Bayesian-lstm-based long-term prediction method for deep-sea creep under in-situ pore pressure observation

Abstract

This invention discloses a long-term prediction method for deep-sea creep based on in-situ pore pressure observation using Bayesian-LSTM, relating to the fields of sediment dynamics and marine engineering. The method includes: deploying pore pressure sensors to collect pore water pressure data in real time, performing standardization processing, and calculating the pore pressure change rate; constructing a creep rate calculation model based on the effective stress principle and power law; building a time-series feature dataset using sliding window technology, constructing and training a Bayesian-LSTM network model, and introducing a Monte Carlo Dropout layer to assess prediction uncertainty; outputting a high-precision creep rate using a multiple sampling averaging strategy; calculating the cumulative creep distance through numerical integration, and constructing confidence intervals to achieve medium- and long-term trend prediction and risk analysis. This invention integrates physical mechanisms and data-driven approaches, effectively handling data noise and prediction uncertainty in the complex environment of the deep sea, significantly improving the accuracy and reliability of long-term creep prediction, and providing a scientific basis for deep-sea engineering safety assessment.

Description

Technical Field

[0001] This invention relates to the fields of sediment dynamics and marine engineering technology, and more specifically, to a long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM. Background Technology

[0002] Creep zones and seafloor deformation are widely distributed in all major sea areas worldwide. With the continuous development of deep-sea exploration technology and seafloor construction, researchers have gradually realized the impact of sediment creep on seafloor structures and deep-sea ecosystems. Sediment creep refers to the gradual flow or deformation of seafloor sediments under long-term pressure. The rate of seafloor deformation is the most direct manifestation of seafloor slope instability, and it is generally monitored using in-situ long-term observation methods. However, due to the complexity of the deep-sea environment, in-situ observation is difficult to implement, the requirements for various indicators of the observation equipment are high, and seafloor deformation and sliding events are highly random, requiring long-term observation to capture minute deformations of the seabed. Currently, in-situ observation of seafloor deformation and sliding is still in the exploratory stage both domestically and internationally, and most seafloor deformation and sliding equipment is only used for nearshore shallow-sea landslide deformation observation. Considering the conditions of the observation equipment, long-term in-situ observation is too difficult. Traditional sediment creep models mostly rely on single experimental data obtained from laboratory experiments after sediment sampling or further combined with simple mechanical models, which cannot take into account the complex factors in the seafloor environment, or even the deep-sea environmental background, such as pore pressure changes, fluid pressure, and storm events. The accuracy and predictive power of these models have significant limitations in practical applications in marine environments.

[0003] In recent years, the development of machine learning and deep learning technologies, especially the application of Long Short-Term Memory (LSTM) networks, has greatly enhanced the ability to model and predict time-series data. LSTM networks can predict future trends using historical data, making them particularly suitable for complex time-series data. However, a drawback of LSTM models is their inability to handle uncertainty in predictions, especially when data is insufficient or noisy. To address this issue, this invention proposes a Bayesian-LSTM model that combines Bayesian learning with LSTM. This model uses Bayesian inference to correct the predictions of LSTM, providing more accurate predictions and prediction ranges, thus overcoming the shortcomings of traditional models in handling uncertainty and long-term prediction. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM.

[0005] This invention is achieved through the following technical solution: a long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM, specifically including the following steps: Step S1, Data Acquisition and Processing: First, pore pressure sensors are deployed: pore pressure sensors are installed in the seabed sediments to monitor the pore water pressure in the sediments in real time, acquiring data continuously. The sensors record pore pressure data hourly. And continuously transmit; perform preliminary preprocessing on the sampled data and standardize the raw pore water pressure data; Step S2, LSTM and Bayesian Learning Corrections: Step S2-1: Constructing a time series feature dataset: Transform the one-dimensional time series data of different variables changing over time after preprocessing in Step S1 into a three-dimensional supervised learning dataset; divide the data into two parts, the first part being the training and prediction of the model, and the second part being the verification and data comparison of the model predictions in the first part. Step S2-2: Build and train the Bayesian-LSTM network model: Construct a Bayesian-LSTM network model to record the historical fluctuation patterns of pore pressure and reflect on itself during prediction to assess uncertainty. Step S2-3, Model Probability Prediction Validation: The built network model is transformed into an executable prediction tool, and a multi-sampling and averaging strategy is used to ensure high accuracy of the prediction results. Step S3, Cumulative Migration Calculation and Medium- to Long-Term Trend Analysis: Convert the instantaneous rate output by the model into cumulative distance and assess long-term risks.

[0006] As a preferred option, step S1 specifically includes the following steps: Step S1-1: Obtain the baseline pressure value through long-term data preprocessing. The baseline pressure represents the static pressure of sediment pore water without external disturbance; the superbaseline pore pressure is calculated as follows: ; Step S1-2: Calculate the pore pressure change rate using the time difference of pore pressure data. : , in, The time interval is in hours; Step S1-3: Obtain the creep rate The creep rate is proportional to the logarithm of the pressure: , in, This is a reference pressure, selected as the initial pressure value; The logarithmic function can be transformed into a power function using the following transformation:

[0007] Steps S1-4: Based on the change in pore pressure, introduce the pressure change rate. The effect of the rate of change of pore water pressure on the creep rate is considered to reflect the influence of the rate of pressure change on the creep rate; the specific calculation formula is as follows:

[0008] in, Creep rate represents the rate at which a sediment deforms at a given moment. The pore water pressure (kPa) for real-time monitoring is data monitored and recorded in real time by a pore water pressure sensor, reflecting the pressure changes of the fluid in the pores of the sediment. It is the creep index, obtained through experiments, which describes the degree of influence of changes in pore water pressure on sediment deformation, and is also the creep sensitivity of sediments; It is the creep rate, obtained through static creep experiments, and represents the basic deformation rate of a sediment under given pressure conditions; The pore pressure change rate reflects the rate of change of pore water pressure.

[0009] Furthermore, step S2-1 specifically includes the following steps: Specifically, the following steps are included: Step S2-1-1: Input Feature Definition: Select two core dimensions as the input basis for the model, namely, the standardized real-time pore water pressure data, including pore pressure. and pore pressure change rate ; Step S2-1-2, Sliding window settings: Set the time step to T; set T=24, with a 24-hour cycle; Step S2-1-3: Construct tensor data: Input set X, construct a three-dimensional matrix of shape (N, T, 2). Where N is the total number of samples, T is the historical length of 24 hours, and 2 represents the two features mentioned above. Input set Y: Construct a two-dimensional vector of shape (N, 1), with the target value being the true creep rate at time t+1.

[0010] Furthermore, step S2-2 specifically includes the following steps: Specifically, the following steps are included: Step S2-2-1: Build the network structure: Input a time-series data matrix with shape (T, 2) and T=24; Step S2-2-2, Long Short-Term Memory Layer (LSTM): Set up 24 LSTM neurons to construct the core temporal feature extraction region of the model; within a complete 24-hour cycle, automatically identify and record important and recurring effective signals of tidal fluctuations, while automatically ignoring those accidental and meaningless momentary interferences; Step S2-2-3, Monte Carlo Random Deactivation Layer (MC Dropout Layer): After the LSTM layer, a Dropout layer with a dropout rate of p is connected. p is set to 0.3, that is, during each forward computation, the system will randomly block 30% of the neuron nodes, causing them to pause work, and keep 70% of the neurons in an active working state. Step S2-2-4, Fully Connected Layer: The output layer contains one neuron, which aggregates all the features extracted earlier and outputs a final value as the predicted creep rate.

[0011] Furthermore, steps S2-3 specifically include the following steps: Specifically, the following steps are included: Step S2-3-1, Import validation data: Import the first part of the prediction data from the validation set data partitioned in S2-1 into the model; Step S2-3-2: Perform Monte Carlo multiple prediction: Use the ensemble averaging method; during the prediction process, while keeping the Dropout layer p=0.3 in the model on, for each time point in the validation set, control the model to repeat the prediction 100 times, and the 100 prediction results form a prediction cluster; Step S2-3-3, Result Correction and Accuracy Verification: Calculate the arithmetic mean of the above 100 predictions as the final prediction result at this moment; Step S2-3-4, High-precision verification: Perform regression analysis between the above average predicted values ​​and the actual observed values.

[0012] Furthermore, step S3 specifically includes the following steps: Step S3-1, Cumulative Migration Integration Calculation: Based on the high-precision predicted rate output in Step S2-3, the cumulative deformation of seabed sediments is calculated using the numerical integration method; the calculation formula is:

[0013] in, for The cumulative creep distance at time t, For the first Predicted creep rate at time t, The sampling time interval is 1 hour. Step S3-2: Construct confidence intervals: Using the distribution data generated from the 100 predictions in step S2-3, calculate the standard deviation at each time point. Construct confidence intervals; Step S3-3, Medium- and Long-Term Trend Prediction: Use the model to simulate the cumulative deformation of submarine landslides over the next 1-10 years, and make the uncertainty of long-term predictions increase cumulatively.

[0014] This invention, by employing the above technical solutions, offers the following advantages compared to existing technologies: It combines the effective stress principle and the power law creep model, utilizing a Bayesian Long Short-Term Memory (LSTM) network model for data-driven creep prediction, overcoming the uncertainties and complexities inherent in traditional methods for handling sediment creep. Furthermore, this invention, through precise pore pressure monitoring data, can calculate creep rates in real time and analyze long-term creep behavior based on the prediction results, providing a more reliable prediction tool for seafloor sediment research, deep-sea engineering, and soil engineering. Additional aspects and advantages of the invention will become apparent in the following description or may be learned by practice of the invention. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is an overall flowchart of a method for predicting creep distance in deep-sea creep zones based on a Bayesian-LSTM model proposed in this invention. Figure 2 This is a schematic diagram of the seabed in-situ observation equipment used in this embodiment; where (a) is a structural diagram of the seabed in-situ observation platform and battery compartment; and (b) is a detailed diagram of the multi-section probe structure equipped with a pore pressure sensor. Figure 3 Time series diagram of annual monitoring data of seabed pore water pressure in the Shenhu Gorge creep zone in the northern South China Sea (October 2024 to October 2025); Figure 4 Detailed diagrams of the multi-dynamic coupling mechanism and periodic cumulative deformation characteristics during a local time period (November 8-12, 2024); (a) shows the correspondence between total pore pressure, internal wave component and creep rate, and (b) shows the corresponding local cumulative displacement. Figure 5 This is a statistical graph of sediment creep behavior throughout the year during the in-situ observation period; where (a) is the daily average creep rate variation curve and (b) is the cumulative creep distance curve for the whole year. Figure 6This is a validation graph showing the prediction performance of the Bayesian-LSTM model during the validation period (August 20, 2025 - September 20, 2025); it demonstrates the high degree of fit between the model's predicted mean and the actual observed values ​​(R²). 2 =0.948); Figure 7 This is a graph showing the prediction results of the medium- to long-term cumulative deformation trend of submarine landslides over the next 1 to 10 years based on the model of this invention (including 95% confidence interval). Detailed Implementation

[0016] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0017] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0018] The following is combined with Figures 1 to 7 The present invention provides a detailed description of the long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM, according to an embodiment of the present invention.

[0019] like Figure 1 As shown, this invention proposes a long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM, characterized by the following steps: Step S1, Data Acquisition and Processing: First, pore pressure sensors are deployed: pore pressure sensors are installed in the seabed sediments to monitor the pore water pressure in the sediments in real time, acquiring data continuously. The sensors record pore pressure data hourly. The data is continuously transmitted; preliminary preprocessing of the sampled data is performed, including standardizing the raw pore water pressure data to ensure that the data is within a reasonable and feasible range to avoid affecting subsequent calculations; specifically, the following steps are included: Step S1-1: The baseline pressure value obtained through long-term data preprocessing The baseline pressure is obtained from long-term monitoring data and typically represents the static pressure of sediment pore water without external disturbance; the superbaseline pore pressure is calculated as follows: (Meaning of P(t)) Step S1-2: Calculate the pore pressure change rate using the time difference of pore pressure data. : , in, The time interval is in hours; Step S1-3: Obtain the creep rate In soil mechanics and geotechnical engineering, the following power-law relationship is often used to describe the relationship between creep rate and pore water pressure:

[0020] The core idea behind these methods is to predict sediment creep rates based on changes in pore water pressure. However, in deep-sea sediment research, given the inherent diversity of marine sediments, creep behavior is not always linear and often exhibits non-linear changes. Typically, creep rates are considered to be proportional to the logarithm of pressure. , in, This is a reference pressure, selected as the initial pressure value; Considering the diversity of marine sediments and the fact that creep behavior exhibits various nonlinear variations, based on the traditional power-law relationship used to describe the relationship between creep rate and pore water pressure, the empirical creep formula for the real marine environment is modified to transform the logarithmic function into a power function through the following transformation:

[0021] Steps S1-4: Based on the change in pore pressure, introduce the pressure change rate. This improvement considers the effect of the rate of change in pore water pressure on the creep rate, reflecting the influence of the pressure change rate on the creep rate. This improvement allows the model to more accurately capture the real-time impact of pressure changes on sediment deformation under actual conditions, enabling a more precise calculation of creep rate. The specific calculation formula is as follows:

[0022] in, Creep rate represents the rate at which a sediment deforms at a given moment. The pore water pressure (kPa) for real-time monitoring is data monitored and recorded in real time by a pore water pressure sensor, reflecting the pressure changes of the fluid in the pores of the sediment. The creep index, obtained experimentally, describes the degree of influence of changes in pore water pressure on sediment deformation and is also known as the creep sensitivity of sediments. This value determines the intensity of the influence of changes in pore water pressure on the deformation rate of sediments. It is the creep rate, obtained through static creep experiments, and represents the basic deformation rate of a sediment under given pressure conditions; The pore pressure change rate reflects the rate of change of pore water pressure.

[0023] Step S2, LSTM and Bayesian Learning Corrections: The model input is constructed based on the time-series data obtained in step S1. The time-series data is divided into two parts: one part is used for model building and training of the Bayesian-LSTM network model, and the other part is used for model probability prediction and verification. The core of this step is to build an intelligent model that can both remember historical patterns and assess prediction risks. The specific implementation process is generally divided into two aspects: data model building and model testing.

[0024] Step S2-1: Constructing a time series feature dataset: In order to meet the format requirements of the LSTM network for time series input, it is necessary to use the sliding window technique to transform the one-dimensional time series data of different variables changing over time after the preprocessing in step S1 into a three-dimensional supervised learning dataset; the data level is divided into two parts, the first part is for model training and prediction, and the second part is for testing the model predictions in the first part and comparing the data. Specifically, the following steps are included: Step S2-1-1, Input Feature Definition: In this embodiment, two core dimensions are selected as the input basis for the model, namely, the standardized real-time pore water pressure data, including pore pressure. and pore pressure change rate ; Step S2-1-2, Sliding Window Settings: Set the time step to T; set T=24, with a 24-hour period. The time step of this layer, Timesteps=24, not only corresponds to 24 hours in physical time but also serves as a complete boundary for the diurnal dynamic evolution cycle. Within this 24-hour closed-loop time window, multi-scale temporal dependencies can be captured simultaneously. It can identify high-frequency nonlinear features such as instantaneous pore pressure abrupt changes at the millisecond level, and also fully extract the low-frequency periodic features of semi-diurnal tides, diurnal tides, and diurnal environmental cycles. This ensures that the model does not miss any long- or short-term dynamic behaviors occurring within a single day.

[0025] Step S2-1-3: Construct tensor data: Input set X, construct a three-dimensional matrix of shape (N, T, 2). Where N is the total number of samples, T is the historical length of 24 hours, and 2 represents the two features mentioned above. Input set Y: Construct a two-dimensional vector of shape (N, 1), with the target value being the true creep rate at time t+1.

[0026] Step S2-2: Build and train the Bayesian-LSTM network model: Construct a Bayesian-LSTM network model to record the historical fluctuations in pore pressure and perform self-reflection during prediction to assess uncertainty; specifically including the following steps: Step S2-2-1: Build the network structure: Input a time-series data matrix with shape (T, 2) and T=24; Step S2-2-2, Long Short-Term Memory Layer (LSTM): Set up 24 LSTM neurons to construct the core temporal feature extraction region of the model; within a complete 24-hour cycle, automatically identify and record important and recurring effective signals of tidal fluctuations, while automatically ignoring those accidental and meaningless momentary interferences; Step S2-2-3, Monte Carlo Randomized Dropout Layer: A Dropout layer with a dropout rate of p (0.3) is added after the LSTM layer. This means that during each forward computation, the system randomly masks 30% of the neurons, pausing their operation and keeping 70% active. This aims to achieve an optimal balance between feature preservation and random perturbation. On one hand, retaining 70% of the nodes ensures the model has sufficient information to capture the subtle and complex periodic changes in deep-sea pore pressure, avoiding underfitting due to excessive dropout. On the other hand, the 30% random masking introduces sufficient randomness to prevent overfitting to specific neurons, ensuring the statistical significance of the output prediction distribution. This causes the network's internal connection structure to change randomly with each prediction. Through repeated predictions, the model can output a series of slightly different predicted values. The variance of these values ​​directly reflects the model's uncertainty about the current prediction result, allowing for the calculation of confidence intervals using statistical methods.

[0027] Step S2-2-4, Fully Connected Layer: The output layer contains one neuron, which aggregates all the features extracted earlier and outputs a final value as the predicted creep rate.

[0028] Step S2-3, Model Probability Prediction Validation: The constructed network model is transformed into an executable prediction tool, and a specific multiple sampling and averaging strategy is used to ensure high accuracy of the prediction results; this includes the following steps: Step S2-3-1, Import validation data: Import the first part of the prediction data from the validation set data partitioned in S2-1 into the model; Step S2-3-2: Perform Monte Carlo multiple forecasts: To eliminate random biases that may arise from a single forecast and ensure forecast accuracy R 2>0.9, the ensemble averaging method is used; during the prediction process, while keeping the Dropout layer in the model with p=0.3 on, the model is controlled to repeat the prediction 100 times for each time point in the validation set; due to the existence of Dropout, the 100 prediction results will fluctuate slightly around the true value, forming a prediction cluster; Step S2-3-3, Result Correction and Accuracy Verification: To eliminate mean denoising, the arithmetic mean of the 100 predicted values ​​is calculated as the final prediction result at that moment. By averaging, the random positive and negative errors generated in a single prediction are canceled out, thus leaving the law closest to the physical essence.

[0029] Step S2-3-4, High-precision verification: Regression analysis was performed on the above average predicted values ​​and the actual observed values; the results showed that the predicted curve accurately depicted every peak and trough of the pore pressure change, proving that the method can effectively overcome deep-sea environmental noise and achieve accurate prediction.

[0030] Step S3, Cumulative Migration Calculation and Medium- to Long-Term Trend Analysis: The instantaneous rate output by the model is converted into the cumulative distance, which is of greater engineering interest, and long-term risks are assessed. This includes the following steps: Step S3-1, Cumulative Migration Integration Calculation: Based on the high-precision predicted rate output in Step S2-3, the cumulative deformation of seabed sediments is calculated using the numerical integration method; the calculation formula is:

[0031] in, for The cumulative creep distance at time t, For the first Predicted creep rate at time t, The sampling time interval is 1 hour. Step S3-2: Construct confidence intervals (quantify risk): Using the distribution data generated from the 100 predictions in step S2-3, calculate the standard deviation at each time point. A confidence interval is constructed; this interval visually represents the range of fluctuation in the prediction results, filling the gap in traditional methods that cannot assess the reliability of predictions.

[0032] Step S3-3, Medium- and Long-Term Trend Prediction: The model is used to simulate the cumulative deformation of submarine landslides over the next 1-10 years, and the uncertainty of long-term prediction is accumulated, providing a quantitative risk reference for the long-term safety assessment of deep-sea engineering facilities.

[0033] This embodiment selects the Shenhu Gorge creep zone in the northern South China Sea as the specific observation and prediction object, and the specific implementation process of this invention strictly follows the work process of the patent.

[0034] Before setting sail for operations, the first step is to... Figure 2 The seabed in-situ observation equipment shown underwent rigorous shore-based debugging and calibration. Specifically, this included the multi-section probe (such as...) Figure 2 (b) As shown, multiple pore water pressure sensors on the device were pressurized in a pressure chamber to ensure accurate readings under high pressure in the deep sea and that the error was controlled within the allowable range. At the same time, the battery compartment and storage unit were connected to test the write frequency and storage stability of the data acquisition module. All electronic compartments were vacuum sealed to prevent deep seawater leakage and to ensure that the equipment could support continuous operation for up to one year.

[0035] After the equipment was debugged, it was transported by ship to the target sea area of ​​Shenhu Canyon. The observation platform was lowered to the seabed surface using the research vessel's winch, and the probe was vertically inserted into the seabed sediment using gravity free-fall penetration. The equipment then entered a long-term in-situ observation period from October 25, 2024 to October 25, 2025, during which the pore pressure sensor recorded real-time fluctuations in pore water pressure within the sediment. Figure 3 As shown, the observation equipment successfully captured the complete changes in pore pressure in the sea area, as well as the significant periodic characteristics of changes with tides and internal waves, further proving the integrity of the data record. After the observation period, the equipment was retrieved via an acoustic release device, and the raw time-series data in the storage unit was exported on the deck.

[0036] After importing the acquired raw data into the computer, the baseline pressure was first determined to be 6005 kPa using long-term observation data. Then, the rate of change of pore pressure at each moment was calculated, and geotechnical experiments were conducted on the sediment samples taken from this location. We placed in-situ sediment columnar samples into a pressure chamber and applied a constant pressure. By measuring the deformation of the samples under long-term constant pressure, the creep rate can be obtained. 2.5×10 −6 / h. Another set of experiments applies different pressures and measures the deformation rate of the sediments to obtain... It is 0.35. Further, using the creep rate formula... Perform calculation Calculate.

[0037]

[0038] As shown in 4(a), using data from November 9, 2024 to November 12, 2024 as an example, this step clearly reveals the correspondence between pore pressure fluctuations and the components affected by pore pressure, the creep rate, and the cumulative creep distance. Based on this... Figure 5The study revealed the annual creep rate variation and distance during the in-situ observation period. Based on this, the cleaned one-dimensional pore pressure data was divided into three-dimensional tensors that meet the requirements of machine learning input using the sliding window technique (window size T=24 hours), thus constructing a dataset.

[0039] Based on the aforementioned dataset, a Bayesian-LSTM neural network model was constructed. In this embodiment, a Dropout layer with a dropout rate of 0.3 was embedded after the LSTM layer, and this layer was configured to remain active throughout subsequent prediction stages to endow the model with the ability to assess uncertainty. Data from the first nine months of 2024 (October 2024) to July 2025 was selected as the training set, allowing the model to repeatedly learn the nonlinear mapping between pore pressure fluctuations and creep rate. Mean squared error (MSE) was used to guide model parameter optimization until convergence.

[0040] When testing the model's performance using the reserved validation set data (August 20th - September 20th, 2025), the model was repeatedly predicted 100 times for each time point. The average of these 100 predictions was then calculated as follows: Figure 6 As shown by the red dashed line, the predicted curve perfectly replicates the peaks and troughs of the actual observed curve (black line), with a coefficient of determination R² as high as 0.948. Simultaneously, the standard deviation of 100 predictions was calculated. Figure 6 The shaded area (95% confidence interval) visually demonstrates the reliability of the current prediction results.

[0041] Finally, the trained model is applied to future risk assessments. Future pore pressure simulation data is input to calculate the cumulative displacement of submarine landslides over the next 1 to 10 years. Figure 7 As shown, the prediction indicates that the cumulative creep distance at this monitoring point will be approximately 5.91 mm in the next year, and will reach 65.23 mm in the next 10 years. The model also provides a confidence interval that gradually widens over time, indicating a risk of cumulative error in long-term predictions. It is recommended that the upper limit of the confidence interval be used as a safety redundancy reference in the design of projects such as submarine optical cables.

[0042] In the description of this invention, the term "a plurality of" refers to two or more. Unless otherwise explicitly defined, the terms "upper," "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. The terms "connection," "installation," "fixing," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a direct connection or an indirect connection through an intermediate medium. For those skilled in the art, the specific meaning of the above terms in this invention can be understood according to the specific circumstances.

[0043] In the description of this specification, the terms "one embodiment," "some embodiments," "specific embodiment," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM, characterized in that... Specifically, it includes the following steps: Step S1, Data Acquisition and Processing: First, pore pressure sensors are deployed: pore pressure sensors are installed in the seabed sediments to monitor the pore water pressure in the sediments in real time, acquiring data continuously. The sensors record pore pressure data hourly. And continuously transmit; perform preliminary preprocessing on the sampled data and standardize the raw pore water pressure data; Step S2, LSTM and Bayesian Learning Corrections: Step S2-1: Constructing a time series feature dataset: Transform the one-dimensional time series data of different variables changing over time after preprocessing in Step S1 into a three-dimensional supervised learning dataset; divide the data into two parts, the first part being the training and prediction of the model, and the second part being the verification and data comparison of the model predictions in the first part. Step S2-2: Build and train the Bayesian-LSTM network model: Construct a Bayesian-LSTM network model to record the historical fluctuation patterns of pore pressure and reflect on itself during prediction to assess uncertainty. Step S2-3, Model Probability Prediction Validation: The built network model is transformed into an executable prediction tool, and a multi-sampling and averaging strategy is used to ensure high accuracy of the prediction results. Step S3, Cumulative Migration Calculation and Medium- to Long-Term Trend Analysis: Convert the instantaneous rate output by the model into cumulative distance and assess long-term risks.

2. The method for long-term prediction of deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM as described in claim 1, characterized in that... Step S1 specifically includes the following steps: Step S1-1: Obtain the baseline pressure value through long-term data preprocessing. The baseline pressure represents the static pressure of sediment pore water without external disturbance; the superbaseline pore pressure is calculated as follows: ; Step S1-2: Calculate the pore pressure change rate using the time difference of pore pressure data. : ， in, The time interval is in hours; Step S1-3: Obtain the creep rate The creep rate is proportional to the logarithm of the pressure: ， in, This is a reference pressure, selected as the initial pressure value; The logarithmic function can be transformed into a power function using the following transformation: ， Steps S1-4: Based on the change in pore pressure, introduce the pressure change rate. The effect of the rate of change of pore water pressure on the creep rate is considered to reflect the influence of the rate of pressure change on the creep rate; the specific calculation formula is as follows: ， in, Creep rate represents the rate at which a sediment deforms at a given moment. The pore water pressure (kPa) for real-time monitoring is data monitored and recorded in real time by a pore water pressure sensor, reflecting the pressure changes of the fluid in the pores of the sediment. It is the creep index, obtained through experiments, which describes the degree of influence of changes in pore water pressure on sediment deformation, and is also the creep sensitivity of sediments; It is the creep rate, obtained through static creep experiments, and represents the basic deformation rate of the sediment under given pressure conditions; The pore pressure change rate reflects the rate of change of pore water pressure.

3. The method for long-term prediction of deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM as described in claim 2, characterized in that... Step S2-1 specifically includes the following steps: Specifically, the following steps are included: Step S2-1-1: Input Feature Definition: Select two core dimensions as the input basis for the model, namely, standardized real-time pore water pressure data, including pore pressure. and pore pressure change rate ; Step S2-1-2, Sliding window settings: Set the time step to T; set T=24, with a 24-hour cycle; Step S2-1-3: Construct tensor data: Input set X, construct a three-dimensional matrix of shape (N, T, 2). Where N is the total number of samples, T is the historical length of 24 hours, and 2 represents the two features mentioned above. Input set Y: Construct a two-dimensional vector of shape (N, 1), with the target value being the true creep rate at time t+1.

4. The method for long-term prediction of deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM as described in claim 3, characterized in that... Step S2-2 specifically includes the following steps: Specifically, the following steps are included: Step S2-2-1: Build the network structure: Input a time-series data matrix with shape (T, 2) and T=24; Step S2-2-2, Long Short-Term Memory Layer (LSTM): Set up 24 LSTM neurons to construct the core temporal feature extraction region of the model; within a complete 24-hour cycle, automatically identify and record important and recurring effective signals of tidal fluctuations, while automatically ignoring those accidental and meaningless momentary interferences; Step S2-2-3, Monte Carlo Random Deactivation Layer (MC Dropout Layer): After the LSTM layer, a Dropout layer with a dropout rate of p is connected. p is set to 0.3, that is, during each forward computation, the system will randomly block 30% of the neuron nodes, causing them to pause work, and keep 70% of the neurons in an active working state. Step S2-2-4, Fully Connected Layer: The output layer contains one neuron, which aggregates all the features extracted earlier and outputs a final value as the predicted creep rate.

5. A long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM, as described in claim 4, is characterized in that... Steps S2-3 specifically include the following steps: Specifically, the following steps are included: Step S2-3-1: Import validation data: Import the validation set data that was partitioned in S2-1 into the first part of the prediction model of the network model; Step S2-3-2: Perform Monte Carlo multiple prediction: Use the ensemble averaging method; during the prediction process, while keeping the Dropout layer p=0.3 in the model on, for each time point in the validation set, control the model to repeat the prediction 100 times, and the 100 prediction results form a prediction cluster; Step S2-3-3, Result Correction and Accuracy Verification: Calculate the arithmetic mean of the above 100 predictions as the final prediction result at this moment; Step S2-3-4, High-precision verification: Perform regression analysis between the above average predicted values ​​and the actual observed values.

6. A long-term prediction method for deep-sea creep variables based on in-situ pore pressure observation using Bayesian-LSTM, as described in claim 5, is characterized in that... Step S3 specifically includes the following steps: Step S3-1, Cumulative Migration Integration Calculation: Based on the high-precision predicted rate output in Step S2-3, the cumulative deformation of seabed sediments is calculated using the numerical integration method; the calculation formula is: ， in, for The cumulative creep distance at time t, For the first Predicted creep rate at time t, The sampling time interval is 1 hour. Step S3-2: Construct confidence intervals: Using the distribution data generated from the 100 predictions in step S2-3, calculate the standard deviation at each time point. Construct confidence intervals; Step S3-3, Medium- and Long-Term Trend Prediction: Use the model to simulate the cumulative deformation of submarine landslides over the next 1-10 years, and make the uncertainty of long-term predictions increase cumulatively.

Images