constructing an adversarial multi-source transfer learning model S31. Constructing the model structure: The adversarial multi-source transfer learning model includes at least a feature extractor. G Domain discriminator C, regressor R Feature extractor G Used to map input feature vectors to latent feature representations. z = G ( x The domain discriminator C takes the latent feature representation z as input and determines the domain origin of a sample. Its output is the discrimination result indicating whether the sample belongs to the source domain or the target domain. In one implementation, this discrimination result can be represented in probabilistic form. Regressor R Used to represent latent features z Mapped to slope stability prediction results y = R ( z ); S32, Adversarial Learning Mechanism: Through adversarial training, the feature extractor... G By learning domain-invariant feature representations, the difference in feature distribution between the source and target domains can be reduced.

[0044] Adversarial training of the model employs joint training optimization using multiple loss functions, the joint loss function of which is expressed as:

[0045] in, This is the distribution difference loss, used to minimize the difference in the distribution of potential features between the source and target domains; This is the domain discrimination loss, used to train the domain discriminator to distinguish between the source domain and the target domain; To predict the loss, the stability label error is minimized; α , β , l These are the loss weighting coefficients, used to balance domain alignment and prediction accuracy.

[0046] During training, the domain discriminator C is made to distinguish domain labels as accurately as possible, while the feature extractor... G During training, by updating the parameters of the feature extractor with reverse gradients, the domain discriminator's ability to distinguish the domain source of the latent feature representation is reduced, thereby prompting the feature extractor to learn a domain-invariant feature representation that has a consistent representational ability for both the source and target domains.

[0047] Specifically, this step aims to establish a deep learning architecture capable of extracting common features between the source and target domains, thereby mitigating inter-domain distribution differences through feature alignment.

[0048] Constructing a three-layer joint network topology: The model includes a feature extractor. G Domain discriminant C and regressor R .

[0049] Feature extractor G The mapping logic is as follows: a feature extraction network is constructed using multi-layer linear neurons and non-linear activation functions to map the input slope feature vector into a high-order latent feature representation z=G(x).

[0050] Adversarial mechanism of domain discriminator C: Construct a discriminative network with latent feature representation as input, output the discriminative probability p=D(z) of the domain to which the sample belongs, and guide feature alignment through binary cross-entropy loss.

[0051] Regressor R Prediction mapping: Construct a regression prediction network to map the latent feature representation to the slope safety factor prediction result y=R(z), thereby achieving accurate capture of nonlinear stability characteristics.

[0052] Multi-loss joint training optimization: Define a joint loss function, including prediction loss. Domain-based loss discrimination and distribution difference loss By updating the feature extractor using reverse gradients, G We learn domain-invariant feature representations, which ensure the accuracy of regression prediction even when the domain discriminator C cannot distinguish the source of the features.

[0053] S4. Optimize the parameters of the adversarial multi-source transfer learning model. S41. Set a set of hyperparameters for the model to be optimized. The set of hyperparameters includes at least one or more of the following: learning rate, number of network layers, number of neurons per layer, dropout rate, batch size, number of training epochs, and loss weight coefficient. S42. Using the prediction error index on the validation set as the fitness function, an intelligent optimization algorithm is used to search for the optimal combination of hyperparameters that minimizes the fitness function of the validation set. S43. Train the adversarial multi-source transfer learning model using the optimal hyperparameter combination and output the optimal model.

[0054] The intelligent optimization algorithm is the walrus optimizer algorithm. The walrus optimizer switches between global exploration and local development through an adaptive adjustment mechanism of danger signals and safety signals to achieve hyperparameter optimization. While training the adversarial multi-source transfer learning model, the validation set loss is monitored in real time. When the validation loss no longer decreases within a preset number of rounds, the training is automatically terminated.

[0055] Specifically, to ensure the robustness of the model in cross-domain prediction scenarios, an adaptive algorithm is used to globally optimize the network hyperparameters.

[0056] Hyperparameter set definition: Determine the parameter space to be optimized, including learning rate, number of network layers, number of hidden layer neurons, dropout rate, batch size, number of training epochs, and loss weight coefficients, etc.

[0057] Adaptive optimization strategy: The Walrus Optimizer algorithm is employed. Utilizing the optimizer's hazard and safety signal adjustment mechanism, a balance is struck between global exploration and local exploitation to search for the optimal parameter combination that minimizes the validation set fitness function (e.g., mean squared error).

[0058] Refined Model Training and Early Stopping Protection: The model is trained using the optimal parameter combination, and the validation set loss is monitored in real time. Early stopping logic is introduced, which automatically terminates training when the validation loss no longer decreases within a preset number of rounds, and saves the optimal model parameters (State Dict) to prevent overfitting.

[0059] S5. Input the sample of the seabed hydrate slope to be evaluated into the parameter-optimized model, and output the slope stability prediction results and the contribution of key factors. S51. Obtain the input features of the seabed hydrate slope to be evaluated (e.g., , c , , E , n , k , H , SA , D , S h , S hd , u ex wait); S52. Input the input features into the optimal model after parameter optimization, and output the predicted safety factor value. F s预 ; S53, Predicted value based on safety factor F s预 The stability assessment result is obtained by comparing the result with a preset threshold, and the contribution or sensitivity ranking result of key influencing factors is output. The preset threshold is a safety factor threshold pre-set based on engineering specifications or historical experience.

[0060] In this embodiment, the contribution of key impact factors can be obtained in any of the following ways: a) Calculate feature contributions based on model interpretability methods, such as permutation importance and SHAP value; b) When the feature extractor includes an attention mechanism, use the attention weights as a reference for feature contributions; c) Perform perturbation sensitivity analysis on the input features and output the sensitivity to perturbation. F s Ranking of the key factors most sensitive to change.

[0061] Specifically, the acquired target domain features to be evaluated are input into the optimal model to perform stability prediction and contribution quantification analysis.

[0062] Automated prediction of safety factor: Input the input features of the submarine hydrate slope to be evaluated (such as soil unit weight, cohesion, hydrate saturation, excess pore water pressure, etc.) into the optimized model, and the feature extractor and regressor work together to output the predicted value of the safety factor.

[0063] Stability classification determination: The predicted safety factor is compared with the preset physical stability threshold, and the stability determination result of the slope (such as stable, unstable or understability) is output.

[0064] Key Influencing Factor Contribution Analysis: Through disturbance sensitivity analysis or interpretability algorithms, the contribution of each physical parameter to the fluctuation of the safety factor is quantified.

[0065] Sensitivity ranking output: The model finally outputs a ranking of the key factors most sensitive to changes in the safety factor, providing auxiliary decision support for disaster prevention and control and hydrate mining risk assessment of seabed slopes.

[0066] In this embodiment, the adversarial multi-source transfer learning model optimized by the walrus optimizer is input, and the predicted safety coefficient value output by the model is: F s预 =1.15, prediction error ε=2.68%, indicating that the method can still maintain high prediction accuracy even when the target domain has few samples and the distribution of the source domain and the target domain are significantly different.

[0067] like Figure 2 As shown, a system for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning includes a dataset construction module, a data preprocessing module, an adversarial multi-source transfer learning module, a parameter optimization module, and a stability prediction output module. The dataset construction module acquires and constructs a multi-source domain dataset containing source domain data and target domain data. The data preprocessing module performs unit and dimension unification, anomaly and missing data processing, standardization, feature filtering or dimensionality reduction, and target domain sample partitioning on the multi-source domain data. The adversarial multi-source transfer learning module constructs and trains an adversarial transfer learning model composed of a feature extractor, a domain discriminator, and a regressor or classifier to achieve feature alignment between the source and target domains. The parameter optimization module adaptively optimizes the model's hyperparameters. The stability prediction output module inputs the submarine hydrate slope samples to be evaluated into the parameter-optimized model and outputs the slope stability prediction results and the contributions of key influencing factors.

[0068] In summary, this invention constructs a multi-source domain slope dataset, then unifies the units and dimensions of the multi-source domain data, detects and processes outliers, and performs standardization and feature filtering or dimensionality reduction. It establishes an adversarial transfer learning network composed of a feature extractor, a domain discriminator, and a regressor. Through adversarial training, it achieves feature alignment between the source and target domains and outputs the predicted safety factor of the seabed hydrate slope in the target domain. A walrus optimizer is used to adaptively optimize the network hyperparameters. Finally, it outputs the slope stability prediction results and the contributions of key influencing factors. This invention can significantly improve the prediction accuracy and generalization ability of the safety factor of seabed hydrate slopes even under conditions of limited sample size, significant differences in the distribution of the source and target domains, and strong interference from outlier data. It has strong engineering applicability and can be widely applied to hydrate mining risk assessment, seabed slope stability analysis, and geological disaster monitoring and early warning.

[0069] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention shall fall within the scope of the technical solution of the present invention.

Claims

1. A method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning, characterized in that, Includes the following steps: S1. Obtain and construct a multi-source domain slope dataset; the multi-source domain slope dataset includes source domain data and target domain data; S2. Preprocessing of multi-source domain dataset samples: S3. Construct an adversarial multi-source transfer learning model; S4. Optimize the parameters of the adversarial multi-source transfer learning model; S5. Input the sample of the seabed hydrate slope to be evaluated into the model after parameter optimization, and output the slope stability prediction results and the contribution of key influencing factors.

2. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Source domain data construction: The source domain dataset is a land slope or conventional soil slope dataset, which includes at least slope geometric parameters, soil mechanical parameters and corresponding stability level data to form the source domain dataset. S12. Target Domain Data Construction: The target domain dataset consists of data obtained from triaxial or direct shear tests using similar materials in indoor experiments, showing the variation of strength parameters with hydrate state; a pore pressure evolution model is established to calculate excess pore water pressure. u ex The input characteristic parameters are sampled within the set range of slope geometric parameters and working parameters; finally, the safety factor is calculated for each set of sampled parameters. F s This forms the target domain dataset.

3. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 2, characterized in that, The input features of a multi-source domain dataset must include at least one or more of the following: soil unit weight Cohesion c internal friction angle Elastic modulus E Poisson's ratio ν Permeability coefficient k , slope height H slope angle SA , water depth D hydrate saturation S h Degree of hydrate decomposition S hd excess pore water pressure u ex .

4. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21. Unify the dimensions and units of the source domain data and the target domain data, and complete or remove missing data. S22. Identify and process abnormal samples; S23. Standardize the input features; S24. Filter or reduce the dimensionality of the input features to obtain the feature vectors used for model training; S25. Divide the target domain samples into a training set and a validation set, and expand the target domain training set when the number of target domain samples is insufficient.

5. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 1, characterized in that, Step S3 specifically includes the following steps: S31. Constructing the model structure: The adversarial multi-source transfer learning model includes at least a feature extractor, a domain discriminator, and a regressor; the feature extractor is used to map the input feature vector to a latent feature representation; the domain discriminator is used to distinguish whether the latent feature representation comes from the source domain or the target domain; the regressor is used to map the latent feature representation to the slope stability prediction result; S32. Adversarial learning mechanism: Through adversarial training, the feature extractor learns domain-invariant feature representations, reducing the feature distribution differences between the source and target domains.

6. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 5, characterized in that, Adversarial training of the model employs joint training optimization using multiple loss functions, the joint loss function of which is expressed as: in, This is the distribution difference loss, used to minimize the difference in the distribution of potential features between the source and target domains; This is the domain discrimination loss, used to train the domain discriminator to distinguish between the source domain and the target domain; To predict the loss, the stability label error is minimized; α , β , λ These are the loss weighting coefficients, used to balance domain alignment and prediction accuracy.

7. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 1, characterized in that, Step S4 specifically includes the following steps: S41. Set a set of hyperparameters for the model to be optimized. The set of hyperparameters includes at least one or more of the following: learning rate, number of network layers, number of neurons per layer, dropout rate, batch size, number of training epochs, and loss weight coefficient. S42. Using the prediction error index on the validation set as the fitness function, the optimal combination of hyperparameters that minimizes the fitness function of the validation set is searched through an intelligent optimization algorithm. S43. Train the adversarial multi-source transfer learning model using the optimal hyperparameter combination and output the optimal model.

8. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 7, characterized in that, The intelligent optimization algorithm is the Walrus Optimizer algorithm. The Walrus Optimizer switches between global exploration and local development through an adaptive adjustment mechanism of danger signals and safety signals to achieve hyperparameter optimization. While training the adversarial multi-source transfer learning model, the validation set loss is monitored in real time. When the validation loss no longer decreases within a preset number of rounds, the training is automatically terminated.

9. The method for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning according to claim 1, characterized in that, Step S5 specifically includes: S51. Obtain the input features of the seabed hydrate slope to be evaluated; S52. Input the above input features into the optimal model after parameter optimization, and output the predicted safety factor value. F s预 ; S53, Predicted value based on safety factor F s预 The stability determination result is obtained by comparing with the preset threshold, and the contribution or sensitivity ranking result of the key influencing factors is output.

10. A system for predicting the stability of submarine hydrate slopes based on adversarial multi-source transfer learning, used to implement the prediction method described in any one of claims 1-9, characterized in that, The system includes a dataset construction module, a data preprocessing module, an adversarial multi-source transfer learning module, a parameter optimization module, and a stability prediction output module. The dataset construction module is used to acquire and construct a multi-source domain dataset containing source domain data and target domain data. The data preprocessing module is used to perform unit and dimension unification, anomaly and missing data processing, standardization, feature selection or dimensionality reduction, and target domain sample partitioning on the multi-source domain data. The adversarial multi-source transfer learning module is used to construct and train an adversarial transfer learning model composed of a feature extractor, a domain discriminator, and a regressor or classifier to achieve feature alignment between the source domain and the target domain. The parameter optimization module is used to adaptively optimize the model hyperparameters; the stability prediction output module is used to input the submarine hydrate slope sample to be evaluated into the parameter-optimized model and output the slope stability prediction results and the contribution of key influencing factors.

Images