Neural network based seabed liquefaction prediction system

Abstract

The present application relates to the technical field of geological disaster prediction, in particular to a seabed liquefaction prediction system based on neural network, which comprises a parameter change extraction module, a response model training module, a disturbance feature division module, a liquefaction probability statistics module and a risk prediction output module.In the present application, the change rates of sediment particle density and effective stress in different depth intervals are extracted layer by layer to build a longitudinal response relationship, which can capture the continuous change characteristics of the sediment layer with the increase of depth, and the response differences under different time sequence disturbance conditions are compared in depth by means of residual neural network, so that the coupling relationship between disturbance information and sediment structure can be dynamically described, and the phased change of disturbance behavior can be effectively corresponded.The liquefaction probability of each depth interval is calculated by combining the disturbance feature and the response model, and the interval risk groups are formed under different time periods, and then the risk amplitude change trend and the jump boundary between disturbance stages are compared.

Description

Technical Field

[0001] This invention relates to the field of geological disaster prediction technology, and in particular to a seabed liquefaction prediction system based on neural networks. Background Technology

[0002] The field of geological hazard prediction technology involves the monitoring, analysis and prediction of potentially dangerous geological phenomena in the natural environment. Core aspects include changes in geological structure, evolution of hydrogeological conditions, distribution of geostress, analysis of seismic activity and assessment of strata stability.

[0003] Among them, the seabed liquefaction prediction system refers to the system used to determine the possibility of seabed sediments liquefying when subjected to earthquakes or other external disturbances, so as to assess the stability of sediment structures under disturbed conditions and their liquefaction risk in complex marine environments.

[0004] Existing technologies rely solely on static assessments of seabed sediment structural stability based on external disturbance conditions. They depend on the mean values ​​of overall sediment parameters at fixed time points, lacking a longitudinal continuity representation of sedimentary layer variations with depth. In multi-layered sedimentary contexts, they struggle to characterize the differences in layered responses. Furthermore, they fail to identify the stages of disturbance processes, making it impossible to effectively correlate disturbance behavior with sedimentary responses at different time intervals. This results in risk assessments lacking temporal continuity and dynamic predictive capabilities. For instance, during actual submarine earthquake disturbances, sedimentary layers may exhibit heterogeneous responses. If a single state value is used to judge liquefaction trends, local liquefaction risks may be masked or delayed in identification, affecting the accuracy and timeliness of liquefaction warnings. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a neural network-based seabed liquefaction prediction system.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a neural network-based seabed liquefaction prediction system comprising:

[0007] The parameter change extraction module obtains the rate of change of sediment particle density and effective stress over time within each fixed thickness range of the seabed sedimentary layer during a specified period, thus obtaining the sequence of seabed sediment layering parameter changes.

[0008] The response model training module inputs the sequence of changes in the seabed sediment stratification parameters into the residual neural network, compares the rate of change of particle density and effective stress of sediments in each fixed thickness range, extracts the sediment change characteristics with increasing depth, and constructs a seabed thickness response model.

[0009] The disturbance feature segmentation module obtains the shear rate and pore pressure change rate of seabed sediments within the same period, and classifies the characteristics of seabed disturbance stages based on the changes.

[0010] The liquefaction probability statistics module inputs the seabed disturbance stage characteristics into the seabed thickness response model, calculates the seabed liquefaction probability of the corresponding disturbance stage based on each fixed thickness interval, and constructs a seabed liquefaction stage risk group.

[0011] The risk prediction output module compares and analyzes the changes in liquefaction risk for each disturbance stage based on the seabed liquefaction stage risk group, and obtains the seabed liquefaction risk prediction results.

[0012] As a further aspect of the present invention, the seabed sediment stratification parameter variation sequence includes a grain density variation rate sequence, an effective stress variation rate sequence, and a thickness interval number sequence; the seabed thickness response model includes sedimentary feature extraction structure, stratification response relationship, and variation rate input weight setting; the seabed disturbance stage features include a shear rate variation rate set, a pore pressure variation rate set, and a disturbance stage corresponding time period; the seabed liquefaction stage risk group includes a liquefaction probability value corresponding to the disturbance stage, a fixed thickness interval calculation result set, and a disturbance stage time index; and the seabed liquefaction risk prediction results include a liquefaction probability variation trend, a liquefaction risk comparison relationship of the disturbance stage, and a risk result corresponding time period.

[0013] As a further aspect of the present invention, the parameter change extraction module includes:

[0014] The time series extraction submodule acquires sediment data at multiple times corresponding to preset observation points within a specified time period of a fixed thickness range of seabed sedimentary layers. It extracts the sampled value sequences of particle density and effective stress in continuous time periods and establishes a set of sediment multi-time parameter sequences.

[0015] The rate of change acquisition submodule calculates the slope of particle density change over time and the slope of effective stress change over time within each fixed thickness interval based on the multi-time parameter sequence group of the sediment, and pairs the two slopes according to the interval number to obtain the interval parameter rate of change pairing results.

[0016] The variation sequence generation submodule integrates the pairing results of the interval parameter change rate according to the arrangement order of the fixed thickness interval of the seabed sedimentary layer, and transforms the pairing value of the particle density change rate and the effective stress change rate into a continuously downward extending seabed sedimentary layer parameter change sequence.

[0017] As a further aspect of the present invention, the response model training module includes:

[0018] The sequence input processing submodule, based on the sequence of changes in the layered parameters of the seabed sediments, takes the rate of change in particle density and the rate of change in effective stress within each fixed thickness interval as input variables, and inputs them into the residual neural network structure according to the arrangement order of the segmented intervals to generate a sequence format input data set;

[0019] The longitudinal feature extraction submodule, based on the input data group in the sequence format, performs depth direction comparison on the fixed thickness interval in the residual neural network, extracts the trend information of the change rate of particle density and the change rate of effective stress with the segment sequence, and identifies the change characteristics with increasing depth to obtain the longitudinal change feature group of sediments.

[0020] The model training execution submodule defines the first third of the fixed thickness intervals as shallow intervals and the last third as deep intervals based on their arrangement in the overall stratification sequence. According to the distribution order of sediment longitudinal variation characteristic groups in the shallow and deep intervals, it applies a suppression coefficient based on the effective stress change rate to the shallow intervals and an enhancement weight based on the particle density change rate to the deep intervals, thereby constructing a seabed thickness response model.

[0021] As a further aspect of the present invention, the disturbance feature division module includes:

[0022] The disturbance parameter extraction submodule obtains the shear rate and pore pressure changes of seabed sediments in continuous time periods within the same period, extracts the slope of shear rate change and pore pressure change rate of adjacent time periods in the time series, and obtains the disturbance parameter change rate sequence.

[0023] The disturbance response construction submodule calls the disturbance parameter change rate sequence, combines and pairs the shear rate change slope and pore pressure change rate in each time period, organizes them into continuous samples according to time order, and generates a disturbance response feature group.

[0024] The disturbance stage segmentation submodule identifies the continuous segment boundaries between similar change samples based on the combined changes in the slope of shear rate change and the rate of change in pore pressure in the disturbance response feature group. It also identifies the disturbance characteristics corresponding to each time period based on similar change trends, thereby obtaining the seabed disturbance stage characteristics.

[0025] As a further aspect of the present invention, the liquefaction probability statistics module includes:

[0026] The disturbance input matching submodule calls the seabed disturbance stage features, extracts the shear rate change slope and pore pressure change rate as input parameters according to the time period index corresponding to the stage, and inputs them into the trained seabed thickness response model in sequence to generate disturbance stage input data set;

[0027] The interval probability generation submodule, based on the input data group of the disturbance stage, calls the response mechanism of each fixed thickness interval in the seabed thickness response model, calculates the liquefaction probability of each fixed thickness interval under the corresponding disturbance stage, and classifies the liquefaction probability of all fixed thickness intervals according to the disturbance stage to obtain the interval liquefaction probability sequence.

[0028] The risk group construction submodule calls the interval liquefaction probability sequence and, combined with the time period index of the disturbance stage, classifies and organizes the liquefaction probability results of all intervals under the same stage to generate a phased risk group for seabed liquefaction.

[0029] As a further aspect of the present invention, the risk prediction output module includes:

[0030] The risk value statistics submodule extracts the liquefaction probability of a fixed thickness range within each disturbance stage in the seabed liquefaction stage risk group, calculates the mean and variance of the liquefaction probability, and obtains the disturbance stage risk statistics group.

[0031] The risk magnitude identification submodule, based on the risk statistics group of the disturbance stage, combines and compares the difference in the mean risk and the difference in the variance of adjacent disturbance stages, calculates the joint risk change magnitude between the two stages, and uses the average of the slope of the shear rate change as the magnitude comparison benchmark to determine whether the magnitude exceeds the magnitude comparison benchmark, and obtains the seabed liquefaction jump indicator sequence.

[0032] The prediction result generation submodule obtains the seabed liquefaction jump identifier sequence and the disturbance stage time index, marks the liquefaction risk jump period, and obtains the seabed liquefaction risk prediction result.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0034] In this invention, by extracting the rate of change of sediment particle density and effective stress at different depth intervals layer by layer and constructing a longitudinal response relationship, the continuous change characteristics of sedimentary layers with increasing depth can be captured. A residual neural network is used to conduct in-depth comparative analysis of response differences under different temporal disturbance conditions, enabling a dynamic characterization of the coupling relationship between disturbance information and sedimentary structure. Simultaneously, the trends of shear rate and pore pressure change rate during disturbance are divided into multiple stages, effectively corresponding to the stage-wise changes in disturbance behavior. The disturbance characteristics are combined with the response model to calculate the liquefaction probability of each depth interval, forming interval risk groups at different time periods. Furthermore, by comparing the risk amplitude change trends and abrupt boundary between disturbance stages, the specific time period of liquefaction risk abrupt change is determined. This extends the assessment of liquefaction risk from static state values ​​to the time-varying evolution process under disturbance sequences, improving the resolution and sensitivity of liquefaction risk prediction in complex disturbance environments, enhancing the continuity and foresight of risk identification in the spatiotemporal distribution dimension, and ultimately achieving accurate prediction and efficient early warning of liquefaction evolution paths under different disturbance conditions. Attached Figure Description

[0035] Figure 1 This is a system flowchart of the present invention;

[0036] Figure 2 This is a flowchart of the parameter change extraction module of the present invention;

[0037] Figure 3 This is a flowchart of the response model training module of the present invention;

[0038] Figure 4 This is a flowchart of the disturbance feature division module of the present invention;

[0039] Figure 5 This is a flowchart of the liquefaction probability statistics module of the present invention;

[0040] Figure 6 This is a flowchart of the risk prediction output module of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0042] Please see Figure 1 The neural network-based seabed liquefaction prediction system includes:

[0043] The parameter change extraction module obtains the rate of change of sediment particle density and effective stress over time within each fixed thickness range of the seabed sedimentary layer during a specified period, thus obtaining the sequence of seabed sediment layering parameter changes.

[0044] The response model training module inputs the sequence of changes in seabed sediment stratification parameters into the residual neural network, compares the rate of change of grain density and effective stress of sediments in each fixed thickness range, extracts sediment change features with increasing depth, and constructs a seabed thickness response model.

[0045] The disturbance feature segmentation module obtains the shear rate and pore pressure change rate of seabed sediments within the same period, and classifies the characteristics of seabed disturbance stages based on the changes.

[0046] The liquefaction probability statistics module inputs the characteristics of the seabed disturbance stage into the seabed thickness response model, calculates the seabed liquefaction probability of the corresponding disturbance stage based on each fixed thickness interval, and constructs a stage-specific risk group for seabed liquefaction.

[0047] The risk prediction output module compares and analyzes the changes in liquefaction risk for each disturbance stage based on the stage-specific risk groups of seabed liquefaction, and obtains the seabed liquefaction risk prediction results.

[0048] The sequence of changes in seabed sediment stratification parameters includes a sequence of changes in grain density rate, a sequence of changes in effective stress rate, and a sequence of thickness interval numbers. The seabed thickness response model includes sedimentary feature extraction structure, stratification response relationship, and input weight settings for the rate of change. The characteristics of seabed disturbance stages include a set of shear rate changes, a set of pore pressure changes, and the corresponding time period for each disturbance stage. The staged risk group for seabed liquefaction includes the liquefaction probability value corresponding to each disturbance stage, a set of calculation results for a fixed thickness interval, and a time index for each disturbance stage. The prediction results for seabed liquefaction risk include the trend of liquefaction probability changes, the comparison relationship of liquefaction risk for each disturbance stage, and the corresponding time period for the risk results.

[0049] Please see Figure 2 The parameter change extraction module includes:

[0050] The time series extraction submodule acquires sediment data at multiple times corresponding to preset observation points within a specified time period of a fixed thickness range of seabed sedimentary layers. It extracts the sampled value sequences of particle density and effective stress in continuous time periods and establishes a set of sediment multi-time parameter sequences.

[0051] In the process of extracting seabed sedimentary layer data, the overall thickness of the seabed sediments is first confirmed using equipment such as sonic profilers, for example, a thickness of 30 meters in nearshore areas. This 30-meter sedimentary layer is divided into 15 vertical intervals of uniform thickness, each interval being 2 meters long. Observation positions are set at the center point of each interval, forming a monitoring grid covering the entire area by deploying multiple fixed points. After setting up the points, sampling is carried out at fixed time intervals, such as once every 3 months for two years, obtaining a total of 8 sampling time points. Each sampling uses a sediment column sampler to obtain samples from the corresponding interval. After being brought back to the laboratory, two key physical parameters are measured: particle density in grams per cubic centimeter and effective stress in kilopascals. All intervals of all observation points have sampled values ​​of these two parameters at different time points. Combining these values ​​according to "interval number, observation point number, and time point sequence" forms a complete structure containing all time-series data.

[0052] The rate of change acquisition submodule calculates the slope of particle density change over time and the slope of effective stress change over time within each fixed thickness interval based on the multi-time parameter sequence group of sediments, and pairs the two slopes according to the interval number to obtain the interval parameter rate of change pairing results.

[0053] When analyzing the evolution trends of grain density and effective stress over time in various fixed-thickness intervals of seabed sedimentary layers, it is necessary to calculate the rate of change from multi-time observation data within each interval. This rate of change is calculated using a linear fitting method, which involves performing a linear regression of the parameter values ​​on time to obtain the slope of change. The formula for calculating the rate of change of grain density is as follows:

[0054] ;

[0055] in, It represents the rate of change of particle density over time, expressed in grams per cubic centimeter per month, reflecting the amount by which particle density increases or decreases per month within a certain interval; Indicates the number of time points involved in the fitting; Indicates the first Each sampling time point is represented in months. Indicates at a point in time The particle density value observed within this interval at that time, in grams per cubic centimeter; This represents the sum of the products of the time value and the particle density value at all time points. Represents the sum of time values ​​at all points in time; This represents the sum of all particle density values; It represents the sum of the squares of all time values; It represents the square of the sum of all time values.

[0056] Set parameters: sampling time (Unit: Month): 0, 3, 6, 9, 12, 15; Particle density observations (Unit: g / cm³): 1.60, 1.62, 1.63, 1.66, 1.67, 1.69; Number of sampling points ,formula: ;

[0057] Calculate each part:

[0058] ;

[0059] ;

[0060] ;

[0061] ;

[0062] Substitute into the formula: ;

[0063] This indicates that the particle density in this range increased by an average of 0.006 grams per cubic centimeter per month.

[0064] To calculate the rate of change of effective stress within a certain thickness range of a seabed sedimentary layer over multiple sampling times, a linear fitting method is needed to determine the slope of the effective stress change over time. This slope is the rate of change of effective stress, reflecting the average trend of effective stress change per unit time. The calculation formula is as follows:

[0065] ;

[0066] in, : Represents the rate of change of effective stress over time in a certain seabed sedimentary region, with the unit being kilopascals per month (kPa / month), used to measure the trend of increase or decrease in effective stress per unit time; The total number of time points involved in sampling, i.e., the number of observed samples in the time series, is a positive integer without a unit. : No. The sampling time corresponding to each observation sample is in "months", representing the cumulative number of months calculated from the start time of the observation; : No. The measured effective stress values ​​at each time point are in kilopascals (kPa) and were obtained using experimental equipment such as a consolidation apparatus. : Represents the sum of the products of the time value and the corresponding effective stress value at each time point, in "month-kilopascal"; The sum of time values ​​at all sampling time points, in "months"; The sum of all effective stress observations, in kilopascals; : The sum of squares of all sampled time values, in "month²"; : The square of the total time, expressed in "months²".

[0067] Set parameters: sampling time (Unit: Month): 0, 3, 6, 9, 12, 15; Effective stress observation values (Unit: kPa): 11.8, 12.0, 12.4, 12.6, 12.9, 13.2; Number of sampling points .

[0068] formula: ;

[0069] Calculate each part:

[0070] ;

[0071] ;

[0072] ;

[0073] ;

[0074] Substitute into the formula: ;

[0075] The final effective stress change rate is obtained as follows: .

[0076] This calculation indicates that within this fixed thickness range, the effective stress increases at an average rate of 0.094 kPa per month. After completing the calculations for all ranges, each rate of change will be combined with the range number.

[0077] The variation sequence generation submodule integrates the results of interval parameter variation rate pairing according to the arrangement order of fixed thickness intervals of seabed sedimentary layers, and transforms the pairing values ​​of particle density variation rate and effective stress variation rate into a continuously downward extending seabed sedimentary layering parameter variation sequence.

[0078] The calculated particle density change rate and effective stress change rate for all intervals need to be integrated in a vertical order. During integration, the intervals are numbered sequentially from top to bottom, and the two change values ​​for each interval are arranged accordingly to form a complete layered change structure. To determine whether the change rate within certain intervals is abnormal, a threshold for the change rate difference needs to be set. This threshold is calculated statistically based on the change rate data for all intervals, specifically by multiplying the average value plus the standard deviation by a proportionality coefficient. ,in, This represents the threshold for the difference in rate of change, with the unit being the same as the rate of change. This represents the average rate of change across all intervals. Indicates standard deviation; This is an empirical coefficient, usually taken as 1.5. If the average rate of change of particle density across all intervals is 0.005 g / cm³ / month, and the standard deviation is 0.0015, then: If the difference in the rate of change between adjacent intervals is greater than 0.00725, then an abrupt change is considered to exist in that interval. For example, if the rate of change in one interval is 0.004 and the rate of change in the next interval is 0.012, the difference is 0.008, which is greater than the set threshold, and it needs to be marked and the data retrieved. Ultimately, the structure of the change values ​​of all intervals combined in sequence represents the vertical rate of change distribution of the seabed sedimentary profile.

[0079] Please see Figure 3 The response model training module includes:

[0080] The sequence input processing submodule, based on the sequence of changes in seabed sediment stratification parameters, takes the rate of change in particle density and the rate of change in effective stress within each fixed thickness interval as input variables, and inputs them into the residual neural network structure according to the order of the segmented intervals to generate a sequence-formatted input data set;

[0081] Based on a complete sequence of sediment parameter variations, the rate of change of particle density and the rate of change of effective stress within each fixed thickness interval are used as a set of input values, corresponding to two feature quantities, representing the average variation of particle density and the average variation of effective stress per unit time within that interval, respectively. When constructing the sequence-formatted input, it is assembled interval by interval in a segmented order from the top to the bottom of the seabed, maintaining spatial consistency. The two rate values ​​of each interval are not processed independently, but paired as a joint input term, constructing an input data sequence of length equal to the number of intervals. This sequence is not expressed numerically, but is input into the residual neural network model as a sequence data source in a "one set per interval, sequentially arranged" structure. To meet the input requirements of the neural network, all input variables need to be normalized. The normalization method is: subtract the average value of the parameter across all intervals from a parameter value, and then divide by its standard deviation, expressed as (parameter value − parameter mean) ÷ parameter standard deviation. This process is performed in parallel for all intervals and two parameters, forming a normalized sequence of particle density change rate and an effective stress change rate sequence. Each interval is arranged in a row in the format of "normalized particle density change rate, normalized effective stress change rate", ultimately forming a sequenced input data set acceptable to the residual neural network.

[0082] The longitudinal feature extraction submodule, based on the input data group in sequence format, performs depth direction comparison on the fixed thickness interval in the residual neural network, extracts the trend information of the change rate of particle density and the change rate of effective stress with the segment sequence, and identifies the change characteristics with increasing depth to obtain the longitudinal change feature group of sediments.

[0083] In the residual neural network structure, input pairs for each fixed thickness interval are input into the network model in a segmented order. The network then extracts depth features from these input pairs in the vertical direction through multiple convolutional layers and residual structures. During extraction, the system identifies the differences in the changing trends of each interval's input within the multi-layered network structure compared to its adjacent intervals, focusing on capturing the trend information of particle density change rate and effective stress change rate in the spatial depth dimension. This feature extraction process is based on the residual propagation mechanism between network layers, performing vertical comparison and residual signal analysis on the features of each depth interval to extract the trend features formed by the gradual change of parameters with depth. Essentially, feature extraction in this process involves determining whether the joint change magnitude of two input parameters between adjacent intervals is continuous, increasing, or decreasing. The judgment logic can be expressed by the following formula: If the difference between the normalized particle density change rate of the next interval and the normalized particle density change rate of the previous interval is greater than a certain set threshold, it is considered a sudden change in the density growth trend; if the difference between the normalized effective stress change rate of the next interval and the normalized effective stress change rate of the previous interval is less than a negative threshold, it is considered a rapid decrease in stress. This trend information along the depth direction is extracted and structured into a longitudinal change feature group, which includes the change rate trend type (e.g., increasing, decreasing, stable) of each interval, as well as intermediate features such as the start and end points of the trend and the trend intensity.

[0084] The model training execution submodule defines the first third of the intervals as shallow intervals and the last third of the intervals as deep intervals based on the arrangement order of the fixed thickness intervals in the overall stratification sequence. According to the distribution order of the sediment longitudinal variation characteristic groups in the shallow and deep intervals, the inhibition coefficient based on the effective stress change rate is applied to the shallow interval, and the enhancement weight based on the particle density change rate is applied to the deep interval to construct the seabed thickness response model.

[0085] When constructing a seabed sedimentary response model, to achieve quantitative adjustment of the influence of characteristics at different depths, different numerical coefficients need to be applied to the input effective stress change rate and grain density change rate according to the numbering position of the thickness interval in the whole layer sequence. The execution process is divided into two parts: shallow effective stress suppression and deep grain density enhancement, as detailed below:

[0086] I. Setting the effective stress suppression coefficient in the shallow layer

[0087] The shallow section refers to the first third of the entire sedimentary layer from the top downwards. Each shallow section is numbered. The effective stress change rate of its input will be multiplied by a suppression coefficient. The formula for calculating this coefficient is as follows:

[0088] ;

[0089] in, Number is The effective stress suppression coefficient (dimensionless) corresponding to the shallow thickness range. : Current thickness interval number (unitless, numbered from the seabed surface downwards, rounded to the nearest positive integer); : Shallow layer starting number, with a value of 1; Total number of shallow intervals, in "segments", representing the total number of intervals. One-third of an integer, that is The original effective stress change rate is in kilopascals per month (kPa / month), and the unit remains unchanged after multiplying by a coefficient.

[0090] Let the total number of intervals be... The number of shallow intervals is The calculation number is... The inhibition coefficient of the shallow region:

[0091] ;

[0092] If the normalized effective stress change rate in this interval is The final input value used for model training is: ;

[0093] This adjustment reflects the weakening of shallow effective stress input during model learning.

[0094] II. Particle density enhancement weighting in deep regions

[0095] The deep section refers to the last third of the thickness of the entire sedimentary layer from the bottom upwards. (Numbered as follows) In the deeper regions, the rate of change of the input particle density will be multiplied by an enhancement weight. The weight calculation formula is as follows:

[0096] ;

[0097] in, Number is The particle density enhancement weight (dimensionless) in the deep region. : The number of the current interval (no unit); : The starting number of the deep layer is equal to ,in This represents the total number of intervals. : Total number of deep intervals (unit: segments), consistent with the shallow intervals, taken as . The original particle density change rate is expressed in grams per cubic centimeter per month, and the unit remains unchanged after multiplying by the weight.

[0098] Still set the total number of intervals The deep layer starting number is The deep layer numbering range is arrive The calculation number is... Enhanced weights for the interval:

[0099] ;

[0100] If the normalized particle density change rate in this interval is Then the input used for network training is: This adjustment enhances the impact of changes in particle density in deeper regions on model learning.

[0101] All thickness ranges are arranged from top to bottom by number, and the effective stress change rate characteristics of the shallow ranges are subject to suppression coefficients as described above. Enhanced weights are applied to the particle density change rate characteristics in the deep layer. The features in the intermediate region retain their normalized original values. The final training input feature matrix maintains its spatial order and is fed into the residual neural network model to drive the modeling and learning process of the seabed sediment layer thickness response structure.

[0102] Throughout the model training process, the calculated results of the two coefficients are used to adjust the input feature weights for different depth intervals. The ultimate goal is to guide the residual neural network to give different levels of attention to shallow and deep sedimentary features during the training phase. In the shallow interval, the calculated effective stress suppression coefficient represents the proportion of the effective stress change rate retained in the model input. The smaller the value, the more significant the weakening of its influence, thereby reducing the participation of shallow stress information in the model. In the deep interval, the calculated particle density enhancement weight represents the amplification factor of the particle density change rate in the model input. The larger the value, the more amplified the role of this feature in training, emphasizing the dominant role of deep particle features in the sedimentary structure. The entire calculation process first divides the entire layer into three segments—shallow, middle, and deep—based on interval numbering. Then, suppression or enhancement factors are calculated sequentially according to the number position. Finally, the normalized feature value of each interval is multiplied by the corresponding coefficient to generate structured, weighted training input data, which is then input into the neural network in sequence.

[0103] Please see Figure 4The perturbation feature segmentation module includes:

[0104] The disturbance parameter extraction submodule obtains the shear rate and pore pressure changes of seabed sediments in continuous time periods within the same period, extracts the slope of shear rate change and pore pressure change rate of adjacent time periods in the time series, and obtains the disturbance parameter change rate sequence.

[0105] Two key indicators reflecting the dynamic changes of sedimentary layers were extracted from continuous time series: the slope of shear rate change and the rate of change of pore pressure. Both are calculated based on the changes in observed values ​​between adjacent time periods, and are defined as follows:

[0106] The slope of the shear rate change describes the increasing or decreasing trend of the shear rate per unit time, and its calculation formula is as follows:

[0107] ;

[0108] in, : Slope of the shear rate change; , : respectively the first Time and the The shear rate at time t; : represents the interval between two time points, in hours; essentially, it is an approximate expression of the first derivative of the shear rate with respect to time.

[0109] If the first time period of a certain observation point Hours, the observed shear rate is ;

[0110] Second time period Hours, shear rate is .

[0111] Substitute into the calculation formula: ;

[0112] This indicates that the hourly shear rate will be higher between these two time periods than before.

[0113] The pore water pressure change rate is used to measure the increase or decrease in pore pressure in a sedimentary layer per unit time. The calculation formula is as follows:

[0114] ;

[0115] in, : Pore pressure change rate, in kilopascals per hour; , : respectively the first With the The pore pressure at any given time, expressed in kilopascals. : The time interval between two time points, in hours; this value is used to describe the dynamic process of pore pressure accumulation or release within the sediment layer.

[0116] If in Hours, pore pressure is ,exist At hour, the pore pressure is given. Substituting this into the formula: ;

[0117] This indicates that the pore pressure increased at a rate of 3.0 kPa per hour during this period, reflecting an enhanced tendency for liquefaction of the sedimentary layer.

[0118] The disturbance response construction submodule calls the disturbance parameter change rate sequence, combines and pairs the shear rate change slope and pore pressure change rate in each time period, organizes them into continuous samples according to time order, and generates disturbance response feature groups.

[0119] The system receives the shear rate change slope sequence and the pore pressure change rate sequence, and maps them one-to-one along the time axis. Within each time period, these two changes are combined to form a bivariate sample. Each sample represents the structural response state of the sediment under external disturbance during that time period. The shear rate slope reflects the strain rate change of the material under shear, and the pore pressure change rate reflects the internal pressure effect of liquid seepage on the structure. The submodule organizes these samples into a continuous structure in chronological order, ensuring a coherent response trajectory across the entire disturbance process. This sample sequence constitutes a complete disturbance response feature set, recording the entire process from the occurrence and transmission of the disturbance to the reaction within the sediment structure.

[0120] The disturbance stage is divided into sub-modules. Based on the joint changes of the slope of shear rate change and the rate of change of pore pressure in the disturbance response feature group, the boundary of continuous segments between similar change samples is identified. The similarity of change trends is used as the basis for division, and the disturbance characteristics corresponding to each time period are identified to obtain the characteristics of the seabed disturbance stage.

[0121] Using a set of disturbance response features as input, the system identifies the similarity in the changing trends of samples across different time periods, and then performs stage division of the disturbance process. The system first analyzes the direction, rate amplitude, and joint trend of the changes in the slope of shear rate and the rate of change of pore pressure among consecutive samples. Under certain sample size conditions, it identifies intervals with common characteristics between adjacent samples. If multiple adjacent samples maintain a consistent trend in these two disturbance indicators, such as simultaneously showing an increase, decrease, or fluctuation convergence, this interval is marked as a disturbance stage. The division boundary typically appears at the location where a joint trend reversal, a sudden change in growth amplitude, or a significant change in fluctuation rhythm occurs between two consecutive samples. After system analysis, the disturbance process is divided into several internally consistent stages, each representing a stable response stage of the seabed sedimentary layer during the disturbance transmission process. The final output includes the disturbance stage identifier and corresponding disturbance characteristics of each time period, constituting complete seabed disturbance stage structure information.

[0122] Please see Figure 5 The liquefaction probability statistics module includes:

[0123] The disturbance input matching submodule calls the features of the seabed disturbance stage, extracts the slope of the shear rate change and the rate of change of pore pressure as input parameters according to the time period index corresponding to the stage, and inputs them into the trained seabed thickness response model in sequence to generate the disturbance stage input data set;

[0124] The identified seabed disturbance stage features are converted into input data that the model can process and matched to a previously trained seabed thickness response model. First, based on the time period index corresponding to each disturbance stage, the system extracts all disturbance parameter values ​​within that time period from the shear rate change slope sequence and the pore pressure change rate sequence. Each pair of shear rate change slopes and pore pressure change rates maintains its chronological order in structure and is grouped according to the stage division results. Next, the system inputs the parameter groups extracted for each disturbance stage sequentially into the trained model, which is capable of recognizing the transmission effects of different disturbance inputs in the thickness direction. During input, the disturbance parameters do not need to be renormalized, but the units must be consistent: the shear rate change slope is in units of seconds per hour, and the pore pressure change rate is in units of kilopascals per hour. These are input in a matched group format, maintaining the original time period order. The final result is a set of formed input samples corresponding to each disturbance stage.

[0125] The interval probability generation submodule, based on the input data set of the disturbance stage, calls the response mechanism of each fixed thickness interval in the seabed thickness response model, calculates the liquefaction probability of each fixed thickness interval under the disturbance stage, and classifies the liquefaction probability of all fixed thickness intervals according to the disturbance stage to obtain the interval liquefaction probability sequence.

[0126] The study evaluates the liquefaction response of the seabed to inputs during the disturbance phase for each fixed thickness interval. In the model, the input samples for the disturbance phase consist of two normalized disturbance parameters: the slope of the shear rate change and the pore pressure change rate. These parameters are input to a pre-trained seabed thickness response model. This model embeds a set of response mechanisms within each thickness interval to map the input disturbance signal and output the liquefaction probability of that interval under the current disturbance state.

[0127] The standard formula for calculating the probability of liquefaction is:

[0128] ;

[0129] in, : This represents the liquefaction probability of a fixed thickness range under a specific disturbance stage, with a value between 0 and 1; : The slope of the normalized shear rate change (unitless, after standard normalization, may be negative or greater than 1); : This is the normalized pore pressure change rate (also a standard normalized result); : This represents the model weight corresponding to the slope of the shear rate change in this interval, indicating the relative influence of this perturbation parameter on the liquefaction response; : The model weight corresponding to the rate of change of pore pressure in this interval; : This is the bias term for this interval, reflecting the basic liquefaction tendency of this interval under the condition of no disturbance input.

[0130] The weight parameters are fitted using the backpropagation algorithm during model training and reflect the actual impact of input perturbations on the liquefaction probability. In the response model:

[0131] If liquefaction in a certain thickness range is more controlled by shear failure, then the corresponding shear rate weight... Larger;

[0132] If liquefaction is mainly driven by pore pressure accumulation, then the weight of the pore pressure change rate... It will be higher;

[0133] During model training, it is often observed that regions closer to the surface are more sensitive to pore pressure, while regions closer to the depth exhibit a more pronounced shear effect, leading to a trend of difference between the two types of weights in the thickness direction.

[0134] The bias term represents the baseline liquefaction probability offset for this thickness range in the absence of input disturbances, and its setting is related to the following factors:

[0135] Compaction degree: Dense sediments have lower bias terms, while loose materials have higher bias terms;

[0136] Consolidation state: The bias term is negative in the fully consolidated range, and may be positive in the unconsolidated or partially saturated range;

[0137] Geological background: Long-term empirical data and field monitoring data can be used to set the initial bias artificially, and then fine-tuned during training.

[0138] During the training phase, a large amount of historical perturbation input and its corresponding liquefied label data are used. The optimal solution is obtained iteratively by minimizing the cross-entropy loss function between the model's predicted probability and the actual label. , , Parameter group.

[0139] Suppose that the normalized input parameters for a certain thickness range under the current disturbance stage are:

[0140] Shear rate change slope: Pore ​​pressure change rate: The model parameters obtained from training in this interval are: shear rate change weights: Weight of pore pressure change rate: Bias term: Substitute the values ​​into the liquefaction probability formula to calculate: ;

[0141] The results indicate that the liquefaction probability of this thickness range under this perturbation stage is approximately 85.7%.

[0142] The risk group construction submodule calls the interval liquefaction probability sequence and combines it with the time period index of the disturbance stage to classify and organize the liquefaction probability results of all intervals under the same stage, and generate the seabed liquefaction stage risk group.

[0143] Using interval liquefaction probability sequences as the data source, the liquefaction probabilities of different thickness intervals within the same disturbance stage are aggregated to generate risk groups that can be used for seabed stability assessment. The specific process is as follows: The system first reads the time period index of each disturbance stage, and then filters out all probability values ​​corresponding to that time period from the interval liquefaction probability sequence. Then, the liquefaction probabilities corresponding to all thickness intervals under the same disturbance stage are categorized and summarized to form a spatial risk set with stage-specific characteristics. During the categorization process, the system identifies whether multiple intervals have liquefaction probabilities reaching high-risk values ​​under the same disturbance stage. If multiple consecutive or adjacent intervals in a certain stage have liquefaction probabilities exceeding a preset threshold (e.g., 70%), the system marks them as a concentrated risk zone. Finally, each disturbance stage is associated with a complete seabed liquefaction risk description, including the risk level of each thickness interval, spatial distribution trend, and corresponding time period index information, forming a system-identifiable seabed liquefaction stage-specific risk group.

[0144] Please see Figure 6 The risk prediction output module includes:

[0145] The risk value statistics submodule extracts the liquefaction probability of a fixed thickness range within each disturbance stage in the seabed liquefaction stage risk group, calculates the mean and variance of the liquefaction probability, and obtains the disturbance stage risk statistics group.

[0146] Using the seabed liquefaction stage risk group as the input data source, the system extracts the liquefaction probability values ​​corresponding to different thickness ranges under each disturbance stage and performs statistical summary analysis. Specifically, the system performs centralized calculations on the liquefaction probability sets of all fixed thickness ranges within the same disturbance stage, calculating the average and variance of the liquefaction probability for that stage. The average value describes the central tendency of the overall liquefaction risk level at that stage, while the variance reflects the dispersion or spatial non-uniformity of the liquefaction risk across different thickness ranges. After processing all disturbance stages, the system unifies the average liquefaction probability and variance corresponding to each stage into a structured result, namely, the disturbance stage risk statistics group.

[0147] The risk magnitude identification submodule, based on the risk statistics group of the disturbance stage, combines and compares the difference in the mean risk and the difference in the variance of adjacent disturbance stages to calculate the joint risk change magnitude between the two stages. It then uses the average of the slope of the shear rate change as the magnitude comparison benchmark to determine whether the magnitude exceeds the magnitude comparison benchmark and obtains the seabed liquefaction jump indicator sequence.

[0148] To identify the fluctuation range of liquefaction risk between different disturbance stages, the system extracts the risk mean and variance of adjacent stages from the risk statistics group of each disturbance stage and performs a combined comparative analysis. The system performs difference processing on the average liquefaction probability between each pair of adjacent stages to obtain the risk mean difference, and simultaneously performs difference calculation on the corresponding variance values ​​to obtain the variance difference value. These two differences are combined into a joint risk change amplitude, representing the comprehensive degree of change in liquefaction risk between different disturbance stages. To determine whether this joint amplitude is significant, the system introduces a benchmark, namely the average slope of the shear rate change in the current stage, as a reference standard for the risk change amplitude. If the joint risk change amplitude exceeds this reference standard, the change is considered a "jump," indicating that the liquefaction risk has entered a new stage in the development of the disturbance. The system organizes all detected jump results in stage order to generate a seabed liquefaction jump identifier sequence.

[0149] The prediction result generation submodule obtains the seabed liquefaction jump identifier sequence and the disturbance stage time index, marks the liquefaction risk jump period, and obtains the seabed liquefaction risk prediction result;

[0150] Based on the seabed liquefaction jump marker sequence and the time index of the disturbance stage, the final risk prediction result is generated. The processing procedure is as follows: the system reads whether there is a risk jump marker for each disturbance stage and matches it with its time period index to mark the corresponding time range. In stages where a jump occurs, the system marks the stage and its corresponding time period as liquefaction risk jump periods, reflecting that the risk level of the seabed sedimentary structure during this period may rapidly evolve or tend to become unstable. All marking results are summarized in chronological order to form a structured liquefaction risk prediction result output. This output includes the start and end times of the jump period, the corresponding risk change type, and the response relationship with the shear rate disturbance trend, which is used for further safety early warning or intervention decision support.

[0151] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A neural network-based seabed liquefaction prediction system, characterized in that, The system includes: The parameter change extraction module obtains the rate of change of sediment particle density and effective stress over time within each fixed thickness range of the seabed sedimentary layer during a specified period, thus obtaining the sequence of seabed sediment layering parameter changes. The response model training module inputs the sequence of changes in seabed sediment stratification parameters into a residual neural network, compares the rate of change of grain density and effective stress of sediments in each fixed thickness range, extracts sediment change features with increasing depth, and constructs a seabed thickness response model. The disturbance feature segmentation module obtains the shear rate and pore pressure change rate of seabed sediments within the same period, and classifies the characteristics of seabed disturbance stages based on the changes. The liquefaction probability statistics module inputs the characteristics of the seabed disturbance stage into the seabed thickness response model, calculates the seabed liquefaction probability of the corresponding disturbance stage based on each fixed thickness interval, and constructs a seabed liquefaction stage risk group. The risk prediction output module compares and analyzes the changes in liquefaction risk for each disturbance stage based on the seabed liquefaction stage risk group, and obtains the seabed liquefaction risk prediction results.

2. The neural network-based seabed liquefaction prediction system according to claim 1, characterized in that, The sequence of changes in seabed sediment stratification parameters includes a sequence of changes in grain density rate, a sequence of changes in effective stress rate, and a sequence of thickness interval numbers. The seabed thickness response model includes sedimentary feature extraction structure, stratification response relationship, and input weight settings for the rate of change. The seabed disturbance stage features include a set of shear rate changes, a set of pore pressure changes, and the corresponding time period for the disturbance stage. The seabed liquefaction stage risk group includes the liquefaction probability value corresponding to the disturbance stage, a set of calculation results for a fixed thickness interval, and a time index for the disturbance stage. The seabed liquefaction risk prediction results include the liquefaction probability change trend, the comparison relationship of liquefaction risk in the disturbance stage, and the time period corresponding to the risk results.

3. The neural network-based seabed liquefaction prediction system according to claim 1, characterized in that, The parameter change extraction module includes: The time series extraction submodule acquires sediment data at multiple times corresponding to preset observation points within a specified time period of a fixed thickness range of seabed sedimentary layers. It extracts the sampled value sequences of particle density and effective stress in continuous time periods and establishes a set of sediment multi-time parameter sequences. The rate of change acquisition submodule calculates the slope of particle density change over time and the slope of effective stress change over time within each fixed thickness interval based on the multi-time parameter sequence group of the sediment, and pairs the two slopes according to the interval number to obtain the interval parameter rate of change pairing result. The variation sequence generation submodule integrates the pairing results of the interval parameter change rate according to the arrangement order of the fixed thickness interval of the seabed sedimentary layer, and transforms the pairing value of the particle density change rate and the effective stress change rate into a continuously downward extending seabed sedimentary layer parameter change sequence.

4. The neural network-based seabed liquefaction prediction system according to claim 3, characterized in that, The response model training module includes: The sequence input processing submodule, based on the sequence of changes in the layered parameters of the seabed sediments, takes the rate of change in particle density and the rate of change in effective stress within each fixed thickness interval as input variables, and inputs them into the residual neural network structure in the order of the segmented intervals to generate a sequence-formatted input data set; The longitudinal feature extraction submodule, based on the input data group in the sequence format, performs depth direction comparison on the fixed thickness interval in the residual neural network, extracts the trend information of the change rate of particle density and the change rate of effective stress with the segment sequence, and identifies the change characteristics with increasing depth to obtain the longitudinal change feature group of sediments. The model training execution submodule defines the first third of the fixed thickness intervals as shallow intervals and the last third as deep intervals based on their arrangement in the overall stratification sequence. According to the distribution order of sediment longitudinal variation characteristic groups in the shallow and deep intervals, it applies a suppression coefficient based on the effective stress change rate to the shallow intervals and an enhancement weight based on the particle density change rate to the deep intervals, thereby constructing a seabed thickness response model.

5. The neural network-based seabed liquefaction prediction system according to claim 4, characterized in that, The disturbance feature segmentation module includes: The disturbance parameter extraction submodule obtains the shear rate and pore pressure changes of seabed sediments in continuous time periods within the same period, extracts the slope of shear rate change and pore pressure change rate of adjacent time periods in the time series, and obtains the disturbance parameter change rate sequence. The disturbance response construction submodule calls the disturbance parameter change rate sequence, combines and pairs the shear rate change slope and pore pressure change rate in each time period, organizes them into continuous samples according to time order, and generates a disturbance response feature group. The disturbance stage segmentation submodule identifies the continuous segment boundaries between similar change samples based on the combined changes in the slope of shear rate change and the rate of change in pore pressure in the disturbance response feature group. It also identifies the disturbance characteristics corresponding to each time period based on similar change trends, thereby obtaining the seabed disturbance stage characteristics.

6. The neural network-based seabed liquefaction prediction system according to claim 5, characterized in that, The liquefaction probability statistics module includes: The disturbance input matching submodule calls the seabed disturbance stage features, extracts the shear rate change slope and pore pressure change rate as input parameters according to the time period index corresponding to the stage, and inputs them into the trained seabed thickness response model in sequence to generate disturbance stage input data set; The interval probability generation submodule, based on the input data group of the disturbance stage, calls the response mechanism of each fixed thickness interval in the seabed thickness response model, calculates the liquefaction probability of each fixed thickness interval under the corresponding disturbance stage, and classifies the liquefaction probability of all fixed thickness intervals according to the disturbance stage to obtain the interval liquefaction probability sequence. The risk group construction submodule calls the interval liquefaction probability sequence and, combined with the time period index of the disturbance stage, classifies and organizes the liquefaction probability results of all intervals under the same stage to generate a phased risk group for seabed liquefaction.

7. The neural network-based seabed liquefaction prediction system according to claim 6, characterized in that, The risk prediction output module includes: The risk value statistics submodule extracts the liquefaction probability of a fixed thickness range within each disturbance stage in the seabed liquefaction stage risk group, calculates the mean and variance of the liquefaction probability, and obtains the disturbance stage risk statistics group. The risk magnitude identification submodule, based on the risk statistics group of the disturbance stage, combines and compares the difference in the mean risk and the difference in the variance of adjacent disturbance stages, calculates the joint risk change magnitude between the two stages, and uses the average of the slope of the shear rate change as the magnitude comparison benchmark to determine whether the magnitude exceeds the magnitude comparison benchmark, and obtains the seabed liquefaction jump indicator sequence. The prediction result generation submodule obtains the seabed liquefaction jump identifier sequence and the disturbance stage time index, marks the liquefaction risk jump period, and obtains the seabed liquefaction risk prediction result.

Images