Search

CN-121980975-A - Submarine slope instability prediction method under submarine groundwater excretion effect based on Bayesian-CNN

CN121980975ACN 121980975 ACN121980975 ACN 121980975ACN-121980975-A

Abstract

The invention provides a submarine slope instability prediction method under the submarine groundwater excretion effect based on Bayesian-CNN, and belongs to the technical field of ocean engineering geology. Aiming at the problem that the existing model is difficult to quantify the nonlinear dynamic evolution of slope instability under the multi-factor coupling action of Submarine Groundwater Drainage (SGD) and waves and the like, the invention constructs a multi-parameter data set comprising water depth, wave height, wave period, salinity, turbidity, SGD rate and erosion depth through in-situ observation, and the safety coefficient Fs is calibrated by combining a fluid-solid coupling mechanism as a supervision tag. And reconstructing the characteristic vector into a 3X 3 matrix, inputting the matrix into a convolutional neural network, introducing Bayesian optimization, automatically adjusting super parameters, extracting the characteristic by adopting a 3X 1 convolutional kernel, realizing probability prediction by using a Dropout layer and Monte Carlo sampling, and quantifying uncertainty. The loss function introduces a critical period weighting mechanism to enhance the destabilizing critical state identification. The invention realizes high-precision prediction and provides reliable support for ocean engineering geological environment evaluation.

Inventors

  • JIA YONGGANG
  • WANG SIMING
  • LIU HANLU
  • JIA ZHENTIAN
  • ZHANG YUMENG
  • JIANG LONG
  • ZHANG ZHENGRONG
  • SHAN HONGXIAN
  • Quan Yongzheng

Assignees

  • 中国海洋大学

Dates

Publication Date
20260505
Application Date
20260408

Claims (8)

  1. 1. A method for predicting submarine slope instability under the submarine groundwater excretion effect based on Bayesian-CNN is characterized by comprising the following steps: S1, multi-parameter coupling observation and data set construction: collecting various submarine environment parameter data of water depth, wave height, wave period, salinity, turbidity, SGD speed and erosion depth through in-situ observation, and constructing a data set; s2, calibrating a submarine slope safety index Fs based on a fluid-solid coupling mechanism: Utilizing the original data acquired in the step S1, calculating and calibrating a slope safety coefficient Fs by utilizing a limit balance criterion based on a fluid-solid coupling theory, upward seepage force generated by SGD, hyperstatic pore water pressure accumulation effect and seabed dynamic erosion feedback, and constructing a tag set; s3, importing and preprocessing data, reconstructing and dividing a sample space: simultaneously reconstructing one-dimensional characteristic vectors into a two-dimensional matrix form to adapt to the structural characteristics of a convolutional neural network, and finally dividing 1000 groups of sample sets into training sets and verification sets according to a ratio of 7:3; s4, constructing a Bayesian-CNN neural network model: the network structure comprises a multi-scale feature extraction layer, a regularization module, a Bayesian random inactivation layer and a regression output layer, and introduces a weighted loss function to enhance the sensitivity to the destabilization critical state; s5, training a Bayesian-CNN neural network: s5-1, training strategy configuration and model learning: performing integrated training on the neural network model constructed in the step S4 by using 700 groups of training set samples divided in the step S3; S5-2, probability prediction implementation based on Monte Carlo sampling: A Monte Carlo sampling strategy is adopted when the model predicts 300 groups of data of the validation set; S6, prediction and result evaluation: Predicting safety coefficient of test sample by using model trained in step S5, and evaluating overall error index including Root Mean Square Error (RMSE), average absolute error (MAE) and decision coefficient To evaluate the predictive performance of the model.
  2. 2. The method for predicting the submarine slope destabilization under the influence of the submarine groundwater excretion based on Bayesian-CNN according to claim 1, wherein the step S1 specifically comprises the following steps: s1-1, submarine groundwater excretion influencing factor coupling effect: the water depth D directly affects the drainage strength of SGD by changing the hydraulic gradient between groundwater and seawater, and the average seepage velocity of groundwater in sediment is expressed as: ; wherein: the equivalent seepage velocity m/s for the drainage of the submarine groundwater, K is the permeability coefficient m/s of the submarine sediment; the land source fresh water head height m; D is the water depth m; Is the density of fresh water; L is the equivalent flow path length m of groundwater from land source to drainage point; In the process of wave near shore propagation, the wave alternately changes along with the wave peaks and wave troughs, and periodic pressure disturbance is formed on the surface of the seabed, so that a pumping effect is generated on pore water, and the instantaneous additional pressure induced by the wave is expressed as: ; wherein: The pressure Pa is induced for waves; the gravity acceleration is m/s2; taking the effective wave height as the wave height m Or maximum wave height Wave number is the number of waves, satisfy ; Is the wavelength m; In order to be of an angular frequency, ; Is the wave period s; is the vertical height m from the seabed surface; Is a horizontal position coordinate; The wave period further affects the pressure decay depth by controlling the wavelength, the relationship of which is determined by the dispersion relationship: ; wherein: The wavelength is m, T is wave period s, D is water depth m; the subsea ramp destabilization safety stability factor under SGD is a function driven by multiple parameters: ; Predicting the submarine slope safety stability coefficient from the complex background of the multi-parameter coupling effect by constructing a neural network; S1-2, in-situ observation and data set construction: obtaining hydrodynamic parameters, pore water pressure, submarine groundwater excretion parameters and seabed erosion depth of a research area, and obtaining data sets of training and testing models, wherein the data sets comprise water depth D, effective wave height and effective wave period Maximum wave height Maximum wave period S, turbidity SGD rate Depth of attack ; Groundwater drainage rate calculation based on pore water pressure signal decomposition: For the collected pore water pressure sequence Performing wavelet transformation: ; wherein: Is a wavelet transform coefficient; Is of a scale Translates into Decomposing the pore water pressure signal into high-frequency and short-period wave action components, low-frequency and long-period tide action components, and residual groundwater runoff background components by reconstructing wavelet scales; based on the decomposed pore water pressure components, the groundwater drainage quantity corresponding to different power mechanisms is calculated respectively and is converted into daily drainage rate in unit shoreline length, total submarine groundwater drainage rate The expression is as follows: ; wherein: The total drainage amount of the groundwater in the sea bottom in the unit shoreline length is m < 3 >. M < -1 >. D < -1 >; net drainage component formed for the direct access of groundwater runoff to the sea; groundwater infiltration or drainage components within a unit shoreline length induced by wave action; is the sea water-groundwater exchange component in unit shoreline length caused by tide fluctuation.
  3. 3. The method for predicting the submarine slope destabilization under the influence of the submarine groundwater excretion based on Bayesian-CNN according to claim 1, wherein the step S2 specifically comprises the following steps: S2-1, effective slope angle correction under dynamic erosion feedback: sediment erosion phenomena accompanying the sea bed in the SGD active period are corrected in real time by using the erosion depth Ds: ; wherein: an effective ramp angle rad for the modified ramp; Is an initial slope angle; characterizing the gradient change rate caused by the unit erosion depth for the erosion-slope correction coefficient, and taking 0.0005rad/mm; the sea bed erosion depth mm is obtained for in-situ observation; s2-2, an ultra-static pore pressure dynamic evolution model of SGD based on seepage mechanical effect: The heterogeneity of the seabed sediment porosity and permeability coefficient, the hyperstatic pore water pressure u generated by SGD is not considered as a simple linear mapping of the excretion rate v, based on the modified expression of darcy's law and the effective stress principle: ; wherein: is the additional pore water pressure kPa induced by SGD dynamic percolation; The drainage rate m3/d of groundwater in the sea floor; initial permeability coefficient for the deposit; is the current effective stress kPa; is the reference effective stress kPa; Is the stress sensitivity coefficient; dynamic viscosity Pa of water s; Is the severe kN/m3 of water; a transient Kong Yabo motion component induced by waves; s2-3, evaluating accumulated stability considering pore pressure dissipation hysteresis: instability of the submarine ramp is caused by accumulated pore pressure accumulation effect caused by long-time high-strength drainage, and accumulated pore pressure factor is defined : ; Wherein: is the cumulative intensity coefficient; for pore pressure dissipation rate, determined by sediment consolidation coefficient; is the instantaneous pore water pressure kPa; f s corresponding to each group of observation samples is calculated based on Mohr-Coulomb limit balance criterion: ; wherein: Effective cohesion kPa for deposit; Is the total normal stress kPa; is static pore water pressure; accumulating the superpore pressure kPa for SGD; Is the wave dynamic pressure component kPa; Effective internal friction angle rad for deposit ; Is slip force kPa; in which the sliding force of the denominator The method comprises the following steps: ; wherein: saturating a heavy object kN/m3 for sediment, wherein D is the water depth m; Total normal stress The method comprises the following steps: ; SGD accumulated superhole pressure The method comprises the following steps: ; Wave dynamic pressure component The method comprises the following steps: 。
  4. 4. the method for predicting the submarine slope destabilization under the influence of the submarine groundwater excretion based on Bayesian-CNN according to claim 1, wherein the step S3 specifically comprises the following steps: s3-1, obtaining the data source determined according to the steps S1 and S2, wherein the data source comprises independent variables: and a target variable A time series data matrix M of (a); S3-2, data cleaning and Z-Score standardization: Performing preprocessing operation on the acquired multidimensional time series data matrix M to ensure robustness of model training, namely performing data cleaning, systematically removing sample records containing missing values or infinite values from the matrix, and performing smoothing treatment; To eliminate the weight imbalance due to physical dimensional differences between different features including water depth m, wave height m and drainage rate m3/d, a Z-Score normalization transform is performed on the input feature vector X: , wherein: Is the normalized characteristic value; Is the characteristic mean value; Standard deviation of the feature; s3-3, dividing a sample set: Dividing the calibrated 1000 groups of samples into two independent subsets according to a fixed ratio of 7:3, namely, using the first 700 groups as training sets for updating the weight of the neural network, and using the last 300 groups as verification sets for model generalization performance evaluation; s3-4, tensor reconstruction: Will be Is quantized and reconstructed into the feature vector according to physical meaning As input to a convolutional neural network: ; S3-5, the training sample set and the test sample set which are finally obtained are formed by the sequences, the data preprocessing is partially completed, and the data are directly input into the Bayesian-CNN neural network for model training and prediction.
  5. 5. The method for predicting subsea slope destabilization under the action of subsea groundwater drainage based on Bayesian-CNN according to claim 4, wherein step S4 comprises the steps of: S4-1, model structural design: the construction of the neural network adopts a convolution CNN structure for predicting the safety stability coefficient of the submarine slope instability: The network input is a time sequence data matrix M obtained by processing in the step S3-1, and the input layer receives Is a two-dimensional matrix of S3-4 Is obtained by reconstructing feature vector Zhang Lianghua of the model, and adopts a multi-scale feature extraction layer and a layer 1 convolution The convolution kernel extracts macroscopic dynamic characteristics and cooperates with packing filling, the output characteristic channel is selected in [16,32,64], meanwhile Batch Normalization batches of standardized layers are introduced to normalize the characteristic mapping, and the layer 2 convolution adopts Extracting fine fluctuation features among parameters by a convolution kernel, and selecting an output feature channel from [32,64,128 ]; regularization and activation module, namely, after convolution of each layer, accessing a ReLU activation function, introducing an L2 regularization penalty term, and leading the coefficients into the matrix Selecting; The Bayes random inactivation layer Dropout layer is connected with the Bayes random inactivation Dropout layer after the convolution layer, and the inactivation rate is selected from [0.2,0.5 ]; the Regression mapping layer outputs a sea-bottom slope safety coefficient Fs prediction result through the full-connection layer of the neurons selected in the step 16,32,64 according to the flattening characteristic of the flat layer and finally through the Regression layer of the single neurons; A key period weighting mechanism is introduced into a loss function, wherein the prediction error of a submarine slope instability abrupt change or a wave height abrupt increase point is given higher weight, the response capability of a model to an extreme event is improved, and the loss function is defined as follows: ; Where N is the number of samples, Is the true value of the subsea ramp safety and stability coefficient, Is the predicted value of the safety and stability coefficient of the submarine slope, Weighting coefficients for critical erosion periods; S4-2, setting a network super-parameter and a Bayesian optimization range, adopting Bayesian optimization to automatically adjust the network super-parameter, and setting the initial learning rate as the value on a logarithmic scale The number of convolution kernels of the first layer is selected from [16,32,64], the number of convolution kernels of the second layer is selected from [32,64 and 128], the Dropout inactivation rate is uniformly distributed in [0.2,0.5], and the L2 regularization coefficient is set to be in a logarithmic scale The number of neurons of the full connection layer is taken [16,32,64], and the batch processing size is taken [16,32,64]; S4-3, bayesian optimization configuration and a composite objective function, performing 30-50 iterations of Bayesian optimization, performing initial random sampling on 5-10 groups of super parameters to construct Gaussian process prior, guiding search by adopting an expected lifting acquisition function, and adopting a composite error of 5-fold cross validation for the objective function: wherein For critical section samples Is weighted to be 2.0, Taking 1.
  6. 6. The method for predicting the submarine slope destabilization under the influence of the submarine groundwater excretion based on Bayesian-CNN according to claim 1, wherein the step S5-1 specifically comprises the following steps: S5-1-1, selecting an optimizer and dispatching a learning rate, namely selecting an Adam optimizer, automatically dispatching key super parameters through Bayesian optimization, wherein the initial learning rate searches a range [ 1X 10- 4 , 1X 10-2] on a logarithmic scale, the first moment attenuation coefficient beta 1 searches in [0.85,0.95], and beta 2 is fixed to be 0.999; S5-1-2, calculating a loss function, wherein the total loss is the sum of weighted mean square error and L2 regularization term, namely Weight of Setting critical area according to sample safety coefficient Get 2.0, neighboring area Taking 1.5 and the rest 1.0, L2 regularization coefficients On the logarithmic scale by Bayesian optimization And (5) internal searching.
  7. 7. The method for predicting the submarine slope destabilization under the influence of the submarine groundwater excretion based on Bayesian-CNN according to claim 1, wherein the step S5-2 specifically comprises the following steps: s5-2-1, repeatedly sampling, namely keeping a Dropout layer of a Bayesian random inactivation layer in an activated state during reasoning, and carrying out K=50 times of random deduction aiming at an input characteristic tensor at the same moment to generate a predicted value set; S5-2-2, extracting expected value by a formula Calculating to obtain a prediction mean value of the safety coefficient; s5-2-3, confidence interval quantization, namely calculating the standard deviation of 50 sampling results 。
  8. 8. The method for predicting the submarine slope destabilization under the influence of the submarine groundwater excretion based on Bayesian-CNN according to claim 1, wherein the step S6 specifically comprises the following steps: S6-1, prediction: After the neural network training is completed, the training set and the test set are predicted to obtain the submarine slope instability safety stability coefficient predicted value corresponding to each sample ; S6-2, result evaluation and error analysis: by integral error index including root mean square error RMSE, mean absolute error MAE and decision coefficient The method is used for comprehensively evaluating the overall prediction performance of the model.

Description

Submarine slope instability prediction method under submarine groundwater excretion effect based on Bayesian-CNN Technical Field The invention relates to the field of ocean engineering geology technology and ocean observation technology, in particular to a submarine slope instability prediction method under the submarine groundwater excretion effect based on Bayesian-CNN. Background Subsea groundwater drainage (Submarine Groundwater Discharge, SGD) is a key process for sea-land groundwater interaction, and its induced percolation effect can alter the stress state of the seabed and thus affect subsea slope stability. Current mainstream seafloor slope destabilization research focuses mainly on near undercut stress models generated by wave-seabed-slope composite hydrodynamic action, and Newmark slider models generated by seismic action. However, none of the existing models incorporate the dynamic impact of SGD on sediment transport, especially lacking the ability to accurately predict the slope instability of the seabed that occurs under extreme hydrodynamic conditions. Recent researches show that the seepage intensity generated by SGD can reach 3-5 times of wave-induced seepage, and the vertical seepage jacking force generated by SGD can directly weaken the structural stability of surface layer sediment. The Chinese yellow river underwater delta is taken as a typical SGD active region, the actual measurement excretory flux reaches 0.5-3.2 m < 3 >/(m.d), the seepage rate is 2.5-15.6 cm/d, and the obvious space-time heterogeneity is presented. In-situ observations carried out by the team in the eastern lone sea area in 2024 months prove that the SGD flux has significant positive correlation with submarine slope destabilization. But is limited by the technical bottleneck of submarine environment monitoring, the existing observation means are difficult to realize long-time sequence, large-scale and real-time monitoring of the destabilization process, and the key period of slope destabilization is difficult to accurately monitor. Traditional physical models are computationally inefficient and difficult to describe the random uncertainty of the subsea environment by a single deterministic number when dealing with multiple field coupling of SGD seepage, wave loading and dynamic evolution of the seabed topography. Therefore, there is a need for a deep learning predictive model that can fuse multisource dynamic observations and give an assessment of risk of instability. Aiming at the technical bottleneck, the research proposes a new model for predicting the submarine slope instability based on deep learning. By fusing the super-parameter optimizing capability of the Bayesian optimizer and the accuracy advantage of probability prediction of the CNN neural network, a submarine slope instability dynamic prediction model under the condition of SGD dominance can be effectively established. The method can avoid the parameter sensitivity problem of the traditional model, break through the space-time limitation of in-situ observation and provide a new method support for submarine engineering geological environment evaluation. Disclosure of Invention In order to make up for the defects of the prior art, the invention provides a submarine slope instability prediction method under the submarine groundwater excretion effect based on Bayesian-CNN. The invention is realized by the following technical scheme that the method for predicting the submarine slope instability under the submarine groundwater excretion effect based on Bayesian-CNN specifically comprises the following steps: S1, multi-parameter coupling observation and data set construction: And acquiring various submarine environment parameter data including water depth, wave height, wave period, salinity, turbidity, SGD speed and erosion depth through in-situ observation, and constructing a data set. S2, calibrating a submarine slope safety index Fs based on a fluid-solid coupling mechanism: and (3) calculating and calibrating a slope safety coefficient Fs by utilizing a limit balance criterion based on the fluid-solid coupling theory, upward seepage force generated by SGD, hyperstatic pore water pressure accumulation effect and seabed dynamic erosion feedback by utilizing the original data acquired in the step (S1), and constructing a tag set. S3, importing and preprocessing data, reconstructing and dividing a sample space: and importing the data obtained in the step S1 and the step S2, and performing cleaning, smoothing and standardization treatment. And reconstructing the one-dimensional characteristic vector into a two-dimensional matrix form to adapt to the structural characteristics of the convolutional neural network, and finally dividing 1000 groups of sample sets into a training set and a verification set according to a ratio of 7:3. S4, constructing a Bayesian-CNN neural network model: The network structure comprises a multi-scale feature extraction layer, a regu